1. Neuroscience
Download icon

The sifting of visual information in the superior colliculus

  1. Kyu Hyun Lee
  2. Alvita Tran
  3. Zeynep Turan
  4. Markus Meister  Is a corresponding author
  1. Division of Biology and Biological Engineering, California Institute of Technology, United States
Research Article
  • Cited 2
  • Views 2,979
  • Annotations
Cite this article as: eLife 2020;9:e50678 doi: 10.7554/eLife.50678

Abstract

Much of the early visual system is devoted to sifting the visual scene for the few bits of behaviorally relevant information. In the visual cortex of mammals, a hierarchical system of brain areas leads eventually to the selective encoding of important features, like faces and objects. Here, we report that a similar process occurs in the other major visual pathway, the superior colliculus. We investigate the visual response properties of collicular neurons in the awake mouse with large-scale electrophysiology. Compared to the superficial collicular layers, neuronal responses in the deeper layers become more selective for behaviorally relevant stimuli; more invariant to location of stimuli in the visual field; and more suppressed by repeated occurrence of a stimulus in the same location. The memory of familiar stimuli persists in complete absence of the visual cortex. Models of these neural computations lead to specific predictions for neural circuitry in the superior colliculus.

Introduction

Whereas the human eye takes in about one gigabit of raw visual information every second, we end up using only a few tens of bits to guide our behavior (Pitkow and Meister, 2014). Of course those bits are carefully selected from the scene, and which specific bits get used depends entirely on the context and goals. All this happens in a processing time of about a tenth of a second (Stanford et al., 2010; Thorpe et al., 1996). How the visual brain sifts the onslaught of visual data for the few behaviorally relevant nuggets has been an enduring mystery. Much research in this area has focused on the primate visual system, and specifically the phenomena of invariant object recognition. For example, certain neurons in the inferotemporal cortex respond selectively to a specific individual’s face regardless of its position or view angle (Freiwald and Tsao, 2010), or to the concept of a specific celebrity regardless of how that concept arises (Quiroga et al., 2005). An impressive body of theory and computational modeling has been developed to explain how this sifting for important bits from the retinal output may be implemented (DiCarlo et al., 2012; Serre et al., 2007). However, empirical evidence regarding the actual biological microcircuits has been difficult to obtain.

In rodent vision, a prominent example of visual sifting is the defensive reaction of a mouse to an approaching aerial predator (De Franceschi et al., 2016; Yilmaz and Meister, 2013). Freezing or escape can be triggered reliably by an overhead display of an expanding dark disk. Effectively, the alarm circuits in the mouse’s visual system extract from the overall visual display just one or two bits of information needed to initiate action. To function properly, such an alarm system must be highly selective for the trigger feature. Indeed the mouse does not respond to expanding white disks, or to dimming dark disks, or to contracting white disks (Yilmaz and Meister, 2013). All these innocuous stimuli share some low-level features with the expanding dark disk, but not the overall configuration. Furthermore, the behavior is invariant to irrelevant features. For example, a mouse will freeze in response to looming stimuli presented anywhere in the upper visual field. It is unknown how this invariance to location arises, and how it can coexist with high selectivity for the local stimulus features.

Recent research on rodents suggests that the visual drive for these defensive behaviors arises not in the thalamo-cortical pathway but in the superior colliculus (Evans et al., 2018; Shang et al., 2018). The superior colliculus (SC) is an evolutionarily ancient midbrain structure that mammals share with birds, fish, and amphibians (Basso and May, 2017; Cang et al., 2018). The superficial layers receive inputs from the retina and in mammals also from the visual cortex, organized in a precise retinotopic map (Seabrook et al., 2017). Neurons there project to the deep layers of the SC as well as other brain areas including the lateral geniculate nucleus and pulvinar. The deep layers also receive signals from other sensory modalities including hearing and touch. Neurons in the deep SC represent pre-motor signals and project broadly to many brain areas in both ascending and descending pathways. Generally speaking neural processing in the SC identifies salient points in the environment and coordinates the orienting of the animal toward or away from such locations. In the primate brain, this has been studied extensively for the special case of eye movements (Kowler, 2011), but the primate SC also helps control head, arm, and body movements. Furthermore, the SC contributes to a type of ‘internal’ orienting, namely when we direct our attention to a specific part of the scene without overt eye movements (Krauzlis et al., 2013).

To better understand how visual sifting proceeds in the SC we recorded spike trains simultaneously from neurons throughout all layers of this structure in the awake mouse. The set of visual displays included visual threats that reliably elicit defensive reactions, and closely related stimuli that do not. We report on three kinds of neural computations that separate behaviorally relevant from irrelevant stimuli, and we trace their emergence from the superficial to the deep layers of the SC: (1) an increasing selectivity for the threat stimulus; (2) an increasing invariance to location of that stimulus; and (3) the suppression of neural responses to a familiar stimulus. In particular, this memory of familiar stimuli is stimulus-specific, lasts for a behaviorally relevant timescale, and does not require input from the visual cortex. To explain these computations we consider several circuit models, some of which can be eliminated based on the population recordings. These results suggest how circuits of the SC can effectively distill the ecologically relevant information that guides behavior.

Results

Emergence of new response properties from superficial to deep layers

To track visual computations in the mouse SC, we recorded from hundreds of neurons simultaneously in all layers of the structure using multi-electrode silicon prongs (Du et al., 2011). The animal was head-fixed, awake, and moving on a running wheel, but not trained to perform any specific task, so we could best observe the autonomous visual functions of the SC. The recording electrodes were aimed at the dorso-medial portion of the SC, which processes stimuli in the upper visual field. Over the course of several hours, we presented a battery of visual displays, ranging from abstract stimuli like flickering checkerboards to those with ecological significance, like overhead looming disks.

In analyzing neuronal responses to these stimuli, we observed a systematic progression from the superficial layers that receive retinal input to the deep layers of the SC. To illustrate the dramatic change in how stimuli are represented, Figure 1 compares recordings from two sample neurons, one in the superficial SC and the other in the deep SC.

Figure 1 with 1 supplement see all
The emergence of selectivity, invariance, and stimulus-specific habituation along the depth of SC.

(A) Left: Experimental setup. Silicon neural probes with 128 channels were implanted into the SC of a headfixed mouse viewing visual stimuli. The mouse was free to run on a circular treadmill. Middle: Diagram of a coronal section showing the anatomically defined layers of the SC (adapted from Paxinos and Franklin, 2001). sSC: superficial SC; dSC: deep SC. Right: Corresponding histological section recovered after neural recording, showing tracks of two electrode prongs. Magenta: DiI; white: anti-Calb1. (B) Extracellular spike waveforms of sample sSC (red) and dSC (blue) neurons recorded simultaneously on the silicon probe. Dots indicate the location of recording sites. Dashed line indicates boundary along the electrode array between sSC and dSC (see Materials and methods and Figure 1—figure supplement 1). (C) Response of neurons from (B) to visual stimuli. The sSC neuron (middle) responds to many types of figural stimuli (left icons: expanding black, expanding white, contracting black, contracting white, dimming, and moving black disk), whereas the dSC neuron (right) is highly selective to the expanding black disk. The sSC neuron responds robustly to every trial, whereas the dSC neuron responds primarily to the first presentation. (D) In an experiment in which looming stimuli appear from many locations (left), the sSC neuron from (B) (middle) is driven only by stimuli that cross its receptive field, whereas the dSC neuron from (B) (right) responds to stimuli placed at many more locations. White: final size of looming stimuli that elicited significant response from the cell; red: one standard deviation outline of spatial receptive field recovered by spike-triggered average method.

The superficial neuron responded well to many different kinds of displays, such as an expanding dark disk (the classic ‘looming’ stimulus), a contracting white disk, a moving disk, or a dimming disk. By contrast, the deep neuron was quite selective for the looming stimulus (Figure 1C). Second, the superficial neuron had a small and precisely circumscribed receptive field roughly 10° in diameter. It responded only when the stimulus invaded that region. By contrast, the deep neuron responded strongly to any looming stimulus presented over a wide region that encompasses much of the visual hemifield (Figure 1D). Third, the superficial neuron responded reliably to repeated presentation of the identical stimulus. By contrast, the deep neuron fired only on the first presentation and failed to respond entirely to the subsequent ones (Figure 1C, top row).

The three characteristics found in the deep SC neuron’s responses – selectivity for the looming feature, spatial invariance, and habituation to familiar stimuli – are all distinct from the signals transmitted by the retina. For example, an ‘approach-sensitive’ retinal ganglion cell (RGC) has been reported in the mouse retina (Münch et al., 2009), but later studies have found that it is actually the Off-transient alpha cell (Roska and Meister, 2014) that responds to many other Off-type stimuli in addition to the looming stimulus (Krieger et al., 2017). RGCs also have local receptive fields ranging up to 10° at most (Krieger et al., 2017), which can be readily mapped with white noise stimuli such as flickering checkerboards or bars (Zhang et al., 2012). Finally, although RGCs show complex adaptation properties, the timescale of adaptation is typically on the order of 0.1 -10 (Baccus and Meister, 2002; Wark et al., 2009), whereas the habituation we find in the deep SC lasts on the order of minutes. In the following sections, we elaborate on these response properties and how they may arise in the circuitry of the SC.

Selectivity for looming stimuli

In an attempt to measure the visual receptive fields of all the recorded neurons, we applied a flickering checkerboard stimulus and then computed the spike-triggered average (STA) stimulus (Chichilnisky, 2001). This is a common procedure that works well for retinal ganglion cells and neurons in the early stages of visual cortex (Meister et al., 1994; Niell and Stryker, 2008). In the superficial SC, the STA analysis yielded linear receptive fields that resembled those of retinal ganglion cells (Figure 2A–B). They were sharply defined in space, with the smallest only ∼5° across. They frequently showed an antagonistic and delayed surround, and some displayed orientation- and direction-selectivity (Feinberg and Meister, 2015; Inayat et al., 2015). The great majority of these neurons (∼90%) were Off cells based on the shape of the STA. By contrast, neurons in the deep SC did not produce sustained responses to the flickering checkerboard (Figure 2A), and thus contained no structure in the STA (Figure 2B). Nevertheless, these same deep SC neurons did respond strongly to certain figural stimuli, like the expanding dark disk (Figure 2A,C–D).

Figure 2 with 1 supplement see all
Selectivity to looming stimulus.

(A) Response of sample sSC (middle) and dSC (right) neurons to looming stimulus (top) and flickering checkerboard (bottom). sSC neuron is driven strongly by both, but dSC neuron is almost completely silent to the checkerboard stimulus. (B) Spatial (top) and temporal (bottom) receptive fields of the sSC (left) and dSC (right) neurons in (A) based on spike-triggered average analysis. In each subpanel, left: spatial center; right: spatial surround; bottom blue: temporal center; bottom red: temporal surround. In the temporal RF panels, the vertical line represents the time of the spike. (C) Population summary of selectivity to looming stimulus over checkerboard stimulus along the depth of SC. Horizontal dashed line indicates the boundary between sSC and dSC. Vertical dashed line separates neurons with high selectivity index (>0.75) from others. The p-value (two-sample Kolmogorov-Smirnov test) indicates that the distributions of sSC and dSC neurons differ significantly. (D) Same as (C), comparing responses to looming stimulus and contracting white disk. Selectivity index is defined as (rL-rO)/(rL+rO) where rL refers to response to looming stimulus and rO refers to response to checkerboard stimulus (C) or contracting white disk (D).

Among the various figural stimuli we tested, many neurons showed some selective tuning (Figure 1C, Figure 2D, Figure 2—figure supplement 1). We focus here on the comparison of an expanding dark disk with a contracting white disk (Figure 2D). These two stimuli are closely related in terms of local features: both contain an advancing dark edge. But the ecological interpretations are quite different: one indicates an approaching dark object and the other a receding white object. Freely moving mice take an evasive action to an expanding dark disk, but are unimpressed by a contracting white disk (Yilmaz and Meister, 2013). Compared to superficial SC, neurons in the deep SC indeed became more selective for the expanding dark disk (Figure 2D). This can be seen as sifting what is likely the most behaviorally relevant signal in the upper visual field from other distracting stimuli.

Invariance to stimulus position

Although superficial SC neurons often had sharp receptive fields just 5-10° in diameter, deep SC neurons generally responded to stimuli over a large part of the visual field. We probed this tendency with expanding dark disks presented at many different locations, as these were the most effective stimuli in the deep SC. With increasing depth in the SC, neurons showed larger receptive fields, growing by a factor of 6 in area or more (Figure 3A–B). Note that the resolution of the receptive field measurement with expanding dark disks is ∼15°, and as a result these receptive fields are larger than those measured by the flickering checkerboard (Figure 2B).

Invariance to stimulus position.

(A) Raster plot of sample sSC (left) and dSC (right) neurons recorded simultaneously during an experiment in which looming stimuli appear randomly in one of 25 locations (small black dots in cartoon) in each trial. These locations are ∼15°apart. The dSC neuron responds to many more locations than the sSC neuron and with an invariant latency. Bottom: The response amplitude at each location is reported by the brightness of the circle. X indicates a location that received no stimulus. (B) Population summary of receptive field size estimated from the experiment in (A). Vertical dashed line is at 60°. (C) Population summary of variability in the timing of the first spike from the experiment in (A). Vertical dashed line is at 75 ms. In both (B) and (C), the horizontal dashed line separates sSC and dSC. The red and blue circles denote the sSC and dSC neurons from (A). The p-values (two-sample Kolmogorov-Smirnov test) indicate that the distributions of sSC and dSC neurons differ significantly.

Despite this wide spatial range, deep SC neurons responded with a remarkably short latency to looming stimuli at any location (Figure 3A). By the time such a neuron starts firing, the expanding dark disk has only covered a few retinal ganglion cells. In contrast, for superficial neurons the latency varied depending on the location of the expanding disk stimulus and it often exceeded the latency of deep SC neurons. (Figure 3A). Figure 3C plots this variation in the latencies across the SC depth. One possible interpretation is that a widefield neuron in the deep SC pools over many local neurons in the superficial SC, such that it becomes sensitive with the same latency at every point in its receptive field. Indeed, such an interlaminar pathway has been demonstrated previously in slice preparations (Lee et al., 1997; Helms et al., 2004). We consider this possibility more thoroughly below.

In any case, it appears that certain widefield neurons in the deep SC have solved the problem of threat detection to a large degree: they signal the looming stimulus rapidly and sensitively without false alarms from stimuli that share some low-level features but not the behavioral significance.

Habituation to familiar stimuli

Neurons in the superficial layers generally produced a spike burst of comparable firing rate with every repeat of the stimulus (Figure 1C). By contrast, some neurons in the deep layers responded with a sharp burst only to the first presentation; the response to all subsequent repeats was suppressed (Figure 4A). The degree of habituation to repeated stimuli was greater in the deeper SC compared to the superficial SC (Figure 4B).

Figure 4 with 3 supplements see all
Stimulus-specific habituation.

(A) Response of a sample dSC neuron to a series of 10 looming stimuli. The first and the 10th trials are shaded in blue. Note that this neuron has a maintained baseline firing rate, which is unchanged by the stimulus on all but the first trial. (B) Population summary of habituation to repeated looming stimulus. The habituation index is defined as 1-r1/r10 where ri refers to the number of spikes fired in in i-th trial after subtracting background activity. The horizontal dashed line separates sSC and dSC. The vertical dashed line is at 0.75. The blue circle is the sample dSC neuron from (A). The p-value (two-sample Kolmogorov-Smirnov test) indicates that the distributions of sSC and dSC differ significantly from each other. (C) Response of a sample dSC neuron to ~100 presentation of looming stimuli delivered in random sequence. Each subpanel represents response to stimuli at one of the 25 locations. Bottom left: two of the response traces from above. Even after the neuron has habituated to stimuli at one location (magenta), it responds strongly to the first stimulus at another location (blue). Bottom right: response of all dSC neurons in this recording, normalized by response to first trial of the magenta trace. Data points are medians and error bars range from 25th to 75th percentiles. (D) Summary of time to recover from habituation for a group of simultaneously recorded sSC and dSC neurons. Even after ~120 s, dSC neurons do not recover beyond 50% of the initial response. Data points are medians and error bars range from 25th to 75th percentiles. (E) Sample sSC (top right) and dSC (middle right) neurons recorded in a mutant mouse that does not develop the neocortex or the hippocampus (left). The dSC neuron in the mutant mouse also shows habituation. Bottom: population response of dSC neurons to 10 presentations of the looming stimulus, normalized by the response to the first presentation. Data points are medians and error bars range from 25th to 75th percentiles.

The onset of this habituation is immediate and already affects the response ∼1 s later (Figure 1C, Figure 4A). The suppression then lasts for minutes: many deep SC neurons showed less than 50% recovery even after ∼120 s (Figure 4D). While we have not measured the exact time course of recovery, we found that the suppression was not permanent. In general, neurons recovered the full sensitivity to the first presentation when probed again about an hour later (Figure 4—figure supplement 1). Furthermore, the burst of spikes was not driven simply by a change in locomotor output or pupil size as a secondary consequence of the visual threat (Figure 4—figure supplement 2).

Remarkably, this habituation was strictly specific to the stimulus that caused the response. As reported above, widefield neurons in the deep SC can be triggered by looming disks at many different locations (Figure 1C, Figure 3A). Figure 4C shows the response of a single neuron to a looming stimulus whose location was chosen randomly on every trial. By comparing the sequence of responses at one location to that at another one can test whether the habituation transfers across space. As shown in the bottom left panel of Figure 4C,a stimulus at one location did not suppress the subsequent response of the same neuron to a stimulus at another location, even separated by as little as 15°. One interpretation is that the habituation takes place in local circuits spanning ∼15° in width before their output gets pooled by the widefield neuron.

Given that the memory for familiar stimuli can last 2 min or longer, we considered whether the hippocampus or the neocortex play a role in storing this information, perhaps by modulating the gain of collicular signals through the extensive projections from visual cortex (Zhao et al., 2014). Thus, we repeated the experiments in a mutant mouse that lacks all the dorsal forebrain, including the hippocampus and most of the neocortex (Kim et al., 2010; Figure 4—figure supplement 3). Intriguingly, the mutant also showed long-lasting suppression of repeated stimuli in deep neurons of the SC (Figure 4E), to a degree that matched the suppression seen in the normal mouse (Figure 4C bottom right and Figure 4D). This is consistent with a local mechanism for habituation within the SC.

The preceding analyses of single-neuron responses suggest that the neural population deep in the SC selectively represents those bits of information that may be of immediate relevance to defensive reactions, while other aspects of the visual display get discarded. To test this directly, we applied a linear decoder to the population vector from neurons in superficial and deep SC. From single stimulus trials, the decoder easily read out the precise location of a visual stimulus from the population in superficial SC, but much less so from neurons in deep SC (Figure 5, left). By contrast, the deep SC represented explicitly whether a stimulus appeared at a novel or a familiar location, whereas that information was barely available in the superficial SC (Figure 5, right). Of course a decoder with access to the entire history of responses could decode stimulus novelty also from the superficial SC. By contrast, in the deep SC that information is available on individual trials. In the next section, we explore how the information about stimulus history may be stored by the collicular circuit.

Population decoding of distinct stimulus features.

Linear decoders were trained with simultaneously recorded sSC and dSC neurons to predict location (left) and novelty (i.e. whether the stimulus has appeared at a location for the first time) (right) of stimuli in the experiment described in Figure 3. Dashed line: chance performance; error bars: one standard deviation across different subsamples of cells.

A working model for circuit mechanisms of visual sifting

The microcircuitry of the SC is still poorly understood, at least compared to that of the retina. One can distinguish about 5 to 10 neuronal types based on morphology and gene expression (Byun et al., 2016; Gale and Murphy, 2014), but their synaptic connectivity is largely unknown. Furthermore, the SC interacts through long-range connections with other brain regions, notably the visual cortex (Seabrook et al., 2017). Nevertheless, it is useful to consider what circuit mechanisms may produce the observed visual responses of SC neurons. The functional evidence we have gathered here makes some potential explanations unlikely, and supports others as a guide in future studies of synaptic connectivity. Here, we focus on explaining three aspects of visual processing encountered in some deep SC neurons: the selectivity for looming stimuli, the invariance to spatial location, and the long-lasting stimulus-specific habituation. None of these phenomena occur in responses of retinal ganglion cells, and thus they must arise from post-retinal circuitry.

One circuit model that accounts for all the observed effects is shown in Figure 6A (‘the working model’). It starts with input signals from retinal ganglion cells. Those are combined to produce neurons selective for a local looming stimulus. The outputs of many such local looming detectors are pooled to produce neurons with widefield sensitivity and position invariance. Finally, the input synapses to those widefield neurons undergo a short-term synaptic depression that accounts for the stimulus-selective habituation.

Figure 6 with 1 supplement see all
Model of selectivity, invariance, and stimulus-specific habituation.

(A) The ‘working model’ of how selectivity, invariance, and habituation arise in the dSC. Looming selectivity is generated by combining fast and slow Off-type retinal inputs (green and pink) in the local looming detector (purple) in sSC. Inset on right shows spatial layout of these inputs. Invariance arises from pooling these local looming detectors to a single global looming detector (cyan) in the deep layers. The stimulus-specific habituation is achieved by synapses that undergo activity-dependent short-term depression (red downward arrows). Solid circles: excitation; open circles: inhibition. (B) An alternative model of looming selectivity based on pooling directionally tuned inputs. (C, D) Alternative models of stimulus-specific habituation: the same input as the excitation drives a persistent inhibition (C) or a facilitating inhibitory synapse (D). (E) Simulation of responses to various figural stimuli. Green: excitation from center; red: inhibition from surround; shaded black: net response. (F) Simulation of stimulus-specific habituation. Each local looming detector connects to the global looming detector with a synapse whose strength w decays rapidly and recovers slowly.

To simulate the function of this circuit we modeled each of the neurons as a Linear-Nonlinear element (Chichilnisky, 2001), and the synapses according to a widely used formalism for short-term plasticity (Tsodyks et al., 1998). This model correctly recapitulates the preference for looming over other stimuli (Figure 6E); the position invariance; and the habituation to familiar stimuli (Figure 6F). It even accounts for detailed dynamics of the looming response in deep neurons, such as the short latency (Figure 3A) and the rapid quenching of the response caused by synaptic depression (Figure 1C, Figure 4A).

While a successful circuit model seems promising, one learns something useful only from comparing different explanations. Here, we consider several alternative microcircuits to account for the looming selectivity and the stimulus-selective habituation.

The working model (Figure 6A) builds on local looming-selective neurons. We encountered multiple cells in the superficial SC that match this profile: a local receptive field, looming selectivity, and little habituation (Figure 6—figure supplement 1). In the working model, this selectivity is achieved by combining signals from retinal ganglion cells (RGCs) with different dynamics: excitation from a fast and transient Off-cell forms the receptive field center, and inhibition from slow and sustained Off-cells forms the surround. Since RGCs are excitatory, the inhibition requires interneurons in the SC, and the slow dynamics of the surround may well result from filtering by those interneurons. In either case, the concentric organization of fast excitation and slow inhibition produces selectivity for looming over contracting white or moving or dimming stimuli (Figure 6E).

As an alternative explanation, could the looming selectivity already originate in RGCs? As we noted previously, the ‘approach-sensitive’ Off-cell that has been previously reported (Münch et al., 2009) is now known to correspond to the Off-transient alpha cell (Roska and Meister, 2014) which – while sensitive to looming stimuli – responds equally well to dimming and flashing spots (Krieger et al., 2017). Therefore, these RGCs do not qualify as the local looming detectors.

Another possibility is that looming selectivity results from a radial organization of direction-selective (DS) neurons, each of which reports a segment of the advancing dark edge (Figure 6B). Supposing those DS inputs come from the retina, the only candidates are the On-Off DS RGCs (Sanes and Masland, 2015), which would be equally sensitive to On edges. Thus, the looming detectors in the SC should respond to an expanding white disk as well, unlike what we observed (Figure 1C). If, on the other hand, the DS signals are generated de novo in the SC, one would expect to find such interneurons with all possible preferred directions. Instead, DS neurons in a given region of the superficial SC have a strong bias for just one or two preferred directions (de Malmazet et al., 2018). In summary, both of the considered alternative microcircuits for looming selectivity seem unlikely given the available evidence.

In the working model (Figure 6A), the stimulus-selective habituation is produced by activity-dependent depression of the synapses that convey the local looming signals to the widefield neuron. A plausible alternative mechanism would involve long-lasting inhibition of the looming detector from a neuron triggered by that same local stimulus (Figure 6C). This neuron would need to exhibit a sustained activity following a single stimulus. In our database of collicular recordings, we never encountered a neuron that matches this description. Another possibility is that local looming detectors – in addition to exciting the widefield neuron – also inhibit it via an interneuron (Figure 6D). Then the long-lasting habituation could be explained by the potentiation of the inhibitory synapse, rather than depression of the excitatory synapse. In that case, one might expect that repeated looming stimuli should produce a suppression of the ongoing baseline firing during later stimulus periods. We never observed such a suppression (Figure 4A). Instead the firing generally increased during stimulus intervals (rstim) compared to inter-stimulus intervals (risi) (for 15 deep SC neurons with baseline firing > 10 spikes/s, median rstim/risi, 25th-75th percentile range: 1.03-1.85).

In summary, several alternative explanations for the basic phenomena observed in deep SC neurons seem less likely than the working model that we propose, based on our database of extracellular recordings. We suggest that the key components of the working model in Figure 6A, namely the microcircuit for looming selectivity and the long-lasting synaptic depression, are fruitful targets for further investigation.

Discussion

Summary

The superior colliculus (SC) presents an interesting interface between purely sensory representations and pre-motor signals. Our goal here was to follow systematically how the sensory inputs from the retina get digested and filtered in the SC. As a guiding problem we chose a robust visually-triggered behavior: the defensive reaction elicited by an overhead looming stimulus. By following visual responses of neurons from superficial to deep layers, we documented three aspects of the sifting process: (1) an increasing selectivity for the behaviorally relevant looming stimulus over other innocuous stimuli with similar low-level features (Figure 2); (2) an increasing invariance to other aspects of the visual display, such as the precise location of the threat stimulus (Figure 3); and (3) an increasing selectivity for novel over familiar stimuli (Figure 4). We considered how this filtering may be achieved by neural circuits and arrived at a plausible model of circuitry in the SC (Figure 6) that accounts for all three of the phenomena of visual sifting considered here. Moreover, several alternative circuit-level mechanisms were found to be inconsistent with the neural signals we encountered.

Relation to earlier work

Some of the phenomena reported here have been described before in a wide range of species. A common theme is that neurons in deep SC respond over larger regions of the visual field, while retaining a preference for small stimulus features within that region (Cynader and Berman, 1972; Dräger and Hubel, 1975; Gordon, 1973; Humphrey, 1968; Ito et al., 2017). Also, the remarkably persistent habituation to repeated stimuli has been noted previously, even in the earliest recordings from optic tectum (Cynader and Berman, 1972; Dräger and Hubel, 1975; Horn and Hill, 1966b; Lettvin et al., 1961; Straschill and Hoffmann, 1969; Woods and Frost, 1977; Reches and Gutfreund, 2008). Another repeated observation is that the visual cortex appears dispensable for many aspects of visual processing in the SC (Horn and Hill, 1966a; Humphrey, 1968; Masland et al., 1971), although it does play a subtle modulatory role (Zhao et al., 2014). Looming stimuli are particularly effective for many neurons in the superficial SC (Zhao et al., 2014). Interestingly, the early literature missed this, perhaps because of the technical difficulty of generating an expanding dark disk with the commonly used hand-held slide projector (Dräger and Hubel, 1975). Our present report places these disjoint observations into a common context, namely the animal’s need to distill a specific signal of ecological value from the broad range of visual stimuli. We show that SC neurons are not only sensitive to looming stimuli but also become increasingly selective in deep layers, an essential requirement for an alarm system. Further we analyze the neural code at the population level throughout this brain region, which reveals the gradual progression of stimulus filtering. Finally, we consider how these aspects of neural representation relate to neural circuitry, and evaluate alternative hypotheses for such circuits. The results allow a broader consideration of how selectivity and invariance come about in brain processing, to be pursued further below.

Ethological significance

The present study focused on stimuli presented in the upper visual field and recordings performed from the corresponding medial region of the SC. Arguably, the most behaviorally relevant event in the upper visual field is the impending arrival of a bigger animal, such as an aerial predator. The imminent threat that these events pose may account for the profuse responses to dark looming stimuli among SC neurons in this region (Figure 2; Zhao et al., 2014). Of course the threats must be distinguished from innocuous events, like the movement of overhead foliage, or the obscuring of the sky when the animal moves under shelter. The increased selectivity to the expanding dark disk in the deeper SC can account for that selectivity (Figure 2C, Figure 4B).

How should one interpret the profound habituation to repeated stimuli in this context? For one, the habituation does not interfere with the alarm response, since the animal must react to the first occurrence of a clear looming stimulus (Yilmaz and Meister, 2013). If the animal escapes or freezes, and the predator approaches a second time, this is likely in a different part of the visual field, and thus unaffected by the location-specific habituation. On the other hand, if the same stimulus recurs periodically in the same location, it is more likely caused by a leaf waving in the wind. Thus, the habituation can be seen as another processing strategy to reject innocuous events from the alarm pathway.

In the lower visual field the animal has different behavioral needs, such as picking out seeds against a cluttered background, following small moving prey (Hoy et al., 2016), perhaps identifying urine marks (Joesch and Meister, 2016), and tracking optic flow. Furthermore the connectivity between SC and other brain areas seems to differ in the upper and lower visual fields (Savage et al., 2017). Thus, one expects a corresponding difference in the rules by which visual stimuli are sifted there, a fertile area for future study.

Selectivity, invariance, and habituation

One remarkable phenomenon in sensory processing is the emergence of neuronal responses that are both highly selective and broadly invariant. For example, certain ‘face cells’ in the primate visual cortex respond selectively to one person’s face regardless of the view angle, scale, or illumination (Freiwald and Tsao, 2010). How do these seemingly conflicting characteristics arise within sensory circuits? In the working model we propose here (Figure 6A) the answer is ‘first selectivity then invariance’. An AND operation across input neurons with different dynamics generates a local looming-selective neuron. These pattern detectors are distributed across the visual field. Then an OR operation pools across many local pattern detectors to produce the position-invariant response of the widefield neurons (Figure 7A).

The logic of selectivity and invariance.

In (A) feature selectivity is accomplished by combining local input signals (red and green) with AND logic (X). Then invariance arises from combining many of those feature signals with OR logic (+). In (B) there is only a single feature computation (X). Invariance is achieved by routing its inputs to local signals in different parts of the visual field. Arrows indicate where the stimulus-specific habituation must take place.

This seems to be the scheme in other neural systems where the circuitry is understood. For example, in the auditory brain of the barn owl certain high-order neurons are selective for a particular interaural time delay, but invariant to the frequency of the sound (Konishi, 2003). These appear to arise from OR pooling over lower-order neurons that are selective for the same time delay but still tuned to different frequency bands. Those delay detectors in turn arise from an AND combination of signals derived from the two ears (Carr and Konishi, 1990). A similar processing scheme applies in the electrolocation circuits of weakly electric fish that exhibit a jamming avoidance response sensitive to frequency but invariant to many other parameters of the electric field (Heiligenberg, 1989).

However, this is not the only solution. In the case of face recognition, for example, it seems implausible that the brain should build separate pattern detectors for each face at each retinal location, and then pool over those to achieve invariance. An alternative scheme produces invariance first and then selectivity (Figure 7B). Here, there exists only a single pattern detector. But the inputs to this neuron are routed to ‘look at’ different spatial locations through a shifting circuit. The sudden appearance of any stimulus could engage these shifter circuits to route the corresponding low-level visual signals into the pattern detector (Olshausen et al., 1993; Ullman and Soloviev, 1999).

The observation of habituation and its specificity to location seems to greatly favor one of these schemes. Recall that habituation is seen prominently among neurons in the deep SC that are already highly pattern-selective. In the ‘selectivity first’ scheme, that places the gain modulation somewhere prior to the output of the pattern detector, which is the last spatially localized signal (Figure 7A). By contrast, the ‘invariance first’ scheme requires the gain modulation to occur in low-level visual neurons that are not yet pattern-selective (Figure 7B). This conflicts with our observations of neurons in the superficial SC that do not show location-specific habituation (Figure 4B). In summary, the robust observation of location-selective habituation in neurons of the deep SC favors a circuit model that develops selectivity before invariance.

Of course one can also envision intermediate solutions. For example, there is speculation that the visual cortex implements an alternation of AND and OR stages through a hierarchy of anatomical areas (DiCarlo et al., 2012; Riesenhuber and Poggio, 1999). Seeing that most vertebrate species do not have a neocortex, yet must solve the same problems of invariant pattern recognition, the SC seems like a promising arena for the study of high-level visual processing.

Circuit mechanisms of sensory sifting

While it is tempting to suppose that the observed reduction of visual data is in fact performed within the SC, such anatomical localization is not a binding conclusion. In the extreme, the neurons of the deep SC, with their selectivity for fine spatio-temporal features and localized habituation, may simply reflect the output of a computation that occurs elsewhere. The SC interacts with many other anatomical structures (Basso and May, 2017; Savage et al., 2017), often in a reciprocal fashion. The most prominent such input, namely from the visual cortex, is likely not required for the essentials of visual sifting, based on our results with mutant mice (Figure 4E) and prior work with cortical silencing (Zhao et al., 2014) and ablation (Horn and Hill, 1966a; Humphrey, 1968). To contribute to sifting, the partner areas should retain a spatial resolution of the stimulus on the order of 10°. This constraint eliminates some small nuclei, but leaves several candidates in place, for example the thalamic area LP (Allen et al., 2016) and the parabigeminal nucleus. Given the position of the SC as a hub of brain pathways, it is an open question whether one can ultimately assign discrete computational functions to discrete anatomical areas.

On a finer level one may ask how the circuit models of Figure 6 map onto neuron types in the SC. About five cell types have been distinguished in the superficial SC of mammals based on morphology alone (Langer and Lund, 1974; May, 2006), and more recent studies have connected these types to visual responses and electrophysiological properties (Gale and Murphy, 2014). The most compelling by their visual appearance are the so-called widefield or bottlebrush neurons. These cells have a dendritic fan that extends towards the surface of the SC and spreads out laterally to cover a large area in the retinorecipient layers. Each dendrite terminates in a bottlebrush-shaped ending, and the overall morphology is startlingly similar across birds and mammals (Luksch et al., 1998; Major et al., 2000). The widefield neurons of mammals project to the pulvinar, and the axon forms multiple collaterals in the SC that could propagate the output to the deep layers (Basso and May, 2017; Major et al., 2000).

By virtue of their broad dendritic tree these widefield neurons offer themselves as the substrate for pooling across spatial locations, as in the working model of Figure 6A. Two further features recommend such an identification: First, the dendrites of widefield neurons generate spikes that propagate to the soma (Endo et al., 2008; Luksch et al., 2004). In this way, the neuron truly implements an OR operation across its inputs (Figure 7A): when any of its inputs fire, the output will fire. Second, experiments on chick tectum showed that each dendritic input undergoes a profound synaptic depression that lasts several seconds, but does not affect the function at another dendrite (Luksch et al., 2004). This could account for the location-specific habituation as in the model of Figure 6A. However, there is some question whether this synaptic depression also happens in the mouse (Gale and Murphy, 2016). Also we found a substantial increase of invariance below the anatomical stratum where the widefield neurons reside (Figure 3).

In summary, the visual response properties of deep SC cells differ dramatically from any signal that emerges from the retina, and it is tempting to associate this transformation with the bottlebrush neuron that is shaped unlike anything in the retina. Some caution is in order, of course. The diagram of Figure 6A should be viewed as a conceptual scheme rather than an explicit circuit with one-to-one corresponding real neurons. Perhaps the selectivity and invariance are accomplished in multiple stages, or with the contribution of other brain areas. Or the local looming detectors may be nonlinear dendrites, and ion channels with long-lasting inactivation (Ulbricht, 2005) may play the role of depressing synapses. The increasing availability of genetic handles for cell types in the SC (Byun et al., 2016; Gale and Murphy, 2014) should help in cracking some of these microcircuits.

Materials and methods

Mouse, surgery, neural recording and spike sorting

Request a detailed protocol

We used C57BL/6 mice (RRID:IMSR_JAX:000664) aged 3-10 months (both males and females, Jackson Labs) for electrophysiological recordings. To prepare an animal for an experiment, we first implanted a metal headplate to the skull with a dental adhesive (3M Scotchbond) under anesthesia (2% isoflurane). After three days of recovery, the animal was habituated to being head-fixed on a circular treadmill for ∼30 min/day for 3 days. On the day of recording, the animal was again anesthetized and a craniotomy (< 1 mm diameter) was made over the SC (0.2-0.4 anterior to lambda, ∼0.5 mm lateral from midline). A small hole was made over the cerebellum to insert a silver reference wire. The craniotomy was then closed with a silicone elastomer (Kwik-Cast, WPI). After 6-8 hours of recovery, the animal was head-fixed and the craniotomy was exposed. A silicon neural probe was then lowered slowly into the brain (< 5µ m/s) and the depth from brain surface was recorded. The craniotomy was then covered with mineral oil to prevent drying of the exposed tissue. We waited a short period (15-30 min) for signals to stabilize before starting the recording. A typical recording session lasted 2-3 hours. All procedures were performed in accordance with institutional guidelines and approved by the Caltech IACUC.

The silicon neural probes were obtained from Sotiris Masmanidis (UCLA) (Du et al., 2011). For the majority of experiments, probe types 128A, 128AN, and 128DN were used. For data acquisition we used the RHD2000 128-channel amplifier board and the RHD2000 USB interface board (Intan). Auxiliary signals including the movement of the running wheel, timing of the stimulus, and timing of pupil video recording were collected concurrently with the neural signal. We used KiloSort (Pachitariu et al., 2016) for spike sorting of the data. The output of the automatic template-matching algorithm of KiloSort was visualized and manually curated on Phy (Rossant et al., 2016; Rossant, 2017).

To test if the long-lasting stimulus-specific habituation requires the neocortex and the hippocampus, we also recorded in mutant mice that developmentally lack these brain areas (Kim et al., 2010). These animals were bred by conditional knockout of exon 3 of Pals1 gene in cortical progenitor cells during embryonic development, achieved by crossing Pals1flox/flox mice with LoxP sites inserted upstream and downstream of exon three with Emx1-Cre animals (Jackson Labs, Strain 005628) expressing Cre recombinase in the cortical progenitor cells. Conditional knockout of both copies of Pals1 due to Cre-mediated recombination during development resulted in Emx1-Cre:Pals1flox/flox homozygous progeny used in this study (Figure 4—figure supplement 3).

Behavioral measures

Request a detailed protocol

The animal’s pupil diameter and locomotion on the circular treadmill were recorded along with the neural signals. The animals were not trained in any particular task and varied in their tendency to run on the treadmill. When looming stimuli were presented, the animals sometimes reacted by stopping (if the stimulus had arrived during a movement bout) or showing an increase in the pupil size (Figure 4—figure supplement 2C), but no characteristic behavioral output was consistently observed. However, we could rule out the possibility that the strong response of deep SC neurons to the first presentation of the looming stimulus is a simple consequence of motor output or change in pupil size, as they were usually not modulated by these factors in the absence of the looming stimulus (Figure 4—figure supplement 2). We also tracked the position of the pupil to monitor the eye movements. In many cases, the eyes were very stable, as demonstrated by the sharp (∼5°), circular receptive fields we recovered (Figure 2B) in the superficial SC by spike-triggered average analysis.

Post-hoc identification of the recorded brain area

Request a detailed protocol

Prior to implanting into the brain, the tip of the silicon probe was covered with fluorescent lipophilic dye (DiD or DiI, Invitrogen). Immediately after recording, the animal was anesthetized and perfused with saline and 4% PFA. The brain was harvested and fixed with 4% PFA (Electron Microscopy Sciences) for 24-48 hours at 4°C, after which it was sectioned coronally at 100 µm thickness with a vibratome (Leica). The sections were then stained with anti-Calb1 antibody (Swant, CB-38a, 1:1000 dilution), which has been previously reported to label the superficial gray layers of the SC (Rousso et al., 2014). Following secondary antibody staining (AlexaFluor 488, donkey-anti-rabbit, 1:1000 dilution), sections were mounted with Vecta-Shield:DAPI and imaged using a confocal microscope (LSM800, Zeiss). From this we could estimate the location of the probe relative to SC layers (Figure 1—figure supplement 1). This histology-based method of localizing the probe relative the SC layers was complemented with current source density (CSD) analysis. First, the raw, broadband recording was low-pass filtered (150 Hz cutoff) to isolate the LFP band. Then the Laplacian of a column of spatially contiguous electrodes was computed and smoothed with a Gaussian kernel. This revealed a series of current sources and sinks in response to visual stimulation (Figure 1—figure supplement 1). By comparing this CSD analysis to the histological localization, we confirmed the results from Stitt et al. (2013) that the inflection point between the current source and sink marks the bottom of the superficial gray layer (SGS). We then defined the boundary between the superficial and deep layers as 100 µm below the inflection point (corresponding to 0 depth in Figure 2C–D, Figure 3B–C, and Figure 4B) to account for the thickness of the optic layer.

Stimuli

Visual stimuli were programmed using the Psychtoolbox (Brainard, 1997; Kleiner et al., 2007) package in MATLAB (Mathworks) and presented on a gamma-corrected monitor (IPS231, LG) at baseline luminance of ∼25 cd/m2. The position of the monitor was slightly adjusted in each experiment such that the receptive fields of the neurons being recorded were located near the center of the monitor. Usually this was at ∼35° in elevation and ∼45° in azimuth (to the left) from the rostro-caudal axis of the animal. The monitor was located 15-20 cm from the animal and covered ∼120°of the horizontal field of view. The visual stimuli were synchronized to the neural recording by using a photodiode to send timing pulses from the monitor to the data acquisition board.

Before presenting the stimuli, we used a small flickering spot to map the part of the monitor that elicited strong neural responses (‘response zone’). Figural stimuli were then presented at these locations. The following is a description of each stimulus type during the stimulus period.

Definition of stimulus period

Request a detailed protocol

Throughout this report, the periods during which the stimulus was presented on the screen are called stimulus periods and are marked as pink sections in the PSTHs. Outside the pink sections, the screen was uniformly gray.

Looming stimulus

Request a detailed protocol

The looming stimulus expanded from 0° to ∼30° at a linear expansion rate of ∼30-60° /s and then remained stationary for another 250 ms before disappearing. It was presented at the full contrast achievable by the monitor. This repeated for 5-10 trials at the same location. The inter-stimulus interval was 1-3 s, except when the time to recover from habituation was explicitly tested (Figure 4D).

Other figural stimuli

Request a detailed protocol

The contracting black, expanding white, and contracting white disks were presented with similar parameters as the looming stimulus. The stationary period of 250 ms was always at the end of the expansion or the contraction. The dimming and the moving dark disks were the same size as the final size of the looming stimulus. The rate of change in contrast of the dimming disk and the trajectory and the speed of the moving dark disk were set such that they had roughly the same duration as the looming stimulus. The moving dark disk traveled at ∼40-70° /s, with the response zone in the middle of the trajectory. Several different movement directions were tried.

Flickering checkerboard

Request a detailed protocol

During the flickering checkerboard stimulus, the entire screen was divided into square checkers (∼3°) whose intensity changed randomly between black and white in every frame at a refresh rate of 60 Hz. The duration ranged from 300 to 600 s, but often 300 s was enough for the spike-triggered average analysis.

Random loom

Request a detailed protocol

In the ‘random loom’ experiment, 25 locations (in a 5 × 5 grid) around the response zone were selected, with ∼15° between adjacent locations (measured from center to center). In each trial, one looming stimulus was presented in one of these locations with the parameters described above. The sequence of stimulus locations was determined with a pseudorandom number generator. The inter-stimulus interval was 3 s and ∼60-120 trials were presented in total.

Analysis

The progression of visual response properties with depth in the SC was discovered in early exploratory experiments. A subsequent round of recordings was performed to validate the initial observations. The present manuscript analyzes data from only these replication experiments. All analysis scripts were written in MATLAB R2016b (Mathworks) unless otherwise noted.

Definition of neural response and background activity

Request a detailed protocol

Throughout our analysis, the neural response is defined as the number of spikes that a neuron fired during the stimulus period (as described above). Some neurons had a maintained baseline firing rate. The background activity is defined as the expected number of spikes contributed by the baseline firing rate during the stimulus period. To compute this, we estimated the baseline firing rate by counting the spikes fired during the ∼5-10 s-long period just preceding the stimulus and dividing by length of this period. We then multiplied this by the stimulus period to get the background activity. The background activity was used to test if the neural response was visually driven (see below).

Identification of visually responsive neurons

Request a detailed protocol

Many of the recorded neurons had no clear response to visual stimuli. In a typical neurophysiology experiment, visually responsive neurons can be separated from others by presenting the stimulus many times and choosing only those that respond consistently across repetitions. In our experiments, we did not have the luxury of repeating the stimuli, as many neurons (esp. in the deep SC) showed significant habituation after just a single presentation (Figure 4A–B). To identify visually responsive neurons from single trials, we instead used a statistical method. First, we computed the neural response and the background activity (see above). We then computed a p-value for the neural response based on a Poisson noise model whose mean was the background activity. If the p-value was less than the pre-set cutoff of 0.005, we considered the response to be visually driven. In cases where the background activity was very low (< 1 spike), the mean of the Poisson model was set to one so that chance firing of 1–2 spikes during the stimulus period would not be considered as a visual response.

This significance criterion was used to select neurons to include in the analysis shown in Figure 2, Figure 3 and Figure 4 (see below). When computing quantities of interest (e.g. selectivity index), we first subtracted the background activity from the neural response. In Figure 3, the analysis required identification of significant responses from a series of stimulus presentations. To compensate for this multiple comparison, we applied a Bonferroni correction by dividing the p-value cutoff by the number of stimulus presentations.

Receptive field analysis with flickering checkerboard

Request a detailed protocol

To measure the spatio-temporal receptive field (Figure 2B), we computed the spike-triggered average stimulus (STA) with the neural response to the flickering checkerboard (Meister et al., 1994). In many neurons that had a strong STA, we could separate the center and the surround of the receptive field by performing singular value decomposition (SVD) on the STA (Wolfe and Palmer, 1998). SVD expresses the spatio-temporal STA as a sum of terms, each of which is a product of a purely spatial and a purely temporal function. The terms are ordered by decreasing contribution to the overall variance in the STA data. We found that often the first term corresponded to the spatial and temporal profile of the center, and the second term to those of the surround.

Stimulus selectivity

Request a detailed protocol

To analyze the selectivity to a looming stimulus over other stimuli (Figure 2C,D), we computed the looming selectivity index defined as (rL-rO)/(rL+rO) with r=r-μ, where r refers to the number of spikes that a neuron fired during the first presentation of the stimulus, μ refers to the number of spikes expected during the stimulus period from the neuron’s baseline firing rate, and the subscripts L and O refer to the looming stimulus and another stimulus (e.g. contracting white disk), respectively. For the comparison to flickering checkerboard (Figure 2C), rO=rCtL-μ, where rC is the average firing rate of the neuron during flickering checkerboard and tL is the duration of the looming stimulus. Only neurons that were significantly responsive to either of the two stimuli being compared based on the Poisson significance criterion outlined above were included in the analysis.

Position invariance

Request a detailed protocol

To analyze the invariance to stimulus location (Figure 3B), we estimated the receptive field of recorded neurons from the results of the ‘random loom’ experiment in which looming stimuli appeared randomly at one of 25 locations (5 × 5 grid) in each presentation. The looming stimulus was chosen because unlike the checkerboard stimulus, it reliably drove both sSC and dSC neurons. First, we defined the function r(𝐱) that specifies the maximum response (in spikes) of a neuron to a stimulus at location 𝐱=(x1,x2). Then we (1) set to zero the responses that did not deviate significantly from baseline activity; and (2) subtracted the expected number of spikes during stimulus period due to baseline activity from r(𝐱). To capture the width of the receptive field given by the remaining responses, we computed the mean radial distance Δ=(𝐱-𝐜r(𝐱))/r(𝐱) where 𝐜=(𝐱r(𝐱))/r(𝐱) is the center of mass of the receptive field and is the Euclidean norm. Then we defined the receptive field size as 2Δ, that is twice the mean radial distance from the center of mass. Based on this method, neurons that respond to stimuli at only a single location would have a receptive field size of zero, as Δ=0. We corrected this by adding the inter-center distance between stimuli (often ~15°) to the estimated receptive field size of all neurons, as this determined the spatial resolution of our experiment.

Variability in response latency

Request a detailed protocol

In addition, we analyzed the variability of response latency during this experiment (Figure 3C). We defined the latency as the timing of the first spike during the stimulus period. We included only the neurons that met the following conditions: (i) background activity (as defined above) is less than 1; and (ii) shows statistically significant response to at least five trials in the random loom experiment. Condition (i) is required by our definition of latency. Condition (ii) is required because we define the variability of latency as the standard deviation of the timing of first spike, and this requires some number of samples to compute. 41 sSC and 128 dSC neurons that met condition (ii) but not (i) were discarded, and the final plot in Figure 3C shows 37 sSC and 70 dSC neurons. Finally, to avoid including spikes not due to visual stimulation, we required that the first spike to not occur earlier than 30ms/ since stimulus onset.

Stimulus-specific habituation

Request a detailed protocol

To analyze the stimulus-specific habituation (Figure 4B), we computed the habituation index defined as 1 – ri/r1 where ri=ri-μ refers to the number of spikes a neuron fired in the i-th repetition of the looming stimulus (ri) after subtracting the expected number of spikes due to baseline activity (μ). Analysis with i= 4, 7, and 10 did not yield significantly different results (Figure 4B uses i = 10). Only the neurons whose initial response to the looming stimuli met the significance criterion were included in the analysis.

Statistical test

Request a detailed protocol

Furthermore, we tested if the empirical distributions of sSC and dSC neurons differ significantly from each other in Figure 2C–D, Figure 3B–C, and Figure 4B. To do so we applied the two-sample Kolmogorov-Smirnov test using the MATLAB function kstest2. In all cases the computed p-values were less than the pre-set cutoff of 0.005 and were reported within the figure panels.

Recovery from habituation

Request a detailed protocol

To analyze the time to recover from the habituation (Figure 4D), a series of looming stimuli was presented at a single location with inter-stimulus intervals of 1.5, 2, 6, 11, 21, 61, and 121 s, in this order. The extent of recovery was defined as ri/r1 where ri=ri-μ refers to the number of spikes a neuron fired in the i-th repetition of this series (ri) after subtracting expected number of spikes due to baseline activity (µ). This was done for simultaneously recorded sSC and dSC neurons that met the significance criterion. The 25th, 50th (median), and 75th percentiles were then computed separately for sSC and dSC neurons and plotted in Figure 4D.

Decoding analysis

Request a detailed protocol

We analyzed the population of neurons from superficial and deep SC to decode stimulus variables in the ‘random loom’ experiment (Figure 5). Specifically, we asked if the population activity contains information about the location (i.e. in which of the 25 possible locations did the stimulus appear?) and novelty (i.e. is this the first stimulus to appear at a location?) of the stimuli.

To do so, we first pooled neurons from three recordings that used similar parameters of the ‘random loom’ experiment. Because of retionotopy in the SC, superficial SC neurons recorded by a single shank of the silicon probe tend to have overlapping receptive fields. As a result, decoding stimulus location from the superficial SC neurons requires sampling them throughout the retinotopic map, which is difficult to do experimentally. Working on the assumption that different parts of the map contain equivalent neural representations, we augmented the data by generating virtual neurons whose response profiles were spatially shifted copies of actual neural responses. Specifically, each copy shifted the response profile to one of the eight adjacent locations in a 3 × 3 grid with the original response profile in the center. The neurons were then divided into two groups (sSC and dSC) based on the depth of the channel with maximum waveform. This augmentation process increased the number of neurons used in this analysis from 106 (38 sSC and 68 dSC) to 963 (342 sSC and 621 dSC). Some neurons whose response profile after shifting lay outside the stimulus presentation area were discarded.

After this, the data consisted of neural response of the augmented sSC and dSC populations in each of the ~100 trials. In the case of the location decoder, the labels were multi-class and ranged from 1 to 25 (one for each stimulus location). In the case of the novelty decoder, the labels were binary (stimuli that were novel, that is the first to appear at a location, were 1; others were 0). The performance measure was the mean four-fold cross validation score. The chance performance for the location decoder is the maximum of the number of times the stimulus appeared at each of the 25 locations, divided by the total number of presentations (i.e. maxi{ni/jnj}, where ni refers to the number of times the stimulus appeared at location i). In our data, this was roughly 10%. The chance performance for the novelty decoder is the number of non-novel presentations divided by the total number of presentations. Given that there were 25 possible locations and 100 trials, this was roughly 75%.

We then subsampled sets of 5, 10, 30, 70, 150, 300 neurons from each of the two groups and used their responses to train the location and novelty decoders. This was done with the LogisticRegression class in the scikit-learn package (Pedregosa et al., 2011) in Python using the following parameters: penalty = ‘l2’, C = 1.0, max_iter = 5000. This process was repeated with 100 random subsamples, and the mean and standard deviation of this ensemble were computed and plotted in Figure 5.

Model

Request a detailed protocol

In the circuit of Figure 6A, we modeled each input neuron as a linear-nonlinear (LN) element. The neuron’s response was calculated as

(1) r(t)=N(g(t))

where

(2) g(t)=s(x,y,t)*k(x,y,t)=xyt=-ts(x,y,t)k(x,y,t-t)dtdydx

is the convolution of the stimulus s with the spatio-temporal receptive field k. The receptive field k(x,y,t) was parametrized as

(3) k(x,y,t)=F(x,y)T(t)
(4) F(x,y)=exp(-x2+y22σ2)
(5) T(t)=(tτ1)n1exp(-n1(t/τ1-1))-b(tτ2)n2exp(-n2(t/τ2-1))

The nonlinear transformation was a half-wave rectifier:

(6) N(g)=max(0,mgθ)

where θ is a threshold and m is a scaling factor. The firing rate of the local looming detector neuron (LD) was computed from the difference between the responses of the center and surround neurons:

(7) rLD(t)=N(rc(t)rs(t))

and the response of the widefield neuron (WF) was computed from the various local detectors as

(8) rWF(t)=iwirLD,i(t)

where wi is the synaptic weight from local neuron i onto the widefield neuron. We modeled the habituation in the synapse between local detectors and the widefield neuron with a differential equation of three parameters for short-term synaptic depression and recovery:

(9) ddtw=1wτa(wwmin)r(t)

where τ is the time constant for synaptic recovery, a is a gain factor for depression, and wmin is a floor on synaptic strength. The simulation in Figure 6F used a=1 and wmin=0.

The temporal kernels used for the center and surround neurons feeding the local looming detector were taken from the measured receptive fields of mouse alpha retinal ganglion cells (Krieger et al., 2017). Table 1 lists the parameter values chosen. We arranged local looming detectors on a grid with 15° spacing between the centers of adjacent cells.

Table 1
Parameter values used for the model in Figure 6, as defined by Equations 3, 4, 5, 6, and 9.
Receptive field (Equations 3, 4 and 5)
ParameterCenterSurround
σ4.00°10.0°
τ1104 ms84.6ms
n12.771.24
τ291.2 ms79.7 ms
n23.941.87
b1.341.33
Nonlinearity (Equation 6)
ParameterValue
m1
θ0
Synaptic depression (Equation 9)
ParameterValue
a1
wmin0

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
    Coding and processing of electrosensory information in gymnotiform fish
    1. W Heiligenberg
    (1989)
    The Journal of Experimental Biology 146:255–275.
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
    Two remarks on the visual system of the frog
    1. J Lettvin
    2. H Maturana
    3. W Pitts
    4. W McCulloch
    (1961)
    In: W. A Rosenblith, editors. Sensory Communication: Contributions to the Symposium on Principles of Sensory Communication. MIT Press. pp. 757–776.
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
    Fast and accurate spike sorting of high-channel count probes with KiloSort
    1. M Pachitariu
    2. NA Steinmetz
    3. SN Kadir
    4. M Carandini
    5. KD Harris
    (2016)
    Advances in Neural Information Processing Systems. pp. 4448–4456.
  50. 50
    The Mouse Brain in Stereotaxic Coordinates
    1. G Paxinos
    2. KB Franklin
    (2001)
    Academic Press.
  51. 51
    Scikit-learn : machine learning in Python to cite this version : scikit-learn : machine learning in Python
    1. F Pedregosa
    2. A Gramfort
    3. V Michel
    4. B Thirion
    5. O Grisel
    6. M Blondel
    7. P Prettenhofer
    8. V Dubourg
    9. F Pedregosa
    10. A Gramfort
    11. V Michel
    12. B Thirion
    13. F Pedregosa
    14. R Weiss
    (2011)
    Journal of Machine Learning Research 12:2825–2830.
  52. 52
    Neural computation in sensory systems
    1. X Pitkow
    2. M Meister
    (2014)
    In: M. S Gazzaniga, G. R Mangun, editors. The Cognitive Neurosciences (5th edition). Cambridge: MIT Press. pp. 305–318.
  53. 53
  54. 54
  55. 55
  56. 56
    The New Visual Neurosciences
    1. B Roska
    2. M Meister
    (2014)
    163–182, The retina dissects the visual scene into distinct features, The New Visual Neurosciences, Cambridge, MIT Press.
  57. 57
  58. 58
  59. 59
    Organization of neuronal subtypes and retinal inputs in the mouse superior colliculus
    1. DL Rousso
    2. M Yamataga
    3. JR Sanes
    (2014)
    Society for Neuroscience.
  60. 60
  61. 61
  62. 62
  63. 63
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
  69. 69
  70. 70
  71. 71
  72. 72
  73. 73
  74. 74
  75. 75
  76. 76
  77. 77

Decision letter

  1. Fred Rieke
    Reviewing Editor; University of Washington, United States
  2. Joshua I Gold
    Senior Editor; University of Pennsylvania, United States
  3. Fred Rieke
    Reviewer; University of Washington, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

The paper shows clear changes in coding properties between the input and deep layers of the superior colliculus. These are interesting particularly in relation to the animal's need to retain information about relevant aspects of the input stimuli while discarded others. Similar changes in coding occur in many sensory circuits, and work like that described in the paper may help guide investigation in other circuits.

Decision letter after peer review:

Thank you for submitting your article "The sifting of visual information in the superior colliculus" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Fred Rieke as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Joshua Gold as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission. While the reviewers were enthusiastic about the questions asked, several issues limited enthusiasm and prevented a complete evaluation of the work:

1) A number of key findings in the paper are supported by data from very few, in some cases single, cells. Specific examples can be found in the individual reviews appended below. The reviewers felt that they could not evaluate the significance of the work without such population analyses. This point was arrived at independently by each reviewer, and emphasized in consultation among reviewers.

2) The population analyses that are presented show substantial overlap between sSC and dSC populations. The text presents a more discrete description of this data, and as a result does not accurately convey the overlap in response properties between sSC and dSC neurons. It is important that the text accurately reflect the degree of similarity/difference in sSC and dSC responses (and overlap of population analyses such as the histograms in Figures 2 and 4).

All three reviewers agreed that adequately dealing with these points will be essential for a successful revision.

Reviewer #1:

This paper investigates how the representation of visual inputs changes as signals traverse the superior colliculus (SC). The emphasis in the paper is on how three aspects of responses to behaviorally-relevant stimuli develop: selectivity, positional invariance, and habituation. Changes in coding properties similar to those characterized here occur in other sensory circuits, and hence the exploration of these changes in the SC is likely to provide broad insight into sensory processing – a point that is made clearly in the paper. The paper is generally in good shape, but I felt that several aspects could be strengthened:

Population analyses.

Most of the main points in the paper are well supported by population analyses. A few other points would benefit from such analyses:

– spatiotemporal noise. The lack of responses to the noise stimuli in the dSC neurons compared to the sSC neurons is further evidence for an increase in selectivity. This currently is shown for single example neurons in Figure 2. Can you add population results – e.g. from something like firing rate to the noise stimulus (perhaps normalized to the response to the expanding dark spot) as a function of depth?

– SC signaling without cortex. Figure 4 shows an example of the persistence of habituation from a SC neuron recorded in a mouse lacking much of cortex. Can you analyze this in all recorded cells? Do positional invariance and selectivity for the looming stimulus persist in these mice? This is an emphasized point in the paper, so it should be explored in more detail and information about cell populations provided.

– Local looming detectors. Figure 6—figure supplement 1 shows on example of a cell with properties consistent with the local looming detectors that form the basis of the proposed model. The model would suggest that there are many such cells. Is this the only one encountered? Some discussion of the required density of these cells, and the number recorded, is important. This is particularly true because the paper argues against some alternative models based on not encountering the requisite cell types.

Description of retinal ganglion cell responses

The paper in several places argues that the recorded neurons have properties that differ from those of the retinal ganglion cell inputs to SC (e.g. subsection “Emergence of new response properties from superficial to deep layers”, last paragraph and subsection “A working model for circuit mechanisms of visual sifting”, first paragraph). It would be quite helpful to provide a summary, ideally early in the paper, of the relevant RGC response properties so that the differences noted can be evaluated/appreciated.

Model.

The model is a nice addition to the paper and provides a clear way to link the results to similar coding properties in other sensory circuits. I felt, however, that the alternative models were dismissed too strongly (e.g. Introduction, last paragraph; subsection “A working model for circuit mechanisms of visual sifting”, last paragraph; Discussion, first paragraph). For example, it would seem possible to cancel the On response of the proposed DS circuit (e.g. using signals from On DS cells). Similarly, the proposed tonically active inhibitory cell that might provide for habituation is rejected because it is not observed (but very few Local Looming Detectors appear to be observed too; see comment above). I would suggest being somewhat less emphatic in ruling these models out (but I would agree that the proposed model is the most likely of those described).

Reviewer #2:

The manuscript tells an interesting story of different visual responses to dark looming stimuli in the superficial and deep layers of the superior colliculus of the mouse. It describes that in the dSC, neurons are more selectively activated by dark looming stimuli compared to receding white stimuli, are less selective to stimulus location, and are more depressed by repeated stimulation, even at novel locations. The authors present a parsimonious three layer network model using synaptic depression to explain the location specific depression. The paper is very easy to read and presents an appealing story. Overall, the manuscript suffers somewhat from sometimes only giving anecdotal evidence and presenting differences in distributions as black-and-white differences between the sSC and dSC.

"The suppression was not permanent". Only a single example is shown. Please add some population description or statistics.

"A stimulus at one location did not suppress the subsequent response.… (Figure 4C)". To support this statement only one example neuron is shown. Please add population description and statistics.

"Remarkably the mutant showed the same long-lasting suppression.…" Again only a single example is shown. Please add quantification and statistics.

"Again, this was not observed in the data". Again, only a single example is shown, and in the image it is actually difficult to observe the ongoing baseline firing. Please add quantification to make this less anecdotal.

Reviewer #3:

The manuscript presents extracellular recordings from superficial and intermediate/deep layers of SC in awake head fixed mice able to run on a treadmill and viewing either a flickering checkerboard stimulus or one of several visual patterns (black or white looming or receding stimuli, and a moving black dot). The question is what selectivity for these visual patterns there is in SC and whether that selectivity increases in deeper layers. The question is motivated by the behavioural observations from this and other laboratories that innate defence behaviours can be elicited by some stimuli (e.g. looming black disc) but not others (e.g. receding white disc).

The fundamental observation is that most of the visual patterns seem to drive robust responses in superficial SC, but they do not drive robust responses in deeper SC, with the exception of the initial presentations of a black looming disc. In addition, responses in superficial layers are confined to small regions of visual space, but in deeper layers can be elicited from large regions of visual space. The authors show that in deeper SC the responses to black looming disc's habituate rapidly, and that this habituation is spatially specific (thus presumably in the input to these deeper SC neurons). A simple model is provided to link these observations.

The paper is easy to read and the figures present data in an accessible way. Most of the analyses presented appear appropriate and well done. However, the presented analyses do not provide enough insight and require fleshing out.

The primary claim is that visual information is sifted in the colliculus, with the deeper layers showing much more selectivity for stimuli that may be behaviourally relevant (looming black disc). However, there is so little analysis of stimulus selectivity that it is impossible to judge the veracity of the claim. The sole quantitative evidence provided is I think the comparison of responses to white receding discs and black looming discs in Figure 2C. There is a mild though significant difference in the distributions in superficial and deeper layers, but the distributions are completely overlapping. This statistical difference in one comparison is not sufficient to buttress the claim of 'sifting'. An analysis of the entire response space is needed – and if the differences remain as subtle as those in Figure 2C then there needs to be some moderation of the claims that there are strong differences in selectivity. Related to this point, response amplitude seems to be substantially reduced in the deeper layers, and one way in which selectivity could be increased is if looming black discs were more capable of driving responses throughout SC, but neurons in deeper layers have a higher threshold for activation.

The second claim is that neurons in deeper SC have much larger activation regions; this is supported by nice analyses in Figure 3, which show a progressive increase in activation area for a black looming stimulus for neurons in deeper layers. I would have liked to know whether this was specific to black looming stimuli, or also the case for other stimuli – in those neurons that responded to them, but that would require a substantial number of experiments and is not a feasible target.

The third claim is that neurons in deeper SC habituate more strongly than those in superficial SC. Figure 4B and D suggest that most neurons in both superficial and deeper layers show habituation over timescales < 10s, but that on average habituation is stronger and longer lasting in deeper layers. Perhaps the nicest result here is the spatial specificity of habituation seen in the deeper layers, which implies a 'bottom-up' desensitisation of inputs to these neurons. The decoding presented in Figure 5 lost me here – the text around the tenth paragraph of the subsection “Analysis” are not particularly informative and I am not sure how the chance levels are derived, nor how the particular choice of analysis (which includes replicating spatially shifted copies of the recorded neurons) might influence the result, nor even how many neurons in each region (superficial/deep) were in the original 3 datasets. This section needs substantial clarification.

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your article "The sifting of visual information in the superior colliculus" for consideration by eLife. Your revised article has been reviewed by three peer reviewers, including Fred Rieke as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Joshua Gold as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission. All the reviewers appreciated the revisions and felt that they made the paper considerably stronger. A few substantive issues remain – as outlined in the individual reviews below.

Reviewer #1:

This is a resubmission of a paper comparing the representation of sensory stimuli in superficial and deep layers of superior colliculus. The primary concern with the original paper was a lack of population data and corresponding statistical tests of the central conclusions. Those issues have been largely dealt with, and the paper is much stronger. I have some remaining suggestions for clarity of presentation.

Reviewer #2:

The authors responded to all my specific questions and suggestions to my satisfaction.

In the addition made in answer to one of my questions, however, it seems a small error was introduced. In the new figure panel Figure 4C bottom right with the 30 dSC cells, it is mentioned that the responses are "normalized by response to first trial of the magenta trace". This indeed seems sensible. However, the first point on the blue curve is exactly 1 (not approximately, I checked in Illustrator). Is an error made in the normalization here, or in the description, or was it so close to 1 that rounding put it exactly at 1? There are no numbers or stats given in the text, so it was not possible for me to check this point.

Reviewer #3:

The revision deals with most of the issues raised by the reviewers including myself. The paper is stronger, and presents a more compelling story. I have some remaining queries.

1) I appreciate the provision of more descriptive overviews of some of the data sets. However, the claims, in the Abstract: "neuronal responses become more selective for behaviorally relevant stimuli") and in the opening of the Discussion "an increasing selectivity for the behaviorally relevant looming stimulus over other innocuous stimuli with similar low-level features (Figure 2)") still rely on only 2 comparisons of activity (black loom vs. checkerboard, and black loom vs. receding white). Another 4 datasets are presented for the example neurons in Figure 2 but are not analysed further. I previously suggested a full analysis over this response space, which the authors argue is not feasible in a single report. But the response metric for these datasets is trivial to derive, and I cannot see why the authors cannot present (perhaps as a table in the supplementary data accompanying Figure 2) an appropriate multivariate analysis. Alternatively, the text could be modified to make clear that the statements rely on population comparisons for the two stimuli as well as a checkerboard.

2) Regardless of the what the authors think of the above, around the subsection “Stimulus selectivity”, I think the authors infer the response to the checkerboard from the mean rate across the entire presentation of the checkerboard stimulus (i.e. several minutes) while the response to the loom stimulus is the number of spikes in the first presentation (i.e. first second). If I am incorrect please clarify. If I am not then if responses habituate to the checkerboard as they do to the loom then this is not a fair comparison and a companion analysis using the same time frames for both stimuli would be appropriate.

3) The use of the word "novelty" in Figure 5 legend and in the last paragraph of the subsection “Habituation to familiar stimuli”, remains unclear to the reader and there needs to be clarity in the main text that novelty here means new location not new shape. This is a non-trivial distinction.

https://doi.org/10.7554/eLife.50678.sa1

Author response

While the reviewers were enthusiastic about the questions asked, several issues limited enthusiasm and prevented a complete evaluation of the work:

1) A number of key findings in the paper are supported by data from very few, in some cases single, cells. Specific examples can be found in the individual reviews appended below. The reviewers felt that they could not evaluate the significance of the work without such population analyses. This point was arrived at independently by each reviewer, and emphasized in consultation among reviewers.

As detailed below, we have added several such population analyses, in part based on new experiments.

2) The population analyses that are presented show substantial overlap between sSC and dSC populations. The text presents a more discrete description of this data, and as a result does not accurately convey the overlap in response properties between sSC and dSC neurons. It is important that the text accurately reflect the degree of similarity/difference in sSC and dSC responses (and overlap of population analyses such as the histograms in Figures 2 and 4).

As detailed below, we have made changes to the test to better reflect those relationships.

All three reviewers agreed that adequately dealing with these points will be essential for a successful revision.

Reviewer #1:

[…] The paper is generally in good shape, but I felt that several aspects could be strengthened:

Population analyses.

Most of the main points in the paper are well supported by population analyses. A few other points would benefit from such analyses:

– spatiotemporal noise. The lack of responses to the noise stimuli in the dSC neurons compared to the sSC neurons is further evidence for an increase in selectivity. This currently is shown for single example neurons in Figure 2. Can you add population results – e.g. from something like firing rate to the noise stimulus (perhaps normalized to the response to the expanding dark spot) as a function of depth?

As suggested, we have added a population analysis comparing the responses of 209 cells to flickering checkerboard vs. expanding dark spot in Figure 2C. dSC neurons strongly prefer the expanding spot.

– SC signaling without cortex. Figure 4 shows an example of the persistence of habituation from a SC neuron recorded in a mouse lacking much of cortex. Can you analyze this in all recorded cells? Do positional invariance and selectivity for the looming stimulus persist in these mice? This is an emphasized point in the paper, so it should be explored in more detail and information about cell populations provided.

We performed additional experiments on acortical mutant mice and added a panel to Figure 4E with population results from dSC neurons. They display persistent habituation comparable to that in the normal mouse. We focus here on the habituation instead of stimulus selectivity or position invariance because neocortex and hippocampus have been invoked as responsible for short-term memory that might underlie the habituation. Regarding stimulus selectivity, the Discussion section cites prior work (Zhao et al., 2014) suggesting that cortex plays a minor role in the tuning of neurons in the SC.

– Local looming detectors. Figure 6—figure supplement 1 shows on example of a cell with properties consistent with the local looming detectors that form the basis of the proposed model. The model would suggest that there are many such cells. Is this the only one encountered?

On further analysis we found 6 additional putative looming detectors in the database that satisfy a strict set of conditions: (i) detected in the superficial SC; (ii) selectivity index (expanding dark vs. contracting white disk) greater than 0.75; (iii) local receptive field (< = 2 adjacent looming locations); and (iv) habituation index less than 0.5. See text changes in the fifth paragraph of the subsection “A working model for circuit mechanisms of visual sifting” and Figure 6—figure supplement 1 legend.

Some discussion of the required density of these cells, and the number recorded, is important. This is particularly true because the paper argues against some alternative models based on not encountering the requisite cell types.

It is difficult to estimate the required density of these cells, as the SC is probably involved in many behaviors beyond the looming reaction. We have weakened the claims in the text about elimination of alternative models. Also, the Discussion mentions the alternative possibility that the local looming detectors might be implemented by nonlinear dendrites within a larger neuron.

Description of retinal ganglion cell responses.

The paper in several places argues that the recorded neurons have properties that differ from those of the retinal ganglion cell inputs to SC (e.g. subsection “Emergence of new response properties from superficial to deep layers”, last paragraph and subsection “A working model for circuit mechanisms of visual sifting”, first paragraph). It would be quite helpful to provide a summary, ideally early in the paper, of the relevant RGC response properties so that the differences noted can be evaluated/appreciated.

We added concrete comparisons to the relevant RGC response properties at the end of the first Results section.

Model.

The model is a nice addition to the paper and provides a clear way to link the results to similar coding properties in other sensory circuits. I felt, however, that the alternative models were dismissed too strongly (e.g. Introduction, last paragraph; subsection “A working model for circuit mechanisms of visual sifting”, last paragraph; Discussion, first paragraph). For example, it would seem possible to cancel the On response of the proposed DS circuit (e.g. using signals from On DS cells). Similarly, the proposed tonically active inhibitory cell that might provide for habituation is rejected because it is not observed (but very few Local Looming Detectors appear to be observed too; see comment above). I would suggest being somewhat less emphatic in ruling these models out (but I would agree that the proposed model is the most likely of those described).

We have softened the language about eliminating the alternative models.

Reviewer #2:

[…] Overall, the manuscript suffers somewhat from sometimes only giving anecdotal evidence and presenting differences in distributions as black-and-white differences between the sSC and dSC.

"The suppression was not permanent". Only a single example is shown. Please add some population description or statistics.

In general most or all cells recovered sensitivity to a stimulus when tested again an hour later. But we have not explored the exact time course of this recovery and simply offer Figure 4—figure supplement 1 as an example. We have amended the text to clarify this point (subsection “Habituation to familiar stimuli”, second paragraph).

"A stimulus at one location did not suppress the subsequent response.… (Figure 4C)". To support this statement only one example neuron is shown. Please add population description and statistics.

As suggested, we have added another panel to Figure 4C showing the response time course at the two locations for all 30 dSC neurons recorded in one experiment.

"Remarkably the mutant showed the same long-lasting suppression.…" Again only a single example is shown. Please add quantification and statistics.

We performed additional experiments on acortical mutant mice and added a panel to Figure 4E with population results from dSC neurons. See also response to reviewer #1 on this point.

"Again, this was not observed in the data". Again, only a single example is shown, and in the image it is actually difficult to observe the ongoing baseline firing. Please add quantification to make this less anecdotal.

We performed additional analysis of 15 dSC neurons with maintained baseline firing (>10 spikes/s), and compared the mean activity during looming stimuli to the mean activity during interstimulus intervals (rstim/risi). We found the median of this ratio was 1.28 and the 25th-75th percentiles ranged from 1.03 to 1.85; i.e. the stimulus periods had slightly more​ activity compared to the baseline. This is inconsistent with the facilitating inhibitory synapse model, which predicts that the stimulus periods would have fewer spikes after habituation. We report this result in the revised text (subsection “A working model for circuit mechanisms of visual sifting”, eighth paragraph).

Reviewer #3:

[…] The paper is easy to read and the figures present data in an accessible way. Most of the analyses presented appear appropriate and well done. However, the presented analyses do not provide enough insight and require fleshing out.

The primary claim is that visual information is sifted in the colliculus, with the deeper layers showing much more selectivity for stimuli that may be behaviourally relevant (looming black disc). However, there is so little analysis of stimulus selectivity that it is impossible to judge the veracity of the claim. The sole quantitative evidence provided is I think the comparison of responses to white receding discs and black looming discs in Figure 2C. There is a mild though significant difference in the distributions in superficial and deeper layers, but the distributions are completely overlapping. This statistical difference in one comparison is not sufficient to buttress the claim of 'sifting'. An analysis of the entire response space is needed – and if the differences remain as subtle as those in Figure 2C then there needs to be some moderation of the claims that there are strong differences in selectivity.

We have added another population analysis of stimulus selectivity in the new Figure 2C, which compares the responses to black expanding disk and flickering checkerboard. As we note in the text, Figure 2D focuses on the comparison between black expanding disk and white contracting disk because they share low-level features (moving dark edge) and yet have different ecological meanings to the animal. Thus it serves as a strong test of selectivity. “An analysis of the entire response space” is not feasible in a single report.

While it is true that the distributions of selectivity index for sSC and dSC overlap (Figure 2), this is consistent with our model, which explicitly posits the existence of local looming detectors in the sSC. Finally, we word the claim carefully: The selectivity generally increases at the population level as one goes deeper into the SC. We do not assert that neurons selective to the looming stimulus are exclusively in the dSC.

Related to this point, response amplitude seems to be substantially reduced in the deeper layers, and one way in which selectivity could be increased is if looming black discs were more capable of driving responses throughout SC, but neurons in deeper layers have a higher threshold for activation.

We agree that any mechanism of stimulus selectivity requires a nonlinear computation. Our working model for looming selectivity (Figure 6) makes use of a threshold as well.

The second claim is that neurons in deeper SC have much larger activation regions; this is supported by nice analyses in Figure 3, which show a progressive increase in activation area for a black looming stimulus for neurons in deeper layers. I would have liked to know whether this was specific to black looming stimuli, or also the case for other stimuli – in those neurons that responded to them, but that would require a substantial number of experiments and is not a feasible target.

In a few studies in which the “random loom” experiment was done with contracting white disks, we found that many dSC neurons were silent or very weakly responsive. This again demonstrates their selectivity for the looming stimuli.

The third claim is that neurons in deeper SC habituate more strongly than those in superficial SC. Figure 4B and D suggest that most neurons in both superficial and deeper layers show habituation over timescales < 10s, but that on average habituation is stronger and longer lasting in deeper layers. Perhaps the nicest result here is the spatial specificity of habituation seen in the deeper layers, which implies a 'bottom-up' desensitisation of inputs to these neurons. The decoding presented in Figure 5 lost me here – the text around the tenth paragraph of the subsection “Analysis” are not particularly informative and I am not sure how the chance levels are derived, nor how the particular choice of analysis (which includes replicating spatially shifted copies of the recorded neurons) might influence the result, nor even how many neurons in each region (superficial/deep) were in the original 3 datasets. This section needs substantial clarification.

We have expanded the text in the Materials and methods section to better describe how the chance levels are derived, the assumptions underlying data augmentation, how the data augmentation was performed, and how many neurons it added to the data set (subsection “Decoding analysis”).

[Editors' note: further revisions were suggested prior to acceptance, as described below.]

Reviewer #2:

The authors responded to all my specific questions and suggestions to my satisfaction.

In the addition made in answer to one of my questions, however, it seems a small error was introduced. In the new figure panel Figure 4C bottom right with the 30 dSC cells, it is mentioned that the responses are "normalized by response to first trial of the magenta trace". This indeed seems sensible. However, the first point on the blue curve is exactly 1 (not approximately, I checked in Illustrator). Is an error made in the normalization here, or in the description, or was it so close to 1 that rounding put it exactly at 1? There are no numbers or stats given in the text, so it was not possible for me to check this point.

The blue trace plots the median of the response ratio for 30 deeper SC neurons normalized by the response to first trial in the magenta trace, with the error bars ranging from 25th to 75th percentiles. The median of the first data point is indeed exactly 1, because this median neuron happened to fire exactly as many spikes in the first trial of the blue trace as in the first trial of the magenta trace (11 spikes). This is not an error, just a coincidence.

Reviewer #3:

The revision deals with most of the issues raised by the reviewers including myself. The paper is stronger, and presents a more compelling story. I have some remaining queries.

1) I appreciate the provision of more descriptive overviews of some of the data sets. However, the claims, in the Abstract: "neuronal responses become more selective for behaviorally relevant stimuli") and in the opening of the Discussion "an increasing selectivity for the behaviorally relevant looming stimulus over other innocuous stimuli with similar low-level features (Figure 2)") still rely on only 2 comparisons of activity (black loom vs. checkerboard, and black loom vs. receding white). Another 4 datasets are presented for the example neurons in Figure 2 but are not analysed further. I previously suggested a full analysis over this response space, which the authors argue is not feasible in a single report. But the response metric for these datasets is trivial to derive, and I cannot see why the authors cannot present (perhaps as a table in the supplementary data accompanying Figure 2) an appropriate multivariate analysis. Alternatively, the text could be modified to make clear that the statements rely on population comparisons for the two stimuli as well as a checkerboard.

[​Does the reviewer mean Figure 1?]

Previously the reviewer suggested “an analysis of the entire response space” which seemed like a tall order. We now understand that the request was more specifically for population-level data on the other figural stimuli in Figure 1C. We have added Figure 2—figure supplement 1, which shows histograms (similar to those at the bottom of Figure 2C and D) for the responses of superficial and deeper SC neurons to four other figural stimuli (expanding white disk, receding dark disk, moving disk, and dimming disk). Although there is a considerable overlap between superficial and deeper SC neurons in all cases, the probability of encountering neurons with high selectivity index for looming motion (> 0.75) is greater in the deeper SC.

2) Regardless of the what the authors think of the above, around the subsection “Stimulus selectivity”, I think the authors infer the response to the checkerboard from the mean rate across the entire presentation of the checkerboard stimulus (i.e. several minutes) while the response to the loom stimulus is the number of spikes in the first presentation (i.e. first second). If I am incorrect please clarify. If I am not then if responses habituate to the checkerboard as they do to the loom then this is not a fair comparison and a companion analysis using the same time frames for both stimuli would be appropriate.

The response to checkerboard in Figure 2C is indeed derived from the average firing rate throughout the stimulus (see Materials and methods for details), but we argue that this is a reasonable comparison. The reviewer raises the possibility that the responses to the checkerboard stimulus could habituate in deeper SC neurons, but habituation (in our work and in general) is defined as a reduction in response to the repeated presentation of a stimulus. The checkerboard stimulus does not repeat – in fact it presents a different visual feature (i.e. randomly drawn pattern of black and white squares) 60 times a second. There are no repeats in the same location, and thus no reason to expect a habituation similar to that observed in the repeated presentation of the expanding dark disk. The key result here (subsection “Selectivity for looming stimuli”, first paragraph) is that the deeper SC neurons respond very little to this whole ensemble of visual features, none of which is ethologically relevant.

3) The use of the word "novelty" in Figure 5 legend and in the last paragraph of the subsection “Habituation to familiar stimuli”, remains unclear to the reader and there needs to be clarity in the main text that novelty here means new location not new shape. This is a non-trivial distinction.

We have changed the text throughout this section to clarify that novelty refers to a novel location.

https://doi.org/10.7554/eLife.50678.sa2

Article and author information

Author details

  1. Kyu Hyun Lee

    Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, United States
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Project administration
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-6483-9444
  2. Alvita Tran

    Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, United States
    Contribution
    Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
  3. Zeynep Turan

    Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, United States
    Contribution
    Data curation, Validation, Investigation, Visualization, Methodology
    Competing interests
    No competing interests declared
  4. Markus Meister

    Division of Biology and Biological Engineering, California Institute of Technology, Pasadena, United States
    Contribution
    Conceptualization, Resources, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Project administration
    For correspondence
    meister@caltech.edu
    Competing interests
    Reviewing editor, eLife
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2136-6506

Funding

Simons Foundation (543015SPI)

  • Markus Meister

National Science Foundation (Graduate Research Fellowship)

  • Alvita Tran

National Institutes of Health (1R01NS111477)

  • Markus Meister

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Sotiris Masmanidis (UCLA) for kindly providing us with silicon probes. The Pals1flox/flox mouse strain was a gift from Seonhee Kim (Temple University) and Christopher A Walsh (Harvard Medical School). We also thank the members of Meister lab for valuable discussions and comments. This work was supported by grants to MM from the Simons Foundation (543015SPI) and the NIH (R01NS111477) and a NSF Graduate Research Fellowship to AT.

Ethics

Animal experimentation: This study was performed according to approved institutional animal care and use committee (IACUC) protocols (#1656) of Caltech. All surgery was performed under isoflurane anesthesia and every effort was made to minimize suffering.

Senior Editor

  1. Joshua I Gold, University of Pennsylvania, United States

Reviewing Editor

  1. Fred Rieke, University of Washington, United States

Reviewer

  1. Fred Rieke, University of Washington, United States

Publication history

  1. Received: July 30, 2019
  2. Accepted: March 9, 2020
  3. Accepted Manuscript published: April 14, 2020 (version 1)
  4. Version of Record published: May 19, 2020 (version 2)

Copyright

© 2020, Lee et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 2,979
    Page views
  • 619
    Downloads
  • 2
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Janna D Lendner et al.
    Research Article Updated

    Deep non-rapid eye movement sleep (NREM) and general anesthesia with propofol are prominent states of reduced arousal linked to the occurrence of synchronized oscillations in the electroencephalogram (EEG). Although rapid eye movement (REM) sleep is also associated with diminished arousal levels, it is characterized by a desynchronized, ‘wake-like’ EEG. This observation implies that reduced arousal states are not necessarily only defined by synchronous oscillatory activity. Using intracranial and surface EEG recordings in four independent data sets, we demonstrate that the 1/f spectral slope of the electrophysiological power spectrum, which reflects the non-oscillatory, scale-free component of neural activity, delineates wakefulness from propofol anesthesia, NREM and REM sleep. Critically, the spectral slope discriminates wakefulness from REM sleep solely based on the neurophysiological brain state. Taken together, our findings describe a common electrophysiological marker that tracks states of reduced arousal, including different sleep stages as well as anesthesia in humans.

    1. Neuroscience
    Özgür Genç et al.
    Research Article Updated

    We identify a set of common phenotypic modifiers that interact with five independent autism gene orthologs (RIMS1, CHD8, CHD2, WDFY3, ASH1L) causing a common failure of presynaptic homeostatic plasticity (PHP) in Drosophila. Heterozygous null mutations in each autism gene are demonstrated to have normal baseline neurotransmission and PHP. However, PHP is sensitized and rendered prone to failure. A subsequent electrophysiology-based genetic screen identifies the first known heterozygous mutations that commonly genetically interact with multiple ASD gene orthologs, causing PHP to fail. Two phenotypic modifiers identified in the screen, PDPK1 and PPP2R5D, are characterized. Finally, transcriptomic, ultrastructural and electrophysiological analyses define one mechanism by which PHP fails; an unexpected, maladaptive up-regulation of CREG, a conserved, neuronally expressed, stress response gene and a novel repressor of PHP. Thus, we define a novel genetic landscape by which diverse, unrelated autism risk genes may converge to commonly affect the robustness of synaptic transmission.