The human brain is made up of billions of neurons, which are organised into networks via trillions of connections. The study of the nature of these connections will be central to understanding how the brain works. In recent years, a number of new methods for imaging the brain have made it possible to visualise and map these connections, generating striking images and creating an additional field of neuroscience known as ‘connectomics’.

However, the sheer volume of data generated by connectomics is now beginning to exceed the capacity of researchers to analyse it. Just as the advent of genome sequencing required the development of statistical techniques to analyse the resulting data, so the emergence of connectomics has created a need for similarly powerful mathematical models in neuroscience.

Jonas and Kording have developed one such algorithm that can classify the component units of circuits, both biological and man-made, and identify the connections between them. When applied to connectomics data for 950 neurons in the mouse retina, the algorithm generated predictions regarding cell types and patterns of connectivity. The predicted cell types agreed closely with those identified by human neuroanatomists. Results were similarly convincing when the algorithm was applied to the nervous system of the nematode worm and genetic model organism, Caenorhabditis elegans, and even when it was asked to classify electronic components and connectivity patterns in a man-made microprocessor.

Algorithms such as that developed by Jonas and Kording will soon be essential for making sense of the vast quantities of data generated by connectomic studies of the human brain. At present, an analysis of 950 neurons requires several hours, thus refinements that make the process faster will likely be required prior to the analysis of larger human datasets. Such algorithms will open up a range of possibilities for examining the structure of the healthy brain, as well as the changes triggered by developmental abnormalities and disease.

Emerging connectomics techniques (Zador et al., 2012; Morgan and Lichtman, 2013) promise to quantify the location and connectivity of each neuron within a tissue volume. These massive datasets will far exceed the capacity of neuroanatomists to manually trace small circuits, thus necessitating computational, quantitative, and automatic methods for understanding neural circuit structure. The impact of this kind of high-throughput transition has been seen before—the rise of sequencing techniques necessitated the development of novel computational methods to understand genomic structure, ushering in bioinformatics as an independent discipline (Koboldt et al., 2013).

The brain consists of multiple kinds of neurons, each of which is hypothesized to have a specific role in overall computation. Neuron types differ in many ways, for example, chemical or morphological, but they also differ in the way they connect to one another (Seung and Sümbül, 2014). In fact, the idea of well defined, type-dependent local connectivity patterns (microcircuits) has a long history (Passingham, 2002), and is prominent in many areas, from sensory (e.g., retina; Masland, 2001) to processing (e.g., neocortex; Mountcastle, 1997) to movement (e.g., spinal cord; Grillner et al., 2005). These types of repeated computing patterns are a common feature of computing systems, even occurring in man-made computing circuits. It remains an important challenge to develop algorithms to use connectivity-based anatomical data (connectomics) to automatically back out underlying microcircuitry.

The discovery of structure is a crucial aspect of network science. Early approaches focused on global graph properties, such as the types of scaling present in the network (Watts and Strogatz, 1998). While this approach leads to an understanding of the global network, more recent work aims at identifying very small-scale repeat patterns, or motifs, in networks (Milo et al., 2002). These motifs are defined not between different node types, but rather represent repeated patterns of topology.

The discovery of structure in probabilistic graphs is a well-known problem in machine learning. Commonly used algorithms include community-based detection methods (Girvan and Newman, 2002) and stochastic block models (Nowicki and Snijders, 2001). While these approaches can incorporate the probabilistic nature of neural connections (Hill et al., 2012), they do not incorporate the additional richer structure present in connectomics data—the location of cell bodies, the spatial distribution of synapses, and the distances between neurons. It is of particular importance that the probability of connections has a strong spatial component, a factor that is hard to reconcile with many other methods. A model attempting to fully capture the variation in the nervous system should take into account the broad set of available features.

When it comes to neuroscience and other computing systems, we expect patterns of connectivity much more complex than traditional motifs, exhibiting a strong spatial dependence arising from the complex genetic, chemical, and activity-based neural development processes.

To address these challenges, here we describe a Bayesian non-parametric model that can discover circuit structure automatically from connectomics data: the cell types, their spatial patterns of interconnection, and the locations of somata and synapses. We show that by incorporating this additional information, our model both accurately predicts the connection as well as agrees with human neuroanatomists as to the identification of cell types. We take as inspiration previous work on identifying cell types automatically from morphology (Guerra et al., 2011) and electrophysiology (Druckmann et al., 2013).

We primarily focus on the recently released mouse retina connectome (Helmstaedter et al., 2013), but additionally examine the Caenorhabditis elegans connectome (White et al., 1986). Comparing the cell types discovered by the algorithms with those obtained manually by human anatomists reveals a high degree of agreement. We thus present a scalable probabilistic approach to infer microcircuitry from connectomics data available today and in the future.

We build a structured probabilistic model which begins with the generic notion of a cell being a member of a single type—and these types affect soma depth, distribution of synapses, as well as a cell type and distance-dependent connection probability. For example, retinal ganglion cells may synapse on nearby, but not far away, amacrine cells, with bipolar cells clearly tessellating space and synapsing on both. In machine learning parlance, our method is unsupervised—it seeks to discover structure in data and make predictions in the absence of training data. Rather than taking in examples of types annotated by human neuroanatomists, we instead start with the weakest possible assumption in an attempt to algorithmically discover this structure. We contrast this with the supervised approaches taken in Guerra et al. (2011), where there is high confidence in the (morphologically defined) types and then a supervised classifier is built, as our goal here is explicit discovery of types.

From these assumptions (priors) we develop a generative Bayesian model that estimates the underlying cell types and how they connect. We take as input (Figure 1A) the connectivity matrix of cells (Figure 1B), a matrix of the distance between cells (Figure 1C), the per-cell soma depth (Figure 1D), and the depth profile of the cell's synapses (Figure 1E). We perform joint probabilistic inference to automatically learn the number of cell types, which cells belong to which type, their type-specific connectivity, and how connections between types vary with distance. We also simultaneously learn the soma depth associated with each type and the typical synaptic density profile (Figure 1F–H).

Figure 1

Download asset Open asset

We start with a model for connectivity, the iSBM (Kemp et al., 2006; Xu et al., 2006), which has been shown to meaningfully cluster connection graphs while learning the number of hidden groups, or types. We extend this approach by adding distance dependence to model salient aspects of microcircuitry via logistic and exponential distance-link functions. We form a unimodial model of cell body depth and a multimodal synapse density profile model (see ‘Materials and methods’ for mathematical details).

As an illustrative example, consider a network with only three cell types, labeled A, B, and C. Assume these cells are uniformly distributed in space, and that the probability of connection between any two cells, c_i and c_j, depends only on their type and their distance, according to a logistic (sigmoidal) function. Let A connect only to nearby B and C cells, but B connect to any C regardless of distance. This is the prior intuition our model is designed to capture.

For the basic link-distance model, we take as input a connectivity matrix R defining the connections between cell e_i and e_j, as well as a distance function d(e_i, e_j) representing a (physical) distance between adjacent cells. See the supplemental material for extension to multiple connectivity matrices. We assume there exist an unknown number K of latent (unobserved) cell types, $k\in \left\{1,2,3,\dots ,K\right\}$, and that each cell e_i belongs to a single cell type. We indicate a cell e_i is of type k using the assignment vector (c), so c_i = k. The observed connectivity between two cells R(e_i, e_j) then depends only on their latent type and their distance through a link function f(⋅, d(e_i, e_j)). We assume f is parameterized based on the latent type, c_i = m and c_j = n, via a parameter η_mn, as well as a set of global hyperparameters θ, such that the link function is f(d(e_i, e_j)|η_mn, θ).

We then jointly infer the posterior distribution of the class assignment vector (c) = {c_i}, the parameter matrix η_mn, and the global model hyperparameters θ:

Our subsequent analysis uses both the full posterior distribution over these parameters as well as the most probable, or maximum a posteriori (MAP), estimate.

For the retina data, we then extend the model with the additional features indicated. Cell soma depth is modeled as a cell-type-dependent Gaussian distribution with latent (unknown) per-type mean and variance. Similarly, each cell has some number N_i of synapses, each of which is drawn from a cell-type-specific density profile with up to three modes.

Inference is performed via MCMC via three composable transition kernels—one for structural, one for per-type parameters, and one for global parameters and hyperparameters. Details of data preprocessing, inference parameters, and runtime can be found in the 'Methods section.

We will first establish that our algorithm works properly and try to understand its properties using simulated data. Subsequently, we will analyze in detail a dataset on the retina. Lastly, we will briefly discuss the analysis of data from the worm C. elegans and from an old man-made microprocessor.

Validation with simulated data where ground truth is known

To validate our model, we performed a series of simulations to test if the model can accurately recover the true underlying network structure and cell type identity. We thus simulate data for which we know the correct structure and compare the estimated structure based on the algorithm (see ‘Materials and methods’) with the one we used for simulation. We find that the model does a good job of recovering the correct number of cell types (Figure 2A), the cell identities (Figure 2B), and the spatial extent of each type (Figure 2C). For comparison, we show the results using the infinite stochastic block model (iSBM) instead (Figure 2A–C, black line) which assumes that only cell type matters, and thus finds small neighborhoods of connected nodes (instead of global connectivity patterns). This contrast shows that while the regular block model can not correctly deal with distance-dependent connectivity, our model can. Our model converges relatively quickly (see 'Mixing of Markov chains') to an estimate of the most probable values for the cell types, which is enabled by using a combination of simulated annealing and parallelized Markov-chain Monte Carlo (MCMC) (see ‘Materials and methods’ for details). Thus our model at least is promising for application to biological datasets.

Figure 2

Download asset Open asset

Model mismatch

We next analyze how our model performs in cases where the data are generated with assumptions different from ours. To understand the properties of our model, we attempt connectivity inference on four sets of synthetic data. This helps us understand what our model would do if the data do not obey our assumptions.

We thus generate 10 sets of synthetic data from each of four existing models. The distance-dependent stochastic block model assumes type depends on distance, the traditional stochastic block model has no notion of distance, the mixed membership block model assumes type is combinatorial, and the latent position cluster model assumes that type is clustered-but-continuous.

If the data are sampled from our model, inference according to our model, unsurprisingly, is good by all measures. It correctly estimates the number of cell types, it is good at predicting connectivity (high area under the curve, AUC), it agrees with human classification (Rand index), it discovers all types, and leads to homogeneous estimates (Figure 3, first row). If the data come from a block model without distance dependence, we see that it still does well on all meaningful measures (Figure 3, second row). This is unsurprising, as our model learns the distance dependence, even its absence. For the mixed membership model (Figure 3, third row), the model grossly overestimates the number of types, by basically allocating a type for each combination of memberships. Otherwise, it still performs relatively well. Lastly, for the latent position clustering model (Figure 3, fourth row), the model does poorly. If type is continuous instead of discrete, then our model is basically trying to cover a continuous set with a discrete scenario leading to rather poor performance. However, as we do expect cell types to have a discrete biological basis, we might expect our model to do well with real data.

Figure 3

Download asset Open asset

Sensitivity to edge effects

Connectomic efforts so far have reconstructed only small sections of neural tissue. Consequently, many connections to cells outside that tissue volume will be lost. We are concerned that this selective elimination of connectivity along the boundary might give the appearance of distance-dependent connectivity when there is none. We thus performed simulations to check if edge effects could destroy spatial structure and if edge effects could introduce artificial, spurious spatial structure. We measure the degree to which distance-dependent effects can arise from selecting regions that are smaller than the ‘scale’ of connectivity (Figure 4). We do this by generating two collections of synthetic datasets—one with distance-dependent connectivity and one without. We then in each dataset randomly examine contiguous circular regions with area varying from zero to the entire volume, and empirically calculate the spatial variance in type-dependent connectivity. We find that, if there is no distance dependence, edge effects do not artificially introduce distance dependence. However, if the section we are examining is too small, our model can miss the distance dependence. Thus with respect to distance-dependent connectivity inference, our model errs on the side of caution. But we also find that for spatial extent that is similar to the currently available datasets, the effects of this are quite limited.

Figure 4

Download asset Open asset

Learning types and circuitry in the retina

The mouse retina (Masland, 2001) is a neural circuit which we expect to have connectivity patterns that are well approximated by our generative model. It is known that there are multiple classes of cells that can be broadly grouped into: ganglion cells that transmit information to the rest of the brain; bipolar cells that connect between different cells; and amacrine cells that feed into the ganglion cells. Recent research (Helmstaedter et al., 2013) has produced a large dataset containing both the types of cells from orthogonal approaches, and also the connectivity matrix between all reconstructed cells (Figure 5A).

Figure 5

Download asset Open asset

The algorithm took 8 hr to perform inference, dividing neurons into a set of cell types which reflect known neuroanatomical distinctions (Figure 5 shows the MAP result). For each pair of neurons there is a specific distance-dependent connection probability (Figure 5D), which is well approximated by the model fit. Moreover, each type of cell is rather isotropically distributed across space (Figure 5C) as should be expected for true cell types.

Comparing the results of the algorithm to other information sources allows evaluation of the quality of the type determination. Our types closely reflect the (anatomist-determined) segmentation of cells into retinal ganglion, narrow amacrine, medium/wide amacrine, and bipolar cells (Figure 6B). We find that the types we find tend to reflect the known laminar distribution in the retina (Figure 6C) as well as the known synaptic density profiles.

Figure 6

Download asset Open asset

The algorithm yields a separation of neurons into a smaller number of types than the fully granular listing of 71 types found by the original authors of the paper (Helmstaedter et al., 2013), although it is still highly correlated with those finer type distinctions (see section ‘Mouse retina’). It is our expectation that, with larger datasets, even closer agreement would be found.

Our fully Bayesian model produces a distribution over probable clusterings. Figure 6 shows this posterior distribution as a cell–cell coassignment matrix, sorted to find maximum block structure. Each large, dark block represents a collection of cells believed with strong probability to be of the same type. When we plot (Figure 6B) the anatomist-derived cell types along the left, we can see that each block consists of a roughly homogeneous collection of types.

We evaluate our model along three sets of parameters (Figure 6): how closely does our clustering agree with neuroanatomists' knowledge? Given two cells, how accurately can our model predict the link between them? And how closely does the spatial extent (within a layer) of our identified types agree with the spatial extent of types identified by neuroanatomists?

For our model we show the receiver operating characteristic (ROC) curve (Figure 6D) which shows how the true and false positive rates trade off. We plot the posterior distribution of the area under this curve in Figure 6E. We then plot the posterior distribution for cluster agreement metrics—completeness, homogeneity, and adjusted Rand index (ARI) (Figure 6F). We see that our model tends to over-cluster—cells which are of distinct type (at the finest granularity of neuroanatomist-identified type) are grouped as a single type by our model.

We compare link-prediction accuracy across the methods, including our own (Figure 6G, AUC, red). We find that given the dataset, many techniques allow for good link-predictive accuracy. All the methods allow decent link prediction with an AUC in the 0.9 range. However, our algorithm clearly outperforms the simple statistical models that only use connectivity.

As a second measure we compare link-prediction accuracy across the methods (Figure 6G, ARI, blue). We find that our algorithm far outperforms the controls. We also find that when it is based on more of the same information used by anatomists, then it gets better at agreeing with these anatomists. In particular, using connectivity, distance, synapse distribution, and soma depth leads to the highest ARI. When using the available information, the algorithm produces a good fit to human anatomist judgments.

Finally we look at the spatial extent of the discovered types both within a layer and between layers (Figure 6H). We see that, in the absence of distance information, mere connectivity information results in types which only span a small region of space—essentially local cliques. Incorporation of distance information results in types which span the entire extent of the layer. The depth variance of all models continues to be substantially larger than that predicted by human anatomists—future directions of work include attempting to more strongly encode this prior belief of laminarity.

Recovering spatial connectivity in multiple graphs simultaneously

Having shown our model to work on the repeating tessellated, laminar structure of the mammalian retina, we then apply our model to a structurally very different connectome—the whole body of a small roundworm: C. elegans is a model system in developmental neuroscience (White et al., 1986), with the location and connectivity of each of 302 neurons developmentally determined, leading to early measurement of the connectome. Unlike the retina, only the motor neurons in C. elegans exhibit regular distribution in space—along the body axis. Most interneurons are concentrated in various ganglia that project throughout the entire animal, and the sensory neurons are primarily located in a small number of anterior ganglia. C. elegans also differs from the retina in that the measured connectome is actually two separate graphs—one of directed chemical synapses and another of undirected electrical synapses. As this is a very different connectome, it allows an interesting generalization test: how well will our model work on such a distinct dataset?

Using both the chemical and electrical connectivity (see ‘Materials and methods’), we determined the underlying cell types explained by connectivity and distance (Figure 7A). A superficial inspection of the results shows clustering into groups consisting roughly homogeneously of motor neurons, sensory neurons, and interneurons. Closer examination reveals agreement with the classifications originally outlined by White in 1986 (White et al., 1986).

Figure 7

Download asset Open asset

Note our clustering does not perfectly reflect known divisions—several combinations of head and sensory neurons are combined, and a difficult-to-explain group of mostly VB and DB motor neuron types, with VC split between various groups. Our identified cell types thus reflect a ‘coarsening’ of known types, based entirely on connectivity and distance information, even when the organism exhibits substantially less spatial regularity than the retina.

Types and connectivity in artificial structures

To show the applicability of our method to other connectome-style datasets, we obtained the spatial location and interconnectivity of the transistors in a classic microprocessor, the MOS Technology 6502 (used in the Apple II) (James et al., 2010). Computer architects use common patterns of transistors when designing circuits, with each transistor having a ‘type’ in the circuit. We identified a region of the processor with complex but known structure containing the primary 8-bit registers X, Y, and S (Figure 8).

Figure 8

Download asset Open asset

Our algorithm identifies areas of spatial homogeneity that mirror the known structure in the underlying architecture of the circuit, segmenting transistor types recognizable to computer architects. Using the original schematics, we see that one identified type contains the ‘clocked’ transistors, which retain digital state. Two other types contain transistors with pins C1 or C2 connected to ground, mostly serving as inverters. An additional identified type controls the behavior of the three registers of interest (X, Y, and S) with respect to the SB data bus, either allowing them to latch or drive data from the bus. The repeat patterns of spatial connectivity are visible in Figure 8C, showing the man-made horizontal and vertical layout of the same types of transistors.

We have presented a machine learning technique that allows cell types and microcircuitry to be discovered from connectomics data. We have shown its applicability to regularly structured laminar neural circuits like the retina, as well as a less structured whole neuronal organism (C. elegans) and a classic processor. When compared to existing methods, we show how the incorporation of all of this data yields results that combine both high link-prediction accuracy and high agreement with human anatomists. We have found that combining the available data types allows us to discover cell types and microcircuitry that were known to exist in the systems based on decades of previous research and allows good prediction of connectivity.

For our probabilistic models, no known solution exists to exactly find the most probable parsing of the neurons into cell types and connectivity patterns. We employ a collection of MCMC techniques (see ‘Materials and methods’), but while different initializations converge to similar ultimate values, we can never realistically obtain the global optimum. There are a broad range of techniques that may offer good approximations to the global optimum and future work could adapt them to find more precise solutions to our problem.

For our probabilistic model, inference becomes slower as the amount of data increases. Our algorithm required several hours for 1000 neurons. Scaling this class of probabilistic model is an active area of research, and recent results in both variational methods (Hoffman et al., 2013) and spectral learning (Anandkumar et al., 2012) and future work could adapt them to find faster approximate solutions to our problem.

Larger datasets will allow algorithms to distinguish more distinct types and we expect closer agreement with existing anatomical knowledge as more data become available. Moreover, in general, for such problems precision increases with the size of the dataset and the cells that we have are not sufficient to statistically distinguish all the cell types known in anatomy (such as the ∼70 in the retina). Still, using only connectivity and distance, it is possible to meaningfully divide neurons into types.

Our small collection of hand-selected distance-dependent likelihood functions is clearly non-exhaustive, and assumes monotonicity of connectivity probability—for a given class, closer cells are never less likely to connect. This is known to be insufficient for various neural systems. Future models could incorporate a wider variety of likelihood functions, or even learn the global functional form from the data.

There are a range of previous approaches to the discovery of neural microcircuitry (Mountcastle, 1957; Douglas and Martin, 1991; Freund and Buzsáki, 1998; Barthó et al., 2004). These generally involve a great deal of manual labor and ad hoc determination of what constitutes a type of cell—to this day there are disagreements in the literature as to the true types in the mammalian retina. Much as phylogenomics has changed our understanding of animal ontologies, modern large scale data will allow the efficient unbiased discovery of cell types and circuits. The sheer amount of available data demands the introduction of algorithmic approaches.

The development of automatic identification and quantification of cell type may also provide a new computational phenotype for quantifying the effect of disease, genetic interventions, and developmentally experienced neural activity. Our method can in principle identify neuron types across non-connected graphs, for example, across animals. For example, the types of neurons in one animal can be associated with the types of neurons in another animal, in the same way as this is already possible through molecular markers (Brown and Hestrin, 2009). This could be particularly important if cell types appear that are due to properties of the stimuli and experience as opposed to just the molecular properties of cells, such as color and orientation selective types in primary visual cortex (Lennie and Movshon, 2005; Sincich and Horton, 2005). This would allow comparative quantitative anatomy across animals, and aid the search for the ultimate causes of connectivity.

Our model combines connectivity, cellular and synaptic properties, and suggests the way towards combining even richer data. Distinct cell types differ in morphology, connectivity, transcriptomics, relation to behavior or stimuli, and many other ways. Algorithms combining these data and type information may allow us to synthesize all the available information from one experiment or even across experiments into a joint model of brain structure and function.

Our work shows how rich probabilistic models can contribute to computational neuroanatomy. Eventually, algorithms will have to become a central tool for anatomists, as it will progressively become impossible for humans to parse the huge datasets. This transition may follow a similar transition to that of molecular biology (with gene-finding algorithms) and evolutionary biology (with computational phylogenetics). Ultimately, computational approaches may help resolve the significant disagreements across human anatomists.

1. 1. Jonas E

   (2015) Connectomics datasets

   Canonical, cleaned-up datasets, publicly available at GitHub, originally published in Varshney et al., 2011 (http://dx.doi.org/10.1371/journal.pcbi.1001066) and Helmstaedter et al., 2013 (http://dx.doi.org/10.1038/nature12346).

   https://github.com/ericmjonas/circuitdata

1. 1. Anandkumar A
   2. Ge R
   3. Hsu D
   4. Kakade SM
   5. Telgarsky M

   (2012)

   Tensor decompositions for learning latent variable models

   Journal of Machine Learning Research 15:1–55.

   • Google Scholar
2. 1. Barthó P
   2. Hirase H
   3. Monconduit L
   4. Zugaro M
   5. Harris KD
   6. Buzsáki G

   (2004) Characterization of neocortical principal cells and interneurons by network interactions and extracellular features

   Journal of Neurophysiology 92:600–608.

   https://doi.org/10.1152/jn.01170.2003

   • Google Scholar
3. 1. Brown SP
   2. Hestrin S

   (2009) Cell-type identity: a key to unlocking the function of neocortical circuits

   Current Opinion in Neurobiology 19:415–421.

   https://doi.org/10.1016/j.conb.2009.07.011

   • Google Scholar
4. 1. Chen BL
   2. Hall DH
   3. Chklovskii DB

   (2006) Wiring optimization can relate neuronal structure and function

   Proceedings of the National Academy of Sciences of USA 103:4723–4728.

   https://doi.org/10.1073/pnas.0506806103

   • Google Scholar
5. 1. Douglas R
   2. Martin K

   (1991) A functional microcircuit for cat visual cortex

   The Journal of Physiology 440:735–769.

   https://doi.org/10.1113/jphysiol.1991.sp018733

   • Google Scholar
6. 1. Druckmann S
   2. Hill S
   3. Schürmann F
   4. Markram H
   5. Segev I

   (2013) A hierarchical structure of cortical interneuron electrical diversity revealed by automated statistical analysis

   Cerebral Cortex 23:2994–3006.

   https://doi.org/10.1093/cercor/bhs290

   • Google Scholar
7. 1. Freund T
   2. Buzsáki G

   (1998) Interneurons of the hippocampus

   Hippocampus 6:347–470.

   https://doi.org/10.1002/(SICI)1098-1063(1996)6:4<347::AID-HIPO1>3.0.CO;2-I

   • Google Scholar
8. 1. Girvan M
   2. Newman ME

   (2002) Community structure in social and biological networks

   Proceedings of the National Academy of Sciences of USA 99:7821–7826.

   https://doi.org/10.1073/pnas.122653799

   • Google Scholar
9. 1. Grillner S
   2. Markram H
   3. De Schutter E
   4. Silberberg G
   5. LeBeau FE

   (2005) Microcircuits in action from CPGs to neocortex

   Trends in Neurosciences 28:525–533.

   https://doi.org/10.1016/j.tins.2005.08.003

   • Google Scholar
10. 1. Guerra L
    2. McGarry LM
    3. Robles V
    4. Bielza C
    5. Larrañaga P
    6. Yuste R

    (2011) Comparison between supervised and unsupervised classifications of neuronal cell types: a case study

    Developmental Neurobiology 71:71–82.

    https://doi.org/10.1002/dneu.20809

    • Google Scholar
11. 1. Helmstaedter M
    2. Briggman KL
    3. Turaga SC
    4. Jain V
    5. Seung HS
    6. Denk W

    (2013) Connectomic reconstruction of the inner plexiform layer in the mouse retina

    Nature 500:168–174.

    https://doi.org/10.1038/nature12346

    • Google Scholar
12. 1. Hill SL
    2. Wang Y
    3. Riachi I
    4. Schürmann F
    5. Markram H

    (2012) Statistical connectivity provides a sufficient foundation for specific functional connectivity in neocortical neural microcircuits

    Proceedings of the National Academy of Sciences of USA 109:E2885–E2894.

    https://doi.org/10.1073/pnas.1202128109

    • Google Scholar
13. 1. Hoffman MD
    2. Blei DM
    3. Wang C
    4. Paisley J

    (2013)

    Stochastic variational inference

    Journal of Machine Learning Research 14:1303–1347.

    • Google Scholar
14. 1. Hubert L
    2. Arabie P

    (1985) Comparing partitions

    Journal of Classification 2:193–218.

    https://doi.org/10.1007/BF01908075

    • Google Scholar
15. 1. James G
    2. Silverman B
    3. Silverman B

    (2010)

    ACM SIGGRAPH 2010 Talks

    1, ACM SIGGRAPH 2010 Talks, New York, USA, ACM Press, 10.1145/1837026.1837061.

    • Google Scholar
16. 1. Kemp C
    2. Tenenbaum J
    3. Griffiths T

    (2006)

    Proceedings of the 21st National Conference on Artificial Intelligence

    381–388, Proceedings of the 21st National Conference on Artificial Intelligence, volume 1, Palo Alto, California, AAAI Press, 10.1145/1837026.1837061.

    • Google Scholar
17. 1. Koboldt DC
    2. Steinberg KM
    3. Larson DE
    4. Wilson RK
    5. Mardis ER

    (2013) The next-generation sequencing revolution and its impact on genomics

    Cell 155:27–38.

    https://doi.org/10.1016/j.cell.2013.09.006

    • Google Scholar
18. 1. Lennie P
    2. Movshon JA

    (2005) Coding of color and form in the geniculostriate visual pathway (invited review)

    Journal of the Optical Society of America. A, Optics, Image Science, and Vision 22:2013–2033.

    https://doi.org/10.1364/JOSAA.22.002013

    • Google Scholar
19. 1. Masland RH

    (2001) The fundamental plan of the retina

    Nature Neuroscience 4:877–886.

    https://doi.org/10.1038/nn0901-877

    • Google Scholar
20. 1. Milo R
    2. Shen-Orr S
    3. Itzkovitz S
    4. Kashtan N
    5. Chklovskii D
    6. Alon U

    (2002) Network motifs: simple building blocks of complex networks

    Science 298:824–827.

    https://doi.org/10.1126/science.298.5594.824

    • Google Scholar
21. 1. Morgan JL
    2. Lichtman JW

    (2013) Why not connectomics?

    Nature Methods 10:494–500.

    https://doi.org/10.1038/nmeth.2480

    • Google Scholar
22. 1. Mountcastle VB

    (1957)

    Modality and topographic properties of single neurons of cat's somatic sensory cortex

    Journal of Neurophysiology 20:408–434.

    • Google Scholar
23. 1. Mountcastle VB

    (1997)

    The columnar organization of the neocortex

    Brain 120:701–722.

    • Google Scholar
24. Book

    1. Murphy KP

    (2012)

    Machine learning: A probabilistic perspective

    Cambridge: The MIT Press.

    • Google Scholar
25. 1. Neal RM

    (2000)

    Markov chain sampling methods for Dirichlet process mixture models

    Journal of Computational and Graphical Statistics 9:249.

    • Google Scholar
26. 1. Neal R

    (2003) Slice sampling

    Annals of Statistics 31:705–741.

    https://doi.org/10.1214/aos/1056562461

    • Google Scholar
27. 1. Nowicki K
    2. Snijders TA

    (2001) Estimation and prediction for stochastic blockstructures

    Journal of the American Statistical Association 96:1077–1087.

    https://doi.org/10.1198/016214501753208735

    • Google Scholar
28. 1. Passingham RE

    (2002) The frontal cortex: does size matter?

    Nature Neuroscience 5:190–192.

    https://doi.org/10.1038/nn0302-190

    • Google Scholar
29. 1. Salter-Townshend M
    2. Murphy TB

    (2013) Variational Bayesian inference for the latent position cluster model for network data

    Computational Statistics & Data Analysis 57:661–671.

    https://doi.org/10.1016/j.csda.2012.08.004

    • Google Scholar
30. 1. Seung HS
    2. Sümbül U

    (2014) Neuronal cell types and connectivity: Lessons from the retina

    Neuron 83:1262–1272.

    https://doi.org/10.1016/j.neuron.2014.08.054

    • Google Scholar
31. 1. Sincich LC
    2. Horton JC

    (2005) The circuitry of V1 and V2: integration of color, form, and motion

    Annual Review of Neuroscience 28:303–326.

    https://doi.org/10.1146/annurev.neuro.28.061604.135731

    • Google Scholar
32. 1. Varshney LR
    2. Chen BL
    3. Paniagua E
    4. Hall DH
    5. Chklovskii DB

    (2011) Structural properties of the Caenorhabditis elegans neuronal network

    PLOS Computational Biology 7:e1001066.

    https://doi.org/10.1371/journal.pcbi.1001066

    • Google Scholar
33. 1. Watts DJ
    2. Strogatz SH

    (1998) Collective dynamics of ‘small-world’ networks

    Nature 393:440–442.

    https://doi.org/10.1038/30918

    • Google Scholar
34. 1. White J
    2. Southgate E
    3. Thomson J
    4. Brenner S

    (1986) The structure of the nervous system of the nematode Caenorhabditis elegans

    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 314:1–340.

    https://doi.org/10.1098/rstb.1986.0056

    • Google Scholar
35. 1. Xu Z
    2. Tresp V
    3. Yu K
    4. Kriegel HP

    (2006)

    Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence

    Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence, Arlington, Virginia, AUAI Press.

    • Google Scholar
36. 1. Zador AM
    2. Dubnau J
    3. Oyibo HK
    4. Zhan H
    5. Cao G
    6. Peikon ID

    (2012) Sequencing the connectome

    PLOS Biology 10:e1001411.

    https://doi.org/10.1371/journal.pbio.1001411

    • Google Scholar

Author details

1. Eric Jonas

   Department of Electrical Engineering and Computer Science, University of California, Berkeley, Berkeley, United States

   Contribution

   EJ, Conception and design, Analysis and interpretation of data, Drafting or revising the article

   For correspondence

   jonas@eecs.berkeley.edu

   Competing interests

   The authors declare that no competing interests exist.

2. Konrad Kording

   1. Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, United States
   2. Department of Physical Medicine and Rehabilitation, Rehabilitation Institute of Chicago, Chicago, United States
   3. Department of Physiology, Northwestern University, Chicago, United States

   Contribution

   KK, Conception and design, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

• 3,936

  views

• 790

  downloads

• 45

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.