Imagine you have lost your cell phone. Your eyes scan the cluttered table in front of you, searching for its familiar blue case. But what is happening within the visual areas of your brain while you search? One possibility is that neurons that represent relevant features such as 'blue' and 'rectangular' increase their activity. This might help you spot your phone among all the other objects on the table.

Paying attention to specific features improves our performance on visual tasks that require detecting those features. The 'feature similarity gain model' proposes that this is because attention increases the activity of neurons sensitive to specific target features, such as ‘blue’ in the example above. But is this how the brain solves such challenges in practice? Previous studies examining this issue have relied on correlations. They have shown that increases in neural activity correlate with improved performance on visual tasks. But correlation does not imply causation.

Lindsay and Miller have now used a computer model of the brain’s visual pathway to examine whether changes in neural activity cause improved performance. The model was trained to use feature similarity gain to detect an object within a set of photographs. As predicted, changes in activity like those that occur in the brain did indeed improve the model’s performance. Moreover, activity changes at later stages of the model's processing pathway produced bigger improvements than activity changes earlier in the pathway. This may explain why attention affects neural activity more at later stages in the visual pathway.

But feature similarity gain is not the only possible explanation for the results. Lindsay and Miller show that another pattern of activity change also enhanced the model’s performance, and propose an experiment to distinguish between the two possibilities. Overall, these findings increase our understanding of how the brain processes sensory information. Work is ongoing to teach computers to process images as efficiently as the human visual system. The computer model used in this study is similar to those used in state-of-the-art computer vision. These findings could thus help advance artificial sensory processing too.

Covert visual attention—applied according to spatial location or visual features—has been shown repeatedly to enhance performance on challenging visual tasks (Carrasco, 2011). To explore the neural mechanisms behind this enhancement, neural responses to the same visual input are compared under different task conditions. Such experiments have identified numerous neural modulations associated with attention, including changes in firing rates, noise levels, and correlated activity (Treue, 2001; Cohen and Maunsell, 2009; Fries et al., 2001; Maunsell and Cook, 2002). But how do these neural activity changes impact performance? Previous theoretical studies have offered helpful insights on how attention may work to enhance performance (Navalpakkam and Itti, 2007; Rolls and Deco, 2006; Tsotsos et al., 1995; Cave, 1999; Hamker and Worcester, 2002; Wolfe, 1994; Hamker, 1999; Eckstein et al., 2009; Borji and Itti, 2014; Whiteley and Sahani, 2012; Bundesen, 1990; Treisman and Gelade, 1980; Verghese, 2001; Chikkerur et al., 2010). However, much of this work is either based on small, hand-designed models or lacks direct mechanistic interpretability. Here, we utilize a large-scale model of the ventral visual stream to explore the extent to which neural changes like those observed experimentally can lead to performance enhancements on realistic visual tasks. Specifically, we use a deep convolutional neural network trained to perform object classification to test effects of the feature similarity gain model of attention (Treue and Martínez Trujillo, 1999).

Deep convolutional neural networks (CNNs) are popular tools in the machine learning and computer vision communities for performing challenging visual tasks (Rawat and Wang, 2017). Their architecture—comprised of layers of convolutions, nonlinearities, and response pooling—was designed to mimic the retinotopic and hierarchical nature of the mammalian visual system (Rawat and Wang, 2017). Models of a similar form have been used to study the biological underpinnings of object recognition for decades (Fukushima, 1988; Riesenhuber and Poggio, 1999; Serre et al., 2007). Recently it has been shown that when these networks are trained to successfully perform object classification on real-world images, the intermediate representations learned are remarkably similar to those of the primate visual system, making CNNs state-of-the-art models of the ventral stream (Yamins et al., 2014; Khaligh-Razavi et al., 2017; Khaligh-Razavi and Kriegeskorte, 2014; Kheradpisheh et al., 2016; Kar et al., 2017; Cadena et al., 2017; Tripp, 2017; Love et al., 2017; Kubilius et al., 2016). A key finding has been the correspondence between different areas in the ventral stream and layers in the deep CNNs, with early convolutional layers best able to capture the representation of V1 and middle and higher layers best able to capture V4 and IT, respectively (Güçlü and van Gerven, 2015; Eickenberg et al., 2017; Yamins et al., 2014). The generalizability of these networks is limited, however, and the models are not able to match all elements of visual behaviour (Ullman et al., 2016; Azulay and Weiss, 2018; Baker et al., 2018). But given that CNNs can reach near-human performance on some visual tasks and have architectural and representational similarities to the visual system, they are well-positioned for exploring how neural correlates of attention can impact behaviour.

One popular framework to describe attention’s effects on firing rates is the feature similarity gain model (FSGM). This model, introduced by Treue and Martinez-Trujillo, claims that a neuron’s activity is multiplicatively scaled up (or down) according to how much it prefers (or doesn’t prefer) the properties of the attended stimulus (Treue and Martínez Trujillo, 1999; Martinez-Trujillo and Treue, 2004). Attention to a certain visual attribute, such as a specific orientation or color, is generally referred to as feature-based attention (FBA). FBA effects are spatially global: if a task performed at one location in the visual field activates attention to a particular feature, neurons that represent that feature across the visual field will be affected (Zhang and Luck, 2009; Saenz et al., 2002). Overall, this leads to a general shift in the representation of the neural population towards that of the attended stimulus (Çukur et al., 2013; Kaiser et al., 2016; Peelen and Kastner, 2011). Spatial attention implies that a particular portion of the visual field is being attended. According to the FSGM, spatial location is treated as an attribute like any other. Therefore, a neuron’s modulation due to attention can be predicted by how well it’s preferred features and spatial receptive field align with the features or location of the attended stimulus. The effects of combined feature and spatial attention have been found to be additive (Hayden and Gallant, 2009).

A debated issue in the attention literature is where in the visual stream attention effects can be seen. Many studies of attention focus on V4 and MT/MST (Treue, 2001), as these areas have reliable attentional effects. Some studies do find effects at earlier areas (Moro et al., 2010), though they tend to be weaker and occur later in the visual response (Kastner and Pinsk, 2004). Therefore, a leading hypothesis is that attention signals, coming from prefrontal areas (Moore and Armstrong, 2003; Monosov et al., 2011; Bichot et al., 2015; Kornblith and Tsao, 2017), target later visual areas, and the feedback connections that those areas send to earlier ones cause the weaker effects seen there later (Buffalo et al., 2010; Luck et al., 1997).

In this study, we define the FSGM of attention mathematically and implement it in a deep CNN. By applying attention at different layers in the network and for different tasks, we see how neural changes at one area propagate through the network and change performance.

The network used in this study—VGG-16, (Simonyan and Zisserman, 2014)—is shown in Figure 1A and explained in Materials and methods, 'Network Model'. Briefly, at each convolutional layer, the application of a given convolutional filter results in a feature map, which is a 2-D grid of artificial neurons that represent how well the bottom-up input at each location aligns with the filter. Each layer has multiple feature maps. Therefore a 'retinotopic’ layout is built into the structure of the network, and the same visual features are represented across that retinotopy (akin to how cells that prefer a given orientation exist at all locations across the V1 retinotopy). This network was explored in (Güçlü and van Gerven, 2015), where it was shown that early convolutional layers of this CNN are best at predicting activity of voxels in V1, while late convolutional layers are best at predicting activity of voxels in the object-selective lateral occipital area (LO).

Figure 1

Download asset Open asset

The relationship between tuning and classification

The feature similarity gain model of attention posits that neural activity is modulated by attention in proportion to how strongly a neuron prefers the attended features, as assessed by its tuning. However, the relationship between a neuron’s tuning and its ability to influence downstream readouts remains a difficult one to investigate biologically. We use our hierarchical model to explore this question. We do so by using back propagation to calculate 'gradient values', which we compare to tuning curves (see Materials and methods, 'Object category gradient calculations' and 'Tuning values' for details). Gradient values indicate the ways in which feature map activities should change in order to make the network more likely to classify an image as being of a certain object category. Tuning values represent the degree to which the feature map responds preferentially to images of a given category. If there is a correspondence between tuning and classification, a feature map that prefers a given object category (that is, responds strongly to it) should also have a high positive gradient value for that category. In Figure 2A we show gradient values and tuning curves for three example feature maps. In Figure 2C, we show the average correlation coefficients between tuning values and gradient values for all feature maps at each of the 13 convolutional layers. As can be seen, tuning curves in all layers show higher correlation with gradient values than expected by chance (as assayed by shuffled controls), but this correlation is relatively low, increasing across layers from about .2 to .5. Overall tuning quality also increases with layer depth (Figure 2B), but less strongly.

Figure 2

Download asset Open asset

Figure 2—source data 1

Object tuning curves and gradients.

https://doi.org/10.7554/eLife.38105.005

Download elife-38105-fig2-data1-v2.txt

Even at the highest layers, there can be serious discrepancies between tuning and gradient values. In Figure 2D, we show the gradient values of feature maps at the final four convolutional layers, segregated according to tuning value. In red are gradient values that correspond to tuning values greater than one (for example, category 12 for the feature map in the middle pane of Figure 2A). As these distributions show, strong tuning values can be associated with weak or even negative gradient values. Negative gradient values indicate that increasing the activity of that feature map makes the network less likely to categorize the image as the given category. Therefore, even feature maps that strongly prefer a category (and are only a few layers from the classifier) still may not be involved in its classification, or even be inversely related to it. This is aligned with a recent neural network ablation study that shows category selectivity does not predict impact on classification (Morcos et al., 2018).

Feature-based attention improves performance on challenging object classification tasks

To determine if manipulation according to tuning values can enhance performance, we created challenging visual images composed of multiple objects for the network to classify. These test images are of two types: merged (two object images transparently overlaid, such as in Serences et al., 2004) or array (four object images arranged on a grid) (see Figure 1C examples). The task for the network is to detect the presence of a given object category in these images. It does so using a series of binary classifiers trained on standard images of these objects, which replace the last layer of the network (Figure 1B). The performance of these classifiers on the test images indicates that this is a challenging task for the network (64.4% on merged images and 55.6% on array, Figure 1D. Chance is 50%), and thus a good opportunity to see the effects of attention.

We implement feature-based attention in this network by modulating the activity of units in each feature map according to how strongly the feature map prefers the attended object category (see Materials and methods, 'Tuning values' and 'How attention is applied'). A schematic of this is shown in Figure 3A. The slope of the activation function of units in a given feature map is scaled according to the tuning value of that feature map for the attended category (positive tuning values increase the slope while negative tuning values decrease it). Thus the impact of attention on activity is multiplicative and bi-directional.

Figure 3 with 2 supplements see all

Download asset Open asset

Figure 3—source data 1

Performance changes with attention.

https://doi.org/10.7554/eLife.38105.009

Download elife-38105-fig3-data1-v2.txt

The effects of attention are measured when attention is applied in this way at each layer individually (Figure 3B; solid lines) or all layers simultaneously (Figure 3—figure supplement 1A, red). For both image types (merged and array), attention enhances performance and there is a clear increase in performance enhancement as attention is applied at later layers in the network (numbering is as in Figure 1A). In particular, attention applied at the final convolutional layer performs best, leading to an 18.8% percentage point increase in binary classification on the merged images task and 22.8% increase on the array images task. Thus, FSGM-like effects can have large beneficial impacts on performance.

Attention applied at all layers simultaneously does not lead to better performance than attention applied at any individual layer (Figure 3—figure supplement 1A). We also performed a control experiment to ensure that nonspecific scaling of activity does not alone enhance performance (Figure 3—figure supplement 1C).

Some components of the FSGM are debated, for example whether attention impacts responses multiplicatively or additively (Boynton, 2009; Baruni et al., 2015; Luck et al., 1997; McAdams and Maunsell, 1999 ), and whether the activity of cells that do not prefer the attended stimulus is actually suppressed (Bridwell and Srinivasan, 2012; Navalpakkam and Itti, 2007). Comparisons of different variants of the FSGM can be seen in Figure 3—figure supplement 2. In general, multiplicative and bidirectional effects work best.

We also measure performance when attention is applied using gradient values rather than tuning values (these gradient values are calculated to maximize performance on the binary classification task, rather than classify the image as a given category; therefore technically they differ from those shown in Figure 2, however in practice they are strongly correlated. See Materials and methods, 'Object category gradient calculations' and 'Gradient values' for details). Attention applied using gradient values shows the same layer-wise trend as when using tuning values. It also reaches the same performance enhancement peak when attention is applied at the final layers. The major difference, however, comes when attention is applied at middle layers of the network. Here, attention applied according to gradient values outperforms that of tuning values.

Attention strength and the trade-off between increasing true and false positives

In the previous section, we examined the best possible effects of attention by choosing the strength for each layer and category that optimized performance. Here, we look at how performance changes as we vary the overall strength ($\beta$) of attention.

In Figure 4A we break the binary classification performance into true and false positive rates. Here, each colored line indicates a different category and increasing dot size represents increasing strength of attention. Ideally, true positives would increase without an equivalent increase (and possibly with a decrease) in false positive rates. If they increase in tandem, attention does not have a net beneficial effect. Looking at the effects of applying attention at different layers, we can see that attention at lower layers is less effective at moving the performance in this space and that movement is in somewhat random directions, although there is an average increase in performance with moderate attentional strength. With attention applied at later layers, true positive rates are more likely to increase for moderate attentional strengths, while substantial false positive rate increases occur only with higher strengths. Thus, when attention is applied with modest strength at layer 13, most categories see a substantial increase in true positives with only modest increases in false positives. As strength continues to increase however, false positives increase substantially and eventually lead to a net decrease in overall classifier performance (representing as crossing the dotted line in Figure 4A).

Figure 4 with 1 supplement see all

Download asset Open asset

Applying attention according to negated tuning values leads to a decrease in true and false positive values with increasing attention strength, which decreases overall performance (Figure 4—figure supplement 1A). This verifies that the effects of attention are not from non-specific changes in activity.

Experimentally, when switching from no or neutral attention, neurons in MT showed an average increase in activity of 7% when attending their preferred motion direction (and similar decrease when attending the non-preferred) (Martinez-Trujillo and Treue, 2004). In our model, when $\beta =.75$ (roughly the value at which performance peaks at later layers; Figure 4—figure supplement 1B), given the magnitude of the tuning values (average magnitude: .38), attention scales activity by an average of 28.5%. This value refers to how much activity is modulated in comparison to the $\beta =0$ condition, which is probably more comparable to passive or anesthetized viewing, as task engagement has been shown to scale neural responses generally (Page and Duffy, 2008). This complicates the relationship between modulation strength in our model and the values reported in the data.

To allow for a more direct comparison, in Figure 4B, we collected the true and false positive rates obtained experimentally during different object detection tasks (explained in Materials and methods, 'Experimental data'), and plotted them in comparison to the model results when attention is applied at layer 13 using tuning values (pink line) or gradient value (brown line). Five experiments (second through sixth studies) are human studies. In all of these, uncued trials are those in which no information about the upcoming visual stimulus is given, and therefore attention strength is assumed to be low. In cued trials, the to-be-detected category is cued before the presentation of a challenging visual stimulus, allowing attention to be applied to that object or category.

The majority of these experiments show a concurrent increase in both true and false positive rates as attention strength is increased. The rates in the uncued conditions (smaller dots) are generally higher than the rates produced by the $\beta =0$ condition in our model, consistent with neutrally cued conditions corresponding to ${\displaystyle \beta >0}$. We find (see Materials and methods, 'Experimental data'), that the average corresponding $\beta$ value for the neutral conditions is .37 and for the attended conditions .51. Because attention scales activity by $1+\beta f_c^{lk}$ (where $f_c^{lk}$ is the tuning value), these changes correspond to a $\approx$5% change in activity.

The first dataset included in the plot (Ori-Change; yellow line in Figure 4B) comes from a macaque change detection study (see Materials and methods, 'Experimental data' for details). Because the attention cue was only 80% valid, attention strength could be of three levels: low (for the uncued stimuli on cued trials), medium (for both stimuli on neutrally-cued trials), or high (for the cued stimuli on cued trials). Like the other studies, this study shows a concurrent increase in both true positive (correct change detection) and false positive (premature response) rates with increasing attention strength. For the model to achieve the performance changes observed between low and medium attention a roughly 12% activity change is needed, but average V4 firing rates recorded during this task show an increase of only 3.6%. This discrepancy may suggest that changes in correlations (Cohen and Maunsell, 2009) or firing rate changes in areas aside from V4 also make important contributions to observed performance changes.

Thus, according to our model, the size of experimentally observed performance changes is broadly consistent with the size of experimentally observed neural changes. While other factors are likely also relevant for performance changes, this rough alignment between the magnitude of firing rate changes and magnitude of performance changes supports the idea that the former could be a major causal factor for the latter. In addition, the fact that the model can capture this relationship provides further support for its usefulness as a model of the biology.

Finally, we show the change in true and false positive rates when the threshold of the final layer binary classifier is varied (a ‘receiver operating characteristic’ analysis, Figure 4B, gray line; no attention was applied during this analysis). Comparing this to the pink line, it is clear that varying the strength of attention applied at the final convolutional layer has more favorable performance effects than altering the classifier threshold (which corresponds to an additive effect of attention at the classifier layer). This points to the limitations that could come from attention targeting only downstream readout areas.

Overall, the model roughly matches experiments in the amount of neural modulation needed to create the observed changes in true and false positive rates. However, it is clear that the details of the experimental setup are relevant, and changes aside from firing rate and/or outside the ventral stream also likely play a role (Navalpakkam and Itti, 2007).

Feature-based attention enhances performance on orientation detection task

Some of the results presented above, particularly those related to the layer at which attention is applied, may be influenced by the fact that we are using an object categorization task. To see if results are comparable using the simpler stimuli frequently used in macaque studies, we created an orientation detection task (Figure 5A). Here, binary classifiers trained on full-field oriented gratings are tested using images that contain two gratings of different orientation and color. The performance of these binary classifiers without attention is above chance (distribution across orientations shown in inset of Figure 5A). The performance of the binary classifier associated with vertical orientation (0 degrees) was abnormally high (92% correct without attention, other orientations average 60.25%. This likely reflects the over-representation of vertical lines in the training images) and this orientation was excluded from further performance analysis.

Figure 5 with 1 supplement see all

Download asset Open asset

Figure 5—source data 1

Performance on orientation detection task.

https://doi.org/10.7554/eLife.38105.014

Download elife-38105-fig5-data1-v2.txt

Figure 5—source data 2

Performance on spatial and feature-based attention task.

https://doi.org/10.7554/eLife.38105.015

Download elife-38105-fig5-data2-v2.txt

Attention is applied according to orientation tuning values of the feature maps (tuning quality by layer is shown in Figure 5B) and tested across layers. We find (Figure 5D, solid line and Figure 3—figure supplement 1B, red) that the trend in this task is similar to that of the object task: applying attention at later layers leads to larger performance increases (14.4% percentage point increase at layer 10). This is despite the fact that orientation tuning quality peaks in the middle layers.

We also calculate the gradient values for this orientation detection task. While overall the correlations between gradient values and tuning values are lower (and even negative for early layers), the average correlation still increases with layer (Figure 5C), as with the category detection task. Importantly, while this trend in correlation exists in both detection tasks tested here, it is not a universal feature of the network or an artifact of how these values are calculated. Indeed, an opposite pattern in the correlation between orientation tuning and gradient values is shown when using attention to orientation to classify the color of a stimulus with the attended orientation (see 'Recordings show how feature similarity gain effects propagate', and Materials and methods, 'Oriented grating attention tasks' and 'Gradient values').

The results of applying attention according to gradient values is shown in Figure 5D (dashed line). Here again, using gradient value creates similar trends as using tuning values, with gradient values performing better in the middle layers.

Recordings show how feature similarity gain effects propagate

To explore how attention applied at one location in the network impacts activity later on, we apply attention at various layers and 'record' activity at others (Figure 6A, in response to full field oriented gratings). In particular, we record activity of feature maps at all layers while applying attention at layers 2, 6, 8, 10, or 12 individually.

Figure 6 with 2 supplements see all

Download asset Open asset

Figure 6—source data 1

Intercepts and slopes from gradient-applied attention.

https://doi.org/10.7554/eLife.38105.019

Download elife-38105-fig6-data1-v2.txt

Figure 6—source data 2

Intercepts and slopes from tuning curve-applied attention.

https://doi.org/10.7554/eLife.38105.020

Download elife-38105-fig6-data2-v2.txt

To understand the activity changes occurring at each layer, we use an analysis from (Martinez-Trujillo and Treue, 2004) that was designed to test for FSGM-like effects and is explained in Figure 6B. Here, the activity of a feature map in response to a given orientation when attention is applied is divided by the activity in response to the same orientation without attention. These ratios are organized according to the feature map’s orientation preference (most to least) and a line is fit to them. According to the FSGM of attention, this ratio should be greater than one for more preferred orientations and less than one for less preferred, creating a line with an intercept greater than one and negative slope.

In Figure 6C, we plot the median value of the slopes and intercepts across all feature maps at a layer, when attention is applied at different layers (indicated by color). When attention is applied directly at a layer according to its tuning values (left), FSGM effects are seen by default (intercept values are plotted in terms of how they differ from one; comparable average values from (Martinez-Trujillo and Treue, 2004) are intercept: .06 and slope: 0.0166, but note we are using $\beta =0$ for the no-attention condition in the model which, as mentioned earlier, is not necessarily the best analogue for no-attention conditions experimentally. Therefore we use these measures to show qualitative effects). As these activity changes propagate through the network, however, the FSGM effects wear off, suggesting that activating units tuned for a stimulus at one layer does not necessarily activate cells tuned for that stimulus at the next. This misalignment between tuning at one layer and the next explains why attention applied at all layers simultaneously isn’t more effective (Figure 3—figure supplement 1). In fact, applying attention to a category at one layer can actually have effects that counteract attention at a later layer (see Figure 6—figure supplement 1).

In Figure 6C (right), we show the same analysis, but while applying attention according to gradient values. The effects at the layer at which attention is applied do not look strongly like FSGM, however FSGM properties evolve as the activity changes propagate through the network, leading to clear FSGM-like effects at the final layer. Finding FSGM-like behaviour in neural data could thus be a result of FSGM effects at that area or non-FSGM effects at an earlier area (here, attention applied according to gradients which, especially at earlier layers, are not aligned with tuning).

An alternative model of the neural effects of attention—the feature matching (FM) model—suggests that the effect of attention is to amplify the activity of a neuron whenever the stimulus in its receptive field matches the attended stimulus. In Figure 6D, we calculate the fraction of feature maps at a given layer that show feature matching behaviour (defined as having activity ratios greater than one when the stimulus orientation matches the attended orientation for both preferred and anti-preferred orientations). As early as one layer post-attention, some feature maps start showing feature matching behaviour. The fact that the attention literature contains conflicting findings regarding the feature similarity gain model versus the feature matching model (Motter, 1994; Ruff and Born, 2015) may result from this finding that FSGM effects can turn into FM effects as they propagate through the network. In particular, this mechanism can explain the observations that feature matching behaviour is observed more in FEF than V4 (Zhou and Desimone, 2011) and that match information is more easily read out from perirhinal cortex than IT (Pagan et al., 2013).

We also investigated the extent to which measures of attention’s neural effects correlate with changes in performance (see Materials and methods, 'Correlating activity changes with performance'). For this we developed a new, experimentally-feasible way of calculating attention’s effects on neural activity that is inspired by the gradient-based approach to attention (that is, it focuses on classification rather than tuning). We show (Figure 6—figure supplement 2) that this new measure better correlates with performance changes than the FSGM measure of activity changes, particularly at earlier layers.

There is a simple experiment that would distinguish whether factors beyond tuning, such as gradients, play a role in guiding attention. It requires using two tasks with very different objectives (which should produce different gradients) but with the same attentional cue. An example is described in Figure 7. Here, the two tasks used would be an orientation-based color classification task (two gratings each with their own color and orientation are simultaneously shown, and the task is to report the color of the grating with the attended orientation) and an orientation detection task (report if the attended orientation is present or absent in the image). In both cases, attention is cued according to orientation. But gradient-based attention will produce different neural modulations for the two tasks, while the FSGM predicts identical modulations (Figure 7C). Thus, an experiment that recorded from the same neurons during both tasks could distinguish between tuning-based and gradient-based attention.

Figure 7

Download asset Open asset

Figure 7—source data 1

Orientation tuning curves and gradients.

https://doi.org/10.7554/eLife.38105.022

Download elife-38105-fig7-data1-v2.txt

In this work, we utilized a deep convolutional neural network (CNN) as a model of the visual system to probe the relationship between modulation of neural activity, as in attention, and performance. Specifically, we formally define the feature similarity gain model (FSGM) of attention (the basic tenets of which have been described in several experimental studies) as a multiplicative modulation of neuronal activity proportional to the neuron’s mean-subtracted feature tuning. This formalization allows us to investigate the FSGM’s ability to enhance a CNN’s performance on challenging visual tasks. We found that, across a variety of tasks, neural activity changes matching the type and magnitude of those observed experimentally can indeed lead to performance changes of the kind and magnitude observed experimentally.

We used the full observability of the model to investigate the relationship between tuning and function. We compared attention applied according to feature tuning (the FSGM) with attention designed to optimally modulate activity to improve performance (as determined by the gradient of performance with respect to the neural activity). Attention applied according to tuning does not successfully propagate from lower or middle to higher layers; that is, enhancing the activity of neurons that most prefer a given category at lower layers need not selectively enhance the activity of neurons preferring that category at higher layers. As a result, attention applied according to the FSGM performs poorly when applied at early to middle layers, while attention applied according to gradients at these layers performs better.

Attention is most effective applied at later layers (e.g., layers 9–13), where tuning and gradient values are better correlated. According to (Güçlü and van Gerven, 2015), these layers correspond most to areas V4 and LO. Such areas are known and studied for reliably showing attentional effects, whereas earlier areas such as V1 are generally not (Luck et al., 1997; Abdelhack and Kamitani, 2018). In a study involving detection of objects in natural scenes, the strength of category-specific preparatory activity in object selective cortex was correlated with performance, whereas such preparatory activity in V1 was anti-correlated with performance (Peelen and Kastner, 2011). This is in line with our finding that feature-based attention effects at earlier areas can counter the beneficial effects of that attention at later areas (Figure 6—figure supplement 1).

Our work raises the question: is attention applied simply according to tuning or is it targeted to best optimize function on a given task? We suggested a simple experiment (Figure 7) that would reveal whether non-tuning factors, such as gradients, guide attentional modulation. In (Chelazzi et al., 1998) the correlation coefficient between an index of tuning and an index of attentional modulation was .52 for a population of V4 neurons, suggesting factors other than selectivity influence attention. Furthermore, many attention studies, including that one, use only preferred and anti-preferred stimuli and therefore don’t include a thorough investigation of the relationship between tuning and attentional modulation. (Martinez-Trujillo and Treue, 2004) uses multiple stimuli to provide support for the FSGM, however the interpretation is limited by the fact that they only report population averages. (Ruff and Born, 2015) investigated the relationship between tuning strength and the strength of attentional modulation on a cell-by-cell basis. While they did find a correlation (particularly for binocular disparity tuning), it was relatively weak, which leaves room for the possibility that tuning is not the primary factor that determines attentional modulation. Local connectivity is also likely to play a role, as a correlation between normalization and attentional modulation has been shown (Ni et al., 2012).

A major challenge for understanding the biological implementation of selective attention is determining how such a precise attentional signal is carried by feedback connections. We believe that it is plausible that the visual system can learn the connections needed to carry out gradient-based attention. For example, if a high-level neuron related to the classification of an image sends a feedback connection to lower areas, an anti-Hebbian post-pre spike timing-dependent learning rule would strengthen the connection from the high level neuron to the low level one, if the lower level one causes the firing of the higher. In this way, neurons in later areas can learn to target the cells in earlier areas that caused them to fire. In contrast, it is actually more difficult to imagine how higher areas could learn the connections needed to target neurons according to their tuning, as in the FSGM. The machine learning literature on attention and learning may inspire other useful hypotheses on underlying brain mechanisms (Xu et al., 2015; Lillicrap et al., 2016).

The concept of attention has been introduced in these models previously in the machine learning literature (Mnih et al., 2014). Generally, this kind of attention relates to what would be called overt spatial attention in the neuroscience literature. That is, the attention mechanism serially selects areas of the input image for further processing, rather than modulating the activity of neurons representing those areas (as in our model of spatial attention). Other work has been done using attention to selectively process image features (Stollenga et al., 2014) and it would be interesting to compare the workings of that model to the feature-based attention used in our study.

While CNNs have representations that are similar to the ventral stream, they lack many biological details including recurrent connections, dynamics, cell types, and noisy responses. Preliminary work has shown that these elements can be incorporated into a CNN structure, and attention can enhance performance in this more biologically-realistic architecture (Lindsay et al., 2017). Furthermore, while the current work does not include neural noise independent of the stimulus, the fact that a given image is presented in many contexts (different merged images or different array images) can be thought of as a form of highly structured noise that does produce variable responses to the same image.

Another biological detail that this model lacks is 'skip connections,’ where one layer feeds into both the layer directly after it and deeper layers after that (He et al., 2016; Huang et al., 2017) as in connections from V2 to V4 or V4 to parietal areas (Ungerleider et al., 2008). Our results regarding propagation of changes through the network suggest that synaptic distance from the classifier is a relevant feature—one that is less straight forward to determine in a network with skip connections.

Because experimenters can easily control the image, defining a cell’s function in terms of how it responds to stimuli makes practical sense. However, it may be that thinking about visual areas in terms of their synaptic distance from decision-making areas such as prefrontal cortex (Heekeren et al., 2004) can be more useful for the study of attention than thinking in terms of their distance from the retina. Thus far, coarse stimulation protocols have found a relationship between the tuning of neural populations and their impact on perception (Moeller et al., 2017; DeAngelis et al., 1998; Salzman et al., 1990). However, studies of the relationship between tuning and choice probabilities suggest that a neuron’s preferred stimulus is not always an indication of its causal role in classification (Zaidel et al., 2017; Purushothaman and Bradley, 2005), though see (Katz et al., 2016). Targeted stimulation protocols and a more fine-grained ability to determine both upstream drivers of, and downstream responses driven by, stimulated neurons will be needed to better address these issues.

The weights for the model used are linked to in the study. The data resulting from simulations have been packaged and are available on Dryad (doi:10.5061/dryad.jc14081). The analysis code are available on GitHub (https://github.com/gwl2108/CNN_attention; copy archived at https://github.com/elifesciences-publications/CNN_attention).

1. 1. Lindsay G
   2. Miller K

   (2018) Available at Dryad Digital Repository under a CC0 Public Domain Dedication

   Data from: How biological attention mechanisms improve task performance in a large-scale visual system model.

   https://doi.org/10.5061/dryad.jc14081

1. 1. Abdelhack M
   2. Kamitani Y

   (2018) Sharpening of hierarchical visual feature representations of blurred images

   Eneuro 5:ENEURO.0443-17.2018.

   https://doi.org/10.1523/ENEURO.0443-17.2018

   • PubMed
   • Google Scholar
2. Preprint

   1. Azulay A
   2. Weiss Y

   (2018)  Why do deep convolutional networks generalize so poorly to small image transformations? 

   arXiv.

   https://arxiv.org/abs/1805.12177

   • Google Scholar
3. 1. Baker N
   2. Lu H
   3. Erlikhman G
   4. Kellman P

   (2018) Deep convolutional networks do not make classifications based on global object shape

   Journal of Vision 18:904.

   https://doi.org/10.1167/18.10.904

   • Google Scholar
4. 1. Bang JW
   2. Rahnev D

   (2017) Stimulus expectation alters decision criterion but not sensory signal in perceptual decision making

   Scientific Reports 7:17072.

   https://doi.org/10.1038/s41598-017-16885-2

   • PubMed
   • Google Scholar
5. 1. Baruni JK
   2. Lau B
   3. Salzman CD

   (2015) Reward expectation differentially modulates attentional behavior and activity in visual area V4

   Nature Neuroscience 18:1656–1663.

   https://doi.org/10.1038/nn.4141

   • PubMed
   • Google Scholar
6. 1. Bichot NP
   2. Heard MT
   3. DeGennaro EM
   4. Desimone R

   (2015) A source for Feature-Based attention in the prefrontal cortex

   Neuron 88:832–844.

   https://doi.org/10.1016/j.neuron.2015.10.001

   • PubMed
   • Google Scholar
7. 1. Borji A
   2. Itti L

   (2014) Optimal attentional modulation of a neural population

   Frontiers in Computational Neuroscience 8:34.

   https://doi.org/10.3389/fncom.2014.00034

   • PubMed
   • Google Scholar
8. 1. Boynton GM

   (2009) A framework for describing the effects of attention on visual responses

   Vision Research 49:1129–1143.

   https://doi.org/10.1016/j.visres.2008.11.001

   • PubMed
   • Google Scholar
9. 1. Bridwell DA
   2. Srinivasan R

   (2012) Distinct attention networks for feature enhancement and suppression in vision

   Psychological Science 23:1151–1158.

   https://doi.org/10.1177/0956797612440099

   • PubMed
   • Google Scholar
10. 1. Buffalo EA
    2. Fries P
    3. Landman R
    4. Liang H
    5. Desimone R

    (2010) A backward progression of attentional effects in the ventral stream

    PNAS 107:361–365.

    https://doi.org/10.1073/pnas.0907658106

    • PubMed
    • Google Scholar
11. 1. Bundesen C

    (1990) A theory of visual attention

    Psychological Review 97:523–547.

    https://doi.org/10.1037/0033-295X.97.4.523

    • PubMed
    • Google Scholar
12. Preprint

    1. Cadena SA
    2. Denfield GH
    3. Walker EY
    4. Gatys LA
    5. Tolias AS
    6. Bethge M
    7. Ecker AS

    (2017) Deep convolutional models improve predictions of macaque v1 responses to natural images

    bioRxiv.

    https://doi.org/10.1101/201764

    • Google Scholar
13. 1. Carrasco M

    (2011) Visual attention: the past 25 years

    Vision Research 51:1484–1525.

    https://doi.org/10.1016/j.visres.2011.04.012

    • PubMed
    • Google Scholar
14. 1. Cave KR

    (1999) The FeatureGate model of visual selection

    Psychological Research 62:182–194.

    https://doi.org/10.1007/s004260050050

    • PubMed
    • Google Scholar
15. 1. Chelazzi L
    2. Duncan J
    3. Miller EK
    4. Desimone R

    (1998) Responses of neurons in inferior temporal cortex during memory-guided visual search

    Journal of Neurophysiology 80:2918–2940.

    https://doi.org/10.1152/jn.1998.80.6.2918

    • PubMed
    • Google Scholar
16. 1. Chikkerur S
    2. Serre T
    3. Tan C
    4. Poggio T

    (2010) What and where: a bayesian inference theory of attention

    Vision Research 50:2233–2247.

    https://doi.org/10.1016/j.visres.2010.05.013

    • PubMed
    • Google Scholar
17. 1. Cohen MR
    2. Maunsell JH

    (2009) Attention improves performance primarily by reducing interneuronal correlations

    Nature Neuroscience 12:1594–1600.

    https://doi.org/10.1038/nn.2439

    • PubMed
    • Google Scholar
18. 1. Cohen MR
    2. Maunsell JH

    (2011) Using neuronal populations to study the mechanisms underlying spatial and feature attention

    Neuron 70:1192–1204.

    https://doi.org/10.1016/j.neuron.2011.04.029

    • PubMed
    • Google Scholar
19. 1. Crapse TB
    2. Lau H
    3. Basso MA

    (2018) A role for the superior colliculus in decision criteria

    Neuron 97:181–194.

    https://doi.org/10.1016/j.neuron.2017.12.006

    • PubMed
    • Google Scholar
20. 1. Çukur T
    2. Nishimoto S
    3. Huth AG
    4. Gallant JL

    (2013) Attention during natural vision warps semantic representation across the human brain

    Nature Neuroscience 16:763–770.

    https://doi.org/10.1038/nn.3381

    • PubMed
    • Google Scholar
21. 1. DeAngelis GC
    2. Cumming BG
    3. Newsome WT

    (1998) Cortical area MT and the perception of stereoscopic depth

    Nature 394:677–680.

    https://doi.org/10.1038/29299

    • PubMed
    • Google Scholar
22. 1. Downing CJ

    (1988) Expectancy and visual-spatial attention: effects on perceptual quality

    Journal of Experimental Psychology: Human Perception and Performance 14:188–202.

    https://doi.org/10.1037/0096-1523.14.2.188

    • Google Scholar
23. 1. Eckstein MP
    2. Peterson MF
    3. Pham BT
    4. Droll JA

    (2009) Statistical decision theory to relate neurons to behavior in the study of covert visual attention

    Vision Research 49:1097–1128.

    https://doi.org/10.1016/j.visres.2008.12.008

    • PubMed
    • Google Scholar
24. 1. Eickenberg M
    2. Gramfort A
    3. Varoquaux G
    4. Thirion B

    (2017) Seeing it all: convolutional network layers map the function of the human visual system

    NeuroImage 152:184–194.

    https://doi.org/10.1016/j.neuroimage.2016.10.001

    • PubMed
    • Google Scholar
25. 1. Fries P
    2. Reynolds JH
    3. Rorie AE
    4. Desimone R

    (2001) Modulation of oscillatory neuronal synchronization by selective visual attention

    Science 291:1560–1563.

    https://doi.org/10.1126/science.1055465

    • PubMed
    • Google Scholar
26. Website

    1. Frossard D

    (2017) VGG in TensorFlow

    Accessed March 1, 2017.

    https://www.cs.toronto.edu/ frossard/post/vgg16
27. 1. Fukushima K

    (1988) Neocognitron: a hierarchical neural network capable of visual pattern recognition

    Neural Networks 1:119–130.

    https://doi.org/10.1016/0893-6080(88)90014-7

    • Google Scholar
28. 1. Güçlü U
    2. van Gerven MA

    (2015) Deep neural networks reveal a gradient in the complexity of neural representations across the ventral stream

    Journal of Neuroscience 35:10005–10014.

    https://doi.org/10.1523/JNEUROSCI.5023-14.2015

    • PubMed
    • Google Scholar
29. 1. Hamker FH

    (1999)

    Connectionist Models in Cognitive Neuroscience

    252–261, The Role of Feedback Connections in Task-Driven Visual Search, Connectionist Models in Cognitive Neuroscience, Springer, 10.1007/978-1-4471-0813-9_22.

    • Google Scholar
30. 1. Hamker FH
    2. Worcester J

    (2002)

    International Workshop on Biologically Motivated Computer Vision

    398, Object Detection in Natural Scenes by Feedback, International Workshop on Biologically Motivated Computer Vision, 407, Springer.

    • Google Scholar
31. 1. Hawkins HL
    2. Hillyard SA
    3. Luck SJ
    4. Mouloua M
    5. Downing CJ
    6. Woodward DP

    (1990) Visual attention modulates signal detectability

    Journal of Experimental Psychology: Human Perception and Performance 16:802–811.

    https://doi.org/10.1037/0096-1523.16.4.802

    • Google Scholar
32. 1. Hayden BY
    2. Gallant JL

    (2009) Combined effects of spatial and feature-based attention on responses of V4 neurons

    Vision Research 49:1182–1187.

    https://doi.org/10.1016/j.visres.2008.06.011

    • PubMed
    • Google Scholar
33. Conference

    1. He K
    2. Zhang X
    3. Ren S
    4. Sun J

    (2016)

    Deep residual learning for image recognition

    IEEE Conference on Computer Vision and Pattern Recognition. pp. 770–778.

    • Google Scholar
34. 1. Heekeren HR
    2. Marrett S
    3. Bandettini PA
    4. Ungerleider LG

    (2004) A general mechanism for perceptual decision-making in the human brain

    Nature 431:859–862.

    https://doi.org/10.1038/nature02966

    • PubMed
    • Google Scholar
35. Conference

    1. Huang G
    2. Liu Z
    3. van der Maaten L
    4. Weinberger KQ

    (2017) Densely connected convolutional networks

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.

    https://doi.org/10.1109/CVPR.2017.243

    • Google Scholar
36. 1. Kaiser D
    2. Oosterhof NN
    3. Peelen MV

    (2016) The neural dynamics of attentional selection in natural scenes

    Journal of Neuroscience 36:10522–10528.

    https://doi.org/10.1523/JNEUROSCI.1385-16.2016

    • PubMed
    • Google Scholar
37. Conference

    1. Kar K
    2. Kubilius J
    3. Issa E
    4. Schmidt K
    5. DiCarlo J

    (2017)

    Evidence that feedback is required for object identity inferences computed by the ventral stream

    Computational and Systems Neuroscience (Cosyne).

    • Google Scholar
38. 1. Kastner S
    2. Pinsk MA

    (2004) Visual attention as a multilevel selection process

    Cognitive, Affective, & Behavioral Neuroscience 4:483–500.

    https://doi.org/10.3758/CABN.4.4.483

    • PubMed
    • Google Scholar
39. 1. Katz LN
    2. Yates JL
    3. Pillow JW
    4. Huk AC

    (2016) Dissociated functional significance of decision-related activity in the primate dorsal stream

    Nature 535:285–288.

    https://doi.org/10.1038/nature18617

    • PubMed
    • Google Scholar
40. 1. Khaligh-Razavi SM
    2. Henriksson L
    3. Kay K
    4. Kriegeskorte N

    (2017) Fixed versus mixed RSA: Explaining visual representations by fixed and mixed feature sets from shallow and deep computational models

    Journal of Mathematical Psychology 76:184–197.

    https://doi.org/10.1016/j.jmp.2016.10.007

    • PubMed
    • Google Scholar
41. 1. Khaligh-Razavi SM
    2. Kriegeskorte N

    (2014) Deep supervised, but not unsupervised, models may explain IT cortical representation

    PLoS Computational Biology 10:e1003915.

    https://doi.org/10.1371/journal.pcbi.1003915

    • PubMed
    • Google Scholar
42. 1. Kheradpisheh SR
    2. Ghodrati M
    3. Ganjtabesh M
    4. Masquelier T

    (2016) Deep networks can resemble human Feed-forward vision in invariant object recognition

    Scientific Reports 6:32672.

    https://doi.org/10.1038/srep32672

    • PubMed
    • Google Scholar
43. 1. Koivisto M
    2. Kahila E

    (2017) Top-down preparation modulates visual categorization but not subjective awareness of objects presented in natural backgrounds

    Vision Research 133:73–80.

    https://doi.org/10.1016/j.visres.2017.01.005

    • PubMed
    • Google Scholar
44. 1. Kornblith S
    2. Tsao DY

    (2017) How thoughts arise from sights: inferotemporal and prefrontal contributions to vision

    Current Opinion in Neurobiology 46:208–218.

    https://doi.org/10.1016/j.conb.2017.08.016

    • PubMed
    • Google Scholar
45. 1. Krauzlis RJ
    2. Lovejoy LP
    3. Zénon A

    (2013) Superior colliculus and visual spatial attention

    Annual Review of Neuroscience 36:165–182.

    https://doi.org/10.1146/annurev-neuro-062012-170249

    • PubMed
    • Google Scholar
46. 1. Kubilius J
    2. Bracci S
    3. Op de Beeck HP

    (2016) Deep neural networks as a computational model for human shape sensitivity

    PLoS Computational Biology 12:e1004896.

    https://doi.org/10.1371/journal.pcbi.1004896

    • PubMed
    • Google Scholar
47. 1. Lillicrap TP
    2. Cownden D
    3. Tweed DB
    4. Akerman CJ

    (2016) Random synaptic feedback weights support error backpropagation for deep learning

    Nature Communications 7:13276.

    https://doi.org/10.1038/ncomms13276

    • PubMed
    • Google Scholar
48. Preprint

    1. Lindsay GW

    (2015) Feature-based attention in convolutional neural networks

    arXiv.

    https://arxiv.org/abs/1511.06408

    • Google Scholar
49. Conference

    1. Lindsay GW
    2. Rubin DB
    3. Miller KD

    (2017)

    The stabilized supralinear network replicates neural and performance correlates of attention

    Computational and Systems Neuroscience (Cosyne).

    • Google Scholar
50. Conference

    1. Love BC
    2. Guest O
    3. Slomka P
    4. Navarro VM
    5. Wasserman E

    (2017)

    Deep networks as models of human and animal categorization

    CogSci 2018.

    • Google Scholar
51. 1. Luck SJ
    2. Chelazzi L
    3. Hillyard SA
    4. Desimone R

    (1997) Neural mechanisms of spatial selective attention in Areas V1, V2, and V4 of macaque visual cortex

    Journal of Neurophysiology 77:24–42.

    https://doi.org/10.1152/jn.1997.77.1.24

    • PubMed
    • Google Scholar
52. 1. Luo TZ
    2. Maunsell JH

    (2015) Neuronal modulations in visual cortex are associated with only one of multiple components of attention

    Neuron 86:1182–1188.

    https://doi.org/10.1016/j.neuron.2015.05.007

    • PubMed
    • Google Scholar
53. 1. Lupyan G
    2. Spivey MJ

    (2010) Making the invisible visible: verbal but not visual cues enhance visual detection

    PLoS ONE 5:e11452.

    https://doi.org/10.1371/journal.pone.0011452

    • PubMed
    • Google Scholar
54. 1. Lupyan G
    2. Ward EJ

    (2013) Language can boost otherwise unseen objects into visual awareness

    PNAS 110:14196–14201.

    https://doi.org/10.1073/pnas.1303312110

    • PubMed
    • Google Scholar
55. 1. Martinez-Trujillo JC
    2. Treue S

    (2004) Feature-based attention increases the selectivity of population responses in primate visual cortex

    Current Biology 14:744–751.

    https://doi.org/10.1016/j.cub.2004.04.028

    • PubMed
    • Google Scholar
56. 1. Maunsell JHR
    2. Cook EP

    (2002) The role of attention in visual processing

    Philosophical Transactions of the Royal Society B: Biological Sciences 357:1063–1072.

    https://doi.org/10.1098/rstb.2002.1107

    • Google Scholar
57. 1. Mayo JP
    2. Cohen MR
    3. Maunsell JH

    (2015) A refined neuronal population measure of visual attention

    PLoS One 10:e0136570.

    https://doi.org/10.1371/journal.pone.0136570

    • PubMed
    • Google Scholar
58. 1. Mayo JP
    2. Maunsell JH

    (2016) Graded neuronal modulations related to visual spatial attention

    The Journal of Neuroscience 36:5353–5361.

    https://doi.org/10.1523/JNEUROSCI.0192-16.2016

    • PubMed
    • Google Scholar
59. 1. McAdams CJ
    2. Maunsell JH

    (1999) Effects of attention on orientation-tuning functions of single neurons in macaque cortical area V4

    The Journal of Neuroscience 19:431–441.

    https://doi.org/10.1523/JNEUROSCI.19-01-00431.1999

    • PubMed
    • Google Scholar
60. 1. Mnih V
    2. Heess N
    3. Graves A

    (2014)

    Advances in Neural Information Processing Systems

    2204–2212, Recurrent Models of Visual Attention, Advances in Neural Information Processing Systems, MIT Press.

    • Google Scholar
61. 1. Moeller S
    2. Crapse T
    3. Chang L
    4. Tsao DY

    (2017) The effect of face patch microstimulation on perception of faces and objects

    Nature Neuroscience 20:743–752.

    https://doi.org/10.1038/nn.4527

    • PubMed
    • Google Scholar
62. 1. Monosov IE
    2. Sheinberg DL
    3. Thompson KG

    (2011) The effects of prefrontal cortex inactivation on object responses of single neurons in the inferotemporal cortex during visual search

    Journal of Neuroscience 31:15956–15961.

    https://doi.org/10.1523/JNEUROSCI.2995-11.2011

    • PubMed
    • Google Scholar
63. 1. Moore T
    2. Armstrong KM

    (2003) Selective gating of visual signals by microstimulation of frontal cortex

    Nature 421:370–373.

    https://doi.org/10.1038/nature01341

    • PubMed
    • Google Scholar
64. Preprint

    1. Morcos AS
    2. Barrett DGT
    3. Rabinowitz NC
    4. Botvinick M

    (2018) On the importance of single directions for generalization

    arXiv .

    https://arxiv.org/abs/1803.06959

    • Google Scholar
65. 1. Moro SI
    2. Tolboom M
    3. Khayat PS
    4. Roelfsema PR

    (2010) Neuronal activity in the visual cortex reveals the temporal order of cognitive operations

    Journal of Neuroscience 30:16293–16303.

    https://doi.org/10.1523/JNEUROSCI.1256-10.2010

    • PubMed
    • Google Scholar
66. 1. Motter BC

    (1994) Neural correlates of feature selective memory and pop-out in extrastriate area V4

    The Journal of Neuroscience 14:2190–2199.

    https://doi.org/10.1523/JNEUROSCI.14-04-02190.1994

    • PubMed
    • Google Scholar
67. 1. Navalpakkam V
    2. Itti L

    (2007) Search goal tunes visual features optimally

    Neuron 53:605–617.

    https://doi.org/10.1016/j.neuron.2007.01.018

    • PubMed
    • Google Scholar
68. 1. Ni AM
    2. Ray S
    3. Maunsell JH

    (2012) Tuned normalization explains the size of attention modulations

    Neuron 73:803–813.

    https://doi.org/10.1016/j.neuron.2012.01.006

    • PubMed
    • Google Scholar
69. 1. Pagan M
    2. Urban LS
    3. Wohl MP
    4. Rust NC

    (2013) Signals in inferotemporal and perirhinal cortex suggest an untangling of visual target information

    Nature Neuroscience 16:1132–1139.

    https://doi.org/10.1038/nn.3433

    • PubMed
    • Google Scholar
70. 1. Page WK
    2. Duffy CJ

    (2008) Cortical neuronal responses to optic flow are shaped by visual strategies for steering

    Cerebral Cortex 18:727–739.

    https://doi.org/10.1093/cercor/bhm109

    • PubMed
    • Google Scholar
71. 1. Peelen MV
    2. Fei-Fei L
    3. Kastner S

    (2009) Neural mechanisms of rapid natural scene categorization in human visual cortex

    Nature 460:94–97.

    https://doi.org/10.1038/nature08103

    • PubMed
    • Google Scholar
72. 1. Peelen MV
    2. Kastner S

    (2011) A neural basis for real-world visual search in human occipitotemporal cortex

    PNAS 108:12125–12130.

    https://doi.org/10.1073/pnas.1101042108

    • PubMed
    • Google Scholar
73. 1. Purushothaman G
    2. Bradley DC

    (2005) Neural population code for fine perceptual decisions in area MT

    Nature Neuroscience 8:99–106.

    https://doi.org/10.1038/nn1373

    • PubMed
    • Google Scholar
74. 1. Rahnev D
    2. Lau H
    3. de Lange FP

    (2011) Prior expectation modulates the interaction between sensory and prefrontal regions in the human brain

    Journal of Neuroscience 31:10741–10748.

    https://doi.org/10.1523/JNEUROSCI.1478-11.2011

    • PubMed
    • Google Scholar
75. 1. Rawat W
    2. Wang Z

    (2017) Deep convolutional neural networks for image classification: a comprehensive review

    Neural Computation 29:2352–2449.

    https://doi.org/10.1162/neco_a_00990

    • PubMed
    • Google Scholar
76. 1. Riesenhuber M
    2. Poggio T

    (1999) Hierarchical models of object recognition in cortex

    Nature Neuroscience 2:1019–1025.

    https://doi.org/10.1038/14819

    • PubMed
    • Google Scholar
77. 1. Rolls ET
    2. Deco G

    (2006) Attention in natural scenes: neurophysiological and computational bases

    Neural Networks 19:1383–1394.

    https://doi.org/10.1016/j.neunet.2006.08.007

    • PubMed
    • Google Scholar
78. 1. Ruff DA
    2. Born RT

    (2015) Feature attention for binocular disparity in primate area MT depends on tuning strength

    Journal of Neurophysiology 113:1545–1555.

    https://doi.org/10.1152/jn.00772.2014

    • PubMed
    • Google Scholar
79. 1. Saenz M
    2. Buracas GT
    3. Boynton GM

    (2002) Global effects of feature-based attention in human visual cortex

    Nature Neuroscience 5:631–632.

    https://doi.org/10.1038/nn876

    • PubMed
    • Google Scholar
80. 1. Sàenz M
    2. Buraĉas GT
    3. Boynton GM

    (2003) Global feature-based attention for motion and color

    Vision Research 43:629–637.

    https://doi.org/10.1016/S0042-6989(02)00595-3

    • PubMed
    • Google Scholar
81. 1. Salzman CD
    2. Britten KH
    3. Newsome WT

    (1990) Cortical microstimulation influences perceptual judgements of motion direction

    Nature 346:174–177.

    https://doi.org/10.1038/346174a0

    • PubMed
    • Google Scholar
82. 1. Seeliger K
    2. Fritsche M
    3. Güçlü U
    4. Schoenmakers S
    5. Schoffelen J-M
    6. Bosch SE
    7. van Gerven MAJ

    (2017) Cnn-based encoding and decoding of visual object recognition in space and time

    bioRxiv.

    https://doi.org/10.1101/118091

    • Google Scholar
83. 1. Serences JT
    2. Schwarzbach J
    3. Courtney SM
    4. Golay X
    5. Yantis S

    (2004) Control of object-based attention in human cortex

    Cerebral Cortex 14:1346–1357.

    https://doi.org/10.1093/cercor/bhh095

    • PubMed
    • Google Scholar
84. 1. Serre T
    2. Wolf L
    3. Bileschi S
    4. Riesenhuber M
    5. Poggio T

    (2007) Robust object recognition with cortex-like mechanisms

    IEEE Transactions on Pattern Analysis and Machine Intelligence 29:411–426.

    https://doi.org/10.1109/TPAMI.2007.56

    • PubMed
    • Google Scholar
85. Preprint

    1. Simonyan K
    2. Zisserman A

    (2014) Very deep convolutional networks for large-scale image recognition.

    arXiv.

    https://arxiv.org/abs/1409.1556

    • Google Scholar
86. 1. Sridharan D
    2. Steinmetz NA
    3. Moore T
    4. Knudsen EI

    (2017) Does the superior colliculus control perceptual sensitivity or choice Bias during attention? evidence from a multialternative decision framework

    The Journal of Neuroscience 37:480–511.

    https://doi.org/10.1523/JNEUROSCI.4505-14.2017

    • PubMed
    • Google Scholar
87. 1. Stein T
    2. Peelen MV

    (2015) Content-specific expectations enhance stimulus detectability by increasing perceptual sensitivity

    Journal of Experimental Psychology: General 144:1089–1104.

    https://doi.org/10.1037/xge0000109

    • PubMed
    • Google Scholar
88. 1. Stein T
    2. Peelen MV

    (2017) Object detection in natural scenes: independent effects of spatial and category-based attention

    Attention, Perception, & Psychophysics 79:738–752.

    https://doi.org/10.3758/s13414-017-1279-8

    • PubMed
    • Google Scholar
89. 1. Stollenga MF
    2. Masci J
    3. Gomez F
    4. Schmidhuber J

    (2014)

    Advances in Neural Information Processing Systems

    3545–3553, Deep networks with internal selective attention through feedback connections, Advances in Neural Information Processing Systems, MIT Press.

    • Google Scholar
90. 1. Treisman AM
    2. Gelade G

    (1980) A feature-integration theory of attention

    Cognitive Psychology 12:97–136.

    https://doi.org/10.1016/0010-0285(80)90005-5

    • PubMed
    • Google Scholar
91. 1. Treue S
    2. Martínez Trujillo JC

    (1999) Feature-based attention influences motion processing gain in macaque visual cortex

    Nature 399:575–579.

    https://doi.org/10.1038/21176

    • PubMed
    • Google Scholar
92. 1. Treue S

    (2001) Neural correlates of attention in primate visual cortex

    Trends in Neurosciences 24:295–300.

    https://doi.org/10.1016/S0166-2236(00)01814-2

    • PubMed
    • Google Scholar
93. Conference

    1. Tripp BP

    (2017)

    Similarities and differences between stimulus tuning in theinferotemporal visual cortex and convolutional networks

    Neural Networks (IJCNN), 2017 International Joint Conference. pp. 3551–3560.

    • Google Scholar
94. 1. Tsotsos JK
    2. Culhane SM
    3. Kei Wai WY
    4. Lai Y
    5. Davis N
    6. Nuflo F

    (1995) Modeling visual attention via selective tuning

    Artificial Intelligence 78:507–545.

    https://doi.org/10.1016/0004-3702(95)00025-9

    • Google Scholar
95. 1. Ullman S
    2. Assif L
    3. Fetaya E
    4. Harari D

    (2016) Atoms of recognition in human and computer vision

    PNAS 113:2744–2749.

    https://doi.org/10.1073/pnas.1513198113

    • PubMed
    • Google Scholar
96. 1. Ungerleider LG
    2. Galkin TW
    3. Desimone R
    4. Gattass R

    (2008) Cortical connections of area V4 in the macaque

    Cerebral Cortex 18:477–499.

    https://doi.org/10.1093/cercor/bhm061

    • PubMed
    • Google Scholar
97. 1. Verghese P

    (2001) Visual search and attention: a signal detection theory approach

    Neuron 31:523–535.

    https://doi.org/10.1016/S0896-6273(01)00392-0

    • PubMed
    • Google Scholar
98. 1. Whiteley L
    2. Sahani M

    (2012) Attention in a bayesian framework

    Frontiers in Human Neuroscience 6:100.

    https://doi.org/10.3389/fnhum.2012.00100

    • PubMed
    • Google Scholar
99. 1. Wolfe JM

    (1994) Guided search 2.0 A revised model of visual search

    Psychonomic Bulletin & Review 1:202–238.

    https://doi.org/10.3758/BF03200774

    • PubMed
    • Google Scholar
100. Conference

     1. Xu K
     2. Ba J
     3. Kiros R
     4. Cho K
     5. Courville A
     6. Salakhudinov R
     7. Zemel R
     8. Bengio Y

     (2015)

     Show, attend and tell: neural image caption generation with visual attention

     International Conference on Machine Learning. pp. 2048–2057.

     • Google Scholar
101. 1. Yamins DL
     2. Hong H
     3. Cadieu CF
     4. Solomon EA
     5. Seibert D
     6. DiCarlo JJ

     (2014) Performance-optimized hierarchical models predict neural responses in higher visual cortex

     PNAS 111:8619–8624.

     https://doi.org/10.1073/pnas.1403112111

     • PubMed
     • Google Scholar
102. 1. Zaidel A
     2. DeAngelis GC
     3. Angelaki DE

     (2017) Decoupled choice-driven and stimulus-related activity in parietal neurons may be misrepresented by choice probabilities

     Nature Communications 8:3.

     https://doi.org/10.1038/s41467-017-00766-3

     • PubMed
     • Google Scholar
103. 1. Zhang W
     2. Luck SJ

     (2009) Feature-based attention modulates feedforward visual processing

     Nature Neuroscience 12:24–25.

     https://doi.org/10.1038/nn.2223

     • PubMed
     • Google Scholar
104. 1. Zhang Y
     2. Meyers EM
     3. Bichot NP
     4. Serre T
     5. Poggio TA
     6. Desimone R

     (2011) Object decoding with attention in inferior temporal cortex

     PNAS 108:8850–8855.

     https://doi.org/10.1073/pnas.1100999108

     • PubMed
     • Google Scholar
105. 1. Zhou H
     2. Desimone R

     (2011) Feature-based attention in the frontal eye field and area V4 during visual search

     Neuron 70:1205–1217.

     https://doi.org/10.1016/j.neuron.2011.04.032

     • PubMed
     • Google Scholar

Thank you for submitting your article "How biological attention mechanisms improve task performance in a large-scale visual system model" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Marcel van Gerven as the Reviewing Editor, and the evaluation has been overseen by Sabine Kastner as the Senior Editor. The following individual involved in review of your submission has agreed to reveal his identity: Marius Peelen (Reviewer #1).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

The authors present an approach to examining the feature similarity model of attention by incorporating attentional modulation in a convolutional neural network for object categorization. They find that task performance enhancements with attention can roughly approximate those found experimentally, but most interestingly, only when attention is applied to the later layers of the network. The authors demonstrate a dissociation between layers in the strength of tuning and the performance enhancement achieved by applying attention. This represents an interesting contribution to theories of attention to the extent that one considers CNNs a useful model for biological vision. Another key contribution is the distinction between gradient-based and tuning-based feedback. The findings obtained here in neural networks make several new predictions for biological experiments and raise fundamental questions about how tuning affects behavior.

Essential revisions:

Introduction, second paragraph: CNNs are introduced as great models of the ventral stream, but an increasing number of studies shows that the features used in these CNNs to classify objects are very different from those used by humans (e.g., Azulay and Weiss, 2018; Baker, Lu, Erlikhman, and Kelleman VSS 2018; Ullman et al., 2016). Some caution may be warranted.

In Figure 3 performance increase across layers is shown. This plot is created by using the best performing weighting parameter β. To dissociate the effect of β and of the attentional modulation f, please add a control condition in which f is set to one. The reasoning is that varying β alone and picking the best β may already induce performance changes.

In Figure 3 the results for modulating all layers are shown. It is felt that the conclusions drawn from these results are unsupported. Picking β at 1/10 of the optimal β for each layer does not constitute an optimal setting for modulating all layers. Also, modulations in early layers may negatively impact activity changes in later layers. Hence, the authors cannot exclude the possibility that modulation of all layers simultaneously could actually help. Results and interpretations of attention applied to all layers should therefore be removed from the paper.

In the subsection “Attention Strength and the Tradeoff between Increasing True and False Positives”, you compare the change in the magnitude of neural activation in the CNN to the changes in primate brains. It was not clear how to interpret these results. Can these magnitudes be meaningfully compared? What can we conclude from this?

In the subsection “Feature-based Attention Primarily Influences Criteria and Spatial Attention Primarily Influences Sensitivity”, it is argued that FBA works through a criterion shift rather than by increasing sensitivity, with FBA shifting the representation of all stimuli in the direction of the attended category. But earlier you show that FBA selectively increases TP (relative to FP), which suggests an increase in sensitivity. (Also, Figure 4E appears to show a positive effect of FBA (L13) on sensitivity). Please clarify.

For many of the analyses results of both types of feedback are shown. However, for some comparisons only tuning-based results are shown (e.g., see the aforementioned subsection). Why? Please ensure consistency throughout.

It is unclear why a new method for quantifying attention is introduced in Figure 7, or how the "FSGM-like" measure is related to feature matching and the activity ratios already discussed. Please motivate or restrict to feature matching and activity ratios. In general, the paper is a dense read due to the various analysis and metrics. Any steps towards simplification of the presentation will aid the reader.

The claim that the new measure of attention (Figure 7A), or the alternative measures of attention for that matter, is experimentally testable seems unsupported. In particular, getting with and without attention activity in response to images that are not classified as the target orientation is not possible to measure in experiments with humans or animals. Subjects are stochastic in their judgments. One could however, measure with and without attention responses to ambiguous stimuli that elicit near chance performance. This metric would then become very similar to a population version of the well-studied "choice probability" metric. This connection should at least be discussed.

It would be more useful to the experimental community to recast the orientation task and analysis more in terms of what would be measured empirically. For instance presenting the task in terms of correctly identified target orientation as a function of the presented orientations rotation from the target. This may be outside the scope of current manuscript though.

What would be the biological mechanism that can account for tuning-based and gradient-based feedback? Especially the gradient based approach seems to be hard to defend from a biological point of view. How would putative decision-related areas have access to this gradient information? Some words should be spent on this in the Discussion section.

Please mention relevant related work:

- Katz et al., 2016, related to the relationship between tuning and influence on decisions.

- Abdelhack and Kamitani, 2018, related to subsection “Recordings Show How Feature Similarity Gain Effects Propagate” (and Figure 7) showing that the activity in response to misclassified stimuli shifts towards the activity in response to correctly classified stimuli when attention is turned on.

- Stein and Peelen, 2015, related to the subsection “Feature-based Attention Primarily Influences Criteria and Spatial Attention Primarily Influences Sensitivity”, arguing that FBA in human experiments does not lead to an increase in sensitivity (see also work by Carrasco on effects of FBA on discrimination tasks).

- Discussion, fifth paragraph: Ni, Ray, and Maunsell, 2012, would appear to be very relevant to this Discussion section. Those authors found that strength of normalization was as strong a factor as tuning in the strength of attentional effects.

- The neural network community developed various models that implement some form of attention. See the Attention section in Hassabis et al., Neuron, 2017. The present work should be contrasted with the papers mentioned there.

Essential revisions:

Introduction, second paragraph: CNNs are introduced as great models of the ventral stream, but an increasing number of studies shows that the features used in these CNNs to classify objects are very different from those used by humans (e.g., Azulay and Weiss, 2018; Baker, Lu, Erlikhman, and Kelleman VSS 2018; Ullman et al., 2016). Some caution may be warranted.

​The language and references have been updated to reflect this (Introduction, second paragraph).

In Figure 3 performance increase across layers is shown. This plot is created by using the best performing weighting parameter β. To dissociate the effect of β and of the attentional modulation f, please add a control condition in which f is set to one. The reasoning is that varying β alone and picking the best β may already induce performance changes.

This control has been added as Figure 3—figure supplement 1C, subsection “Feature-based Attention Improves Performance on Challenging Object Classification Tasks”, fourth paragraph.

In Figure 3 the results for modulating all layers are shown. It is felt that the conclusions drawn from these results are unsupported. Picking β at 1/10 of the optimal β for each layer does not constitute an optimal setting for modulating all layers. Also, modulations in early layers may negatively impact activity changes in later layers. Hence, the authors cannot exclude the possibility that modulation of all layers simultaneously could actually help. Results and interpretations of attention applied to all layers should therefore be removed from the paper.

We shifted the all-layer results to supplementary figures. Although it is true that we have not exhaustively explored the possibilities with modulation at all layers, we believe that the particular case we did study is of sufficient interest to readers to be included in the supplement. What we have shown is that modulating all layers, with 1/10 the strength of a single layer, improves performance about as much as modulation of the best single layer. Figure 3—figure supplements 1A and B, subsection “Feature-based Attention Improves Performance on Challenging Object Classification Tasks”, fourth paragraph.

In the subsection “Attention Strength and the Tradeoff between Increasing True and False Positives”, you compare the change in the magnitude of neural activation in the CNN to the changes in primate brains. It was not clear how to interpret these results. Can these magnitudes be meaningfully compared? What can we conclude from this?

We feel that the main take-away from that calculation is that the neural changes needed to cause observed performance chances are neither absurdly large nor absurdly small when compared to the what is observed experimentally in neurons. Therefore, we believe that as an order-of-magnitude comparison it provides some insight and reassurance that the model may be working similarly to the biology. We have shortened this section to make it clearer that this is the intended interpretation (subsection “Attention Strength and the Tradeoff between Increasing True and False Positives”, eighth paragraph).

In the subsection “Feature-based Attention Primarily Influences Criteria and Spatial Attention Primarily Influences Sensitivity”, it is argued that FBA works through a criterion shift rather than by increasing sensitivity, with FBA shifting the representation of all stimuli in the direction of the attended category. But earlier you show that FBA selectively increases TP (relative to FP), which suggests an increase in sensitivity. (Also, Figure 4E appears to show a positive effect of FBA (L13) on sensitivity). Please clarify.

We’d like to make two points of clarification to the reviewers on this regard:

First, the claim we make about feature-based attention (FBA) vs. spatial attention is that FBA has a larger impact on criteria than sensitivity and vice versa for spatial attention. Thus, we do not claim that there are no sensitivity changes coming from FBA, only that the criteria changes are more dominant.

Second, the fact that FBA first (with increasing modulation strength) increases TP (true positives) much more than FP (false positives), and only later (for much stronger modulation strength) increases FP more than TP, need not suggest an increase in sensitivity (where increasing sensitivity means increasing the separation between the distribution of TP and FP). This outcome could result solely from a criteria change (a decrease in threshold), as shown by the example in Author response image 1.

Author response image 1

Download asset Open asset

Here, the shaded area representing true positives will increase as the threshold moves leftward, before the false positives increase.

The text reflects these points (subsection “Feature-based Attention Primarily Influences Criteria and Spatial Attention Primarily Influences Sensitivity”, second paragraph).

For many of the analyses results of both types of feedback are shown. However, for some comparisons only tuning-based results are shown (e.g., see the aforementioned subsection). Why? Please ensure consistency throughout.

The gradient-based results were omitted in Figure 5E only due to the already crowded nature of that graph. Gradient results have now been added to that figure.

It is unclear why a new method for quantifying attention is introduced in Figure 7, or how the "FSGM-like" measure is related to feature matching and the activity ratios already discussed. Please motivate or restrict to feature matching and activity ratios. In general, the paper is a dense read due to the various analysis and metrics. Any steps towards simplification of the presentation will aid the reader.

The feature-similarity gain model (FSGM) measure in Figure 6 was taken to match a measure used in previous experimental work. It requires the presentation of multiple orientations in order to calculate the impact of attention on a given cell. That is, it measures how the effects of attention change across stimuli for a single cell. In Figure 7, we wanted to test for FSGM-like activity changes for a single stimulus, because this allows us to best study the correlation of FSGM-like activity changes with performance changes (because performance varies from stimulus to stimulus). Therefore we had to measure how the effects of attention change across cells, i.e. across the population as a whole, for responses to a single orientation (or even a single image). That is why we introduce a new “FSGM-like” measure in Figure 7. Both measures test whether attention multiplicatively increases (decreases) the activity of cells that are tuned to prefer (anti-prefer) the attended stimulus. The difference is simply whether the attention vs. no-attention activity ratios used to fit the line are taken from the same cell (which for our model means the same feature map) in response to different orientations (Figure 6B), or from all the feature maps in response to the same orientation (Figure 7). Our new population measure, like the single-cell measure which was used previously in experiments, can be easily implemented by experimentalists. We agree with the reviewers that we did not make our reasons for using a new measure, and the precise difference between the two measures, clear, and we have now made these points clear in our revision.

Having a measure of FSGM-like activity that could be calculated per-orientation allowed us to make a more fine-grained analysis of the correlation between FSGM-like activity changes and performance changes (because we did not have to average performance over all orientations). We show these correlations in Figure 7C. In that figure we also introduced a second measure of attention-induced activity changes (vector angle). This measure was introduced to see if a in hopes of finding a different measure of activity changes that better correlates with performance changes than the FSGM measure. The vector angle measure was meant to be an experimentally tractable measure that better reflects gradients rather than FSGM-like tuning curves, following our findings that the gradients should be most potent in affecting performance. We show that, at early layers, the vector angle measure of attention-induced activity changes does indeed better correlate with performance changes than the FSGM-like measure.

However, we agree with the reviewers that Figure 7 and the associated new analyses are complex and could be confusing. Yet we believe the proposal of different measures of activity changes and the exploration of their correlation with performance changes are valuable contributions from this kind of modeling work. Furthermore the results of Figure 7 were modest rather than striking. We therefore referenced the findings of these analyses in the main text and moved Figure 7, and a detailed explanation of the different measures and their motivations, to the supplementary materials.

In its place, we have included, as Figure 7, a figure illustrating the point we think will be of most interest to experimentalists at this point in our paper, namely a figure illustrating an experiment to clearly distinguish whether attention is applied according to gradients rather than (as always assumed up to this point) according to tuning. We previously only addressed this experiment in the Discussion, but we believe it is best to make it a prominent and illustrated part of the Results.

This new figure is now Figure 7, and the old Figure 7 has been moved to Figure 6—figure supplement 2. The analysis techniques in Figure 6—figure supplement 2 are referenced in the sixth paragraph of the subsection “Recordings Show How Feature Similarity Gain Effects Propagate”, and further explanation has been added in the Materials and methods subsection “Assessment of Feature Similarity Gain Model and Feature Matching Behavior” as well as to the caption of Figure 6—figure supplement 2. The experiment (Figure 7) is described in the last paragraph of the subsection “Recordings Show How Feature Similarity Gain Effects Propagate”.

The claim that the new measure of attention (Figure 7A), or the alternative measures of attention for that matter, is experimentally testable seems unsupported. In particular, getting with and without attention activity in response to images that are not classified as the target orientation is not possible to measure in experiments with humans or animals. Subjects are stochastic in their judgments. One could however, measure with and without attention responses to ambiguous stimuli that elicit near chance performance. This metric would then become very similar to a population version of the well-studied "choice probability" metric. This connection should at least be discussed.

The alternative measures (we take this to mean the FSGM measures of Figure 7 vs. Figure 6) have already, in the Figure 6 version, been used experimentally, and we see no reason why the population-based version of Figure 7 should not also be usable. We understand the rest of this comment to be referring to the “new measure”, namely the vector-angle method of assaying attention-induced activity changes. We agree with the reviewers that behavioral responses will be stochastic, whereas those in our model are deterministic, but disagree that this renders the vector-angle method unusable. As the reviewers suggest in referring to a threshold task, so long as the method is applied to images in which, without attention, there is a significant percentage of mis-categorizations, and attention improves performance, the measure will be usable: the method only requires that population responses be measured (1) without attention for mis-categorization (2) without attention for correct categorization and (3) with attention for both cases. We believe introducing this measure to experimentalists, which should better test the ability of attention to modulate gradients rather than tuning, will be a valuable contribution, although again, we have largely put this in the supplement.

Furthermore, it is not necessary for the blue and red vectors in Figure 7A to be calculated only using images that were negatively-classified without attention. The same result would be expected if all images were used. To be specific, a collection of images, some of which contained the desired orientation and some that did not, could be shown to a subject (who is not deploying attention to the orientation). The positive-classification vector (gray vector in 7A) would be made from the responses when an image was positively-classified (regardless of whether the desired orientation was present). The blue vector would be made from the responses to all images that contained the desired orientation (regardless of how they were classified), and the red vector would be from the responses to images that contained the desired orientation (regardless of how they were classified) but when attention was applied to that orientation. In this setting we would still expect the red vector (attention present) to be closer to positive classification vector than the blue (attention absent).

The results of calculating these vectors this way are shown in the green lines in Author response image 2 (with original analyses for comparison), and give results virtually identical to those of our previously-used definition of vector angle (blue lines).

Author response image 2

Download asset Open asset

Figure 6—figure supplement 2 now uses the new definition of the Vector Angle measure (that which uses all images, green line in Author response image 2). The last paragraph of the subsection “Assessment of Feature Similarity Gain Model and Feature Matching Behavior” has been changed to reflect that.

It would be more useful to the experimental community to recast the orientation task and analysis more in terms of what would be measured empirically. For instance presenting the task in terms of correctly identified target orientation as a function of the presented orientations rotation from the target. This may be outside the scope of current manuscript though.

We agree that relating the performance trends in these models more to behavioral data from various experimentally-used tasks would be interesting, but indeed outside the scope of the current manuscript.

What would be the biological mechanism that can account for tuning-based and gradient-based feedback? Especially the gradient based approach seems to be hard to defend from a biological point of view. How would putative decision-related areas have access to this gradient information? Some words should be spent on this in the Discussion section.

We believe there are means by which the gradient-based feedback could be learned biologically, through correlation-based or spike-timing-based learning rules in which, if lower-level activity is followed by high-level activity, the feedback from those higher-level units to the corresponding lower-level units is strengthened. We have fleshed this out in the Discussion (fifth paragraph).

Please mention relevant related work:

- Katz et al., 2016, related to the relationship between tuning and influence on decisions.

- Abdelhack and Kamitani, 2018, related to subsection “Recordings Show How Feature Similarity Gain Effects Propagate” (and Figure 7) showing that the activity in response to misclassified stimuli shifts towards the activity in response to correctly classified stimuli when attention is turned on.

- Stein and Peelen, 2015, related to the subsection “Feature-based Attention Primarily Influences Criteria and Spatial Attention Primarily Influences Sensitivity”, arguing that FBA in human experiments does not lead to an increase in sensitivity (see also work by Carrasco on effects of FBA on discrimination tasks).

- Discussion, fifth paragraph: Ni, Ray, and Maunsell, 2012, would appear to be very relevant to this Discussion section. Those authors found that strength of normalization was as strong a factor as tuning in the strength of attentional effects.

- The neural network community developed various models that implement some form of attention. See the Attention section in Hassabis et al., Neuron, 2017. The present work should be contrasted with the papers mentioned there.

These citations have been included in the Discussion.

Author details

1. Grace W Lindsay

   1. Center for Theoretical Neuroscience, College of Physicians and Surgeons, Columbia University, New York, United States
   2. Mortimer B. Zuckerman Mind Brain Behaviour Institute, Columbia University, New York, United States

   Contribution

   Conceptualization, Funding acquisition, Investigation, Visualization, Methodology, Writing—original draft

   For correspondence

   gracewlindsay@gmail.com

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-9904-7471

2. Kenneth D Miller

   1. Center for Theoretical Neuroscience, College of Physicians and Surgeons, Columbia University, New York, United States
   2. Mortimer B. Zuckerman Mind Brain Behaviour Institute, Columbia University, New York, United States
   3. Swartz Program in Theoretical Neuroscience, Kavli Institute for Brain Science, New York, United States
   4. Department of Neuroscience, Columbia University, New York, United States

   Contribution

   Supervision, Funding acquisition, Methodology, Writing—review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-1433-0647

• 4,785

  Page views

• 622

  Downloads

• 24

  Citations

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