Sparse coding in striate and extrastriate visual cortex

Stimulus-dependent synaptic plasticity underlies neuronal circuitry refinement in the mouse primary visual cortex

Perceptual learning improves our ability to interpret sensory stimuli present in our environment through experience. Despite its importance, the underlying mechanisms that enable perceptual learning in our sensory cortices are still not fully understood. In this study, we used in vivo two-photon imaging to investigate the functional and structural changes induced by visual stimulation in the mouse primary visual cortex (V1). Our results demonstrate that repeated stimulation leads to a refinement of V1 circuitry by decreasing the number of responsive neurons while potentiating their response. At the synaptic level, we observe a reduction in the number of dendritic spines and an overall increase in spine AMPA receptor levels in the same subset of neurons. In addition, visual stimulation induces synaptic potentiation in neighboring spines within individual dendrites. These findings provide insights into the mechanisms of synaptic plasticity underlying information processing in the neocortex.

Representational maps in the brain: concepts, approaches, and applications

Neural systems have evolved to process sensory stimuli in a way that allows for efficient and adaptive behavior in a complex environment. Recent technological advances enable us to investigate sensory processing in animal models by simultaneously recording the activity of large populations of neurons with single-cell resolution, yielding high-dimensional datasets. In this review, we discuss concepts and approaches for assessing the population-level representation of sensory stimuli in the form of a representational map. In such a map, not only are the identities of stimuli distinctly represented, but their relational similarity is also mapped onto the space of neuronal activity. We highlight example studies in which the structure of representational maps in the brain are estimated from recordings in humans as well as animals and compare their methodological approaches. Finally, we integrate these aspects and provide an outlook for how the concept of representational maps could be applied to various fields in basic and clinical neuroscience.

Odorant detection in a locust exhibits unusually low redundancy

Olfactory coding, from insects to humans, is canonically considered to involve considerable across-fiber coding already at the peripheral level, thereby allowing recognition of vast numbers of odor compounds. We show that the migratory locust has evolved an alternative strategy built on highly specific odorant receptors feeding into a complex primary processing center in the brain. By collecting odors from food and different life stages of the locust, we identified 205 ecologically relevant odorants, which we used to deorphanize 48 locust olfactory receptors via ectopic expression in Drosophila. Contrary to the often broadly tuned olfactory receptors of other insects, almost all locust receptors were found to be narrowly tuned to one or very few ligands. Knocking out a single receptor using CRISPR abolished physiological and behavioral responses to the corresponding ligand. We conclude that the locust olfactory system, with most olfactory receptors being narrowly tuned, differs from the so-far described olfactory systems.

Beyond ℓ1 sparse coding in V1

Growing evidence indicates that only a sparse subset from a pool of sensory neurons is active for the encoding of visual stimuli at any instant in time. Traditionally, to replicate such biological sparsity, generative models have been using the ℓ1 norm as a penalty due to its convexity, which makes it amenable to fast and simple algorithmic solvers. In this work, we use biological vision as a test-bed and show that the soft thresholding operation associated to the use of the ℓ1 norm is highly suboptimal compared to other functions suited to approximating ℓp with 0 ≤ p < 1 (including recently proposed continuous exact relaxations), in terms of performance. We show that ℓ1 sparsity employs a pool with more neurons, i.e. has a higher degree of overcompleteness, in order to maintain the same reconstruction error as the other methods considered. More specifically, at the same sparsity level, the thresholding algorithm using the ℓ1 norm as a penalty requires a dictionary of ten times more units compared to the proposed approach, where a non-convex continuous relaxation of the ℓ0 pseudo-norm is used, to reconstruct the external stimulus equally well. At a fixed sparsity level, both ℓ0- and ℓ1-based regularization develop units with receptive field (RF) shapes similar to biological neurons in V1 (and a subset of neurons in V2), but ℓ0-based regularization shows approximately five times better reconstruction of the stimulus. Our results in conjunction with recent metabolic findings indicate that for V1 to operate efficiently it should follow a coding regime which uses a regularization that is closer to the ℓ0 pseudo-norm rather than the ℓ1 one, and suggests a similar mode of operation for the sensory cortex in general.

A Redundant Cortical Code for Speech Envelope

Animal communication sounds exhibit complex temporal structure because of the amplitude fluctuations that comprise the sound envelope. In human speech, envelope modulations drive synchronized activity in auditory cortex (AC), which correlates strongly with comprehension (Giraud and Poeppel, 2012; Peelle and Davis, 2012; Haegens and Zion Golumbic, 2018). Studies of envelope coding in single neurons, performed in nonhuman animals, have focused on periodic amplitude modulation (AM) stimuli and use response metrics that are not easy to juxtapose with data from humans. In this study, we sought to bridge these fields. Specifically, we looked directly at the temporal relationship between stimulus envelope and spiking, and we assessed whether the apparent diversity across neurons' AM responses contributes to the population representation of speech-like sound envelopes. We gathered responses from single neurons to vocoded speech stimuli and compared them to sinusoidal AM responses in auditory cortex (AC) of alert, freely moving Mongolian gerbils of both sexes. While AC neurons displayed heterogeneous tuning to AM rate, their temporal dynamics were stereotyped. Preferred response phases accumulated near the onsets of sinusoidal AM periods for slower rates (<8 Hz), and an over-representation of amplitude edges was apparent in population responses to both sinusoidal AM and vocoded speech envelopes. Crucially, this encoding bias imparted a decoding benefit: a classifier could discriminate vocoded speech stimuli using summed population activity, while higher frequency modulations required a more sophisticated decoder that tracked spiking responses from individual cells. Together, our results imply that the envelope structure relevant to parsing an acoustic stream could be read-out from a distributed, redundant population code.SIGNIFICANCE STATEMENT Animal communication sounds have rich temporal structure and are often produced in extended sequences, including the syllabic structure of human speech. Although the auditory cortex (AC) is known to play a crucial role in representing speech syllables, the contribution of individual neurons remains uncertain. Here, we characterized the representations of both simple, amplitude-modulated sounds and complex, speech-like stimuli within a broad population of cortical neurons, and we found an overrepresentation of amplitude edges. Thus, a phasic, redundant code in auditory cortex can provide a mechanistic explanation for segmenting acoustic streams like human speech.

Inference via sparse coding in a hierarchical vision model

Sparse coding has been incorporated in models of the visual cortex for its computational advantages and connection to biology. But how the level of sparsity contributes to performance on visual tasks is not well understood. In this work, sparse coding has been integrated into an existing hierarchical V2 model (Hosoya & Hyvärinen, 2015), but replacing its independent component analysis (ICA) with an explicit sparse coding in which the degree of sparsity can be controlled. After training, the sparse coding basis functions with a higher degree of sparsity resembled qualitatively different structures, such as curves and corners. The contributions of the models were assessed with image classification tasks, specifically tasks associated with mid-level vision including figure–ground classification, texture classification, and angle prediction between two line stimuli. In addition, the models were assessed in comparison with a texture sensitivity measure that has been reported in V2 (Freeman et al., 2013) and a deleted-region inference task. The results from the experiments show that although sparse coding performed worse than ICA at classifying images, only sparse coding was able to better match the texture sensitivity level of V2 and infer deleted image regions, both by increasing the degree of sparsity in sparse coding. Greater degrees of sparsity allowed for inference over larger deleted image regions. The mechanism that allows for this inference capability in sparse coding is described in this article.

Cortical hierarchy, dual counterstream architecture and the importance of top-down generative networks

Hierarchy is a major organizational principle of the cortex and underscores modern computational theories of cortical function. The local microcircuit amplifies long-distance inter-areal input, which show distance-dependent changes in their laminar profiles. Statistical modeling of these changes in laminar profiles demonstrates that inputs from multiple hierarchical levels to their target areas show remarkable consistency, allowing the construction of a cortical hierarchy based on a principle of hierarchical distance. The statistical modeling that is applied to structure can also be applied to laminar differences in the oscillatory coherence between areas thereby determining a functional hierarchy of the cortex. Close examination of the anatomy of inter-areal connectivity reveals a dual counterstream architecture with well-defined distance-dependent feedback and feedforward pathways in both the supra- and infragranular layers, suggesting a multiplicity of feedback pathways with well-defined functional properties. These findings are consistent with feedback connections providing a generative network involved in a wide range of cognitive functions. A dynamical model constrained by connectivity data sheds insight into the experimentally observed signatures of frequency-dependent Granger causality for feedforward versus feedback signaling. Concerted experiments capitalizing on recent technical advances and combining tract-tracing, high-resolution fMRI, optogenetics and mathematical modeling hold the promise of a much improved understanding of lamina-constrained mechanisms of neural computation and cognition. However, because inter-areal interactions involve cortical layers that have been the target of important evolutionary changes in the primate lineage, these investigations will need to include human and non-human primate comparisons.

Pseudosparse neural coding in the visual system of primates

When measuring sparseness in neural populations as an indicator of efficient coding, an implicit assumption is that each stimulus activates a different random set of neurons. In other words, population responses to different stimuli are, on average, uncorrelated. Here we examine neurophysiological data from four lobes of macaque monkey cortex, including V1, V2, MT, anterior inferotemporal cortex, lateral intraparietal cortex, the frontal eye fields, and perirhinal cortex, to determine how correlated population responses are. We call the mean correlation the pseudosparseness index, because high pseudosparseness can mimic statistical properties of sparseness without being authentically sparse. In every data set we find high levels of pseudosparseness ranging from 0.59-0.98, substantially greater than the value of 0.00 for authentic sparseness. This was true for synthetic and natural stimuli, as well as for single-electrode and multielectrode data. A model indicates that a key variable producing high pseudosparseness is the standard deviation of spontaneous activity across the population. Consistently high values of pseudosparseness in the data demand reconsideration of the sparse coding literature as well as consideration of the degree to which authentic sparseness provides a useful framework for understanding neural coding in the cortex.

Natural images are reliably represented by sparse and variable populations of neurons in visual cortex

Natural scenes sparsely activate neurons in the primary visual cortex (V1). However, how sparsely active neurons reliably represent complex natural images and how the information is optimally decoded from these representations have not been revealed. Using two-photon calcium imaging, we recorded visual responses to natural images from several hundred V1 neurons and reconstructed the images from neural activity in anesthetized and awake mice. A single natural image is linearly decodable from a surprisingly small number of highly responsive neurons, and the remaining neurons even degrade the decoding. Furthermore, these neurons reliably represent the image across trials, regardless of trial-to-trial response variability. Based on our results, diverse, partially overlapping receptive fields ensure sparse and reliable representation. We suggest that information is reliably represented while the corresponding neuronal patterns change across trials and collecting only the activity of highly responsive neurons is an optimal decoding strategy for the downstream neurons.

Face Representations via Tensorfaces of Various Complexities

Neurons selective for faces exist in humans and monkeys. However, characteristics of face cell receptive fields are poorly understood. In this theoretical study, we explore the effects of complexity, defined as algorithmic information (Kolmogorov complexity) and logical depth, on possible ways that face cells may be organized. We use tensor decompositions to decompose faces into a set of components, called tensorfaces, and their associated weights, which can be interpreted as model face cells and their firing rates. These tensorfaces form a high-dimensional representation space in which each tensorface forms an axis of the space. A distinctive feature of the decomposition algorithm is the ability to specify tensorface complexity. We found that low-complexity tensorfaces have blob-like appearances crudely approximating faces, while high-complexity tensorfaces appear clearly face-like. Low-complexity tensorfaces require a larger population to reach a criterion face reconstruction error than medium- or high-complexity tensorfaces, and thus are inefficient by that criterion. Low-complexity tensorfaces, however, generalize better when representing statistically novel faces, which are faces falling beyond the distribution of face description parameters found in the tensorface training set. The degree to which face representations are parts based or global forms a continuum as a function of tensorface complexity, with low and medium tensorfaces being more parts based. Given the computational load imposed in creating high-complexity face cells (in the form of algorithmic information and logical depth) and in the absence of a compelling advantage to using high-complexity cells, we suggest face representations consist of a mixture of low- and medium-complexity face cells.

Elementary motion sequence detectors in whisker somatosensory cortex

How somatosensory cortex (S1) encodes complex patterns of touch, as occur during tactile exploration, is poorly understood. In mouse whisker S1, temporally dense stimulation of local whisker pairs revealed that most neurons are not classical single-whisker feature detectors, but instead are strongly tuned to 2-whisker sequences involving the columnar whisker (CW) and one, specific surround whisker (SW), usually in SW-leading-CW order. Tuning was spatiotemporally precise and diverse across cells, generating a rate code for local motion vectors defined by SW-CW combinations. Spatially asymmetric, sublinear suppression for suboptimal combinations and near-linearity for preferred combinations sharpened combination tuning relative to linearly predicted tuning. This resembles computation of motion direction selectivity in vision. SW-tuned neurons, misplaced in the classical whisker map, had the strongest combination tuning. Thus, each S1 column contains a rate code for local motion sequences involving the CW, providing a basis for higher-order feature extraction.

On the Sparse Structure of Natural Sounds and Natural Images: Similarities, Differences, and Implications for Neural Coding

Sparse coding models of natural images and sounds have been able to predict several response properties of neurons in the visual and auditory systems. While the success of these models suggests that the structure they capture is universal across domains to some degree, it is not yet clear which aspects of this structure are universal and which vary across sensory modalities. To address this, we fit complete and highly overcomplete sparse coding models to natural images and spectrograms of speech and report on differences in the statistics learned by these models. We find several types of sparse features in natural images, which all appear in similar, approximately Laplace distributions, whereas the many types of sparse features in speech exhibit a broad range of sparse distributions, many of which are highly asymmetric. Moreover, individual sparse coding units tend to exhibit higher lifetime sparseness for overcomplete models trained on images compared to those trained on speech. Conversely, population sparseness tends to be greater for these networks trained on speech compared with sparse coding models of natural images. To illustrate the relevance of these findings to neural coding, we studied how they impact a biologically plausible sparse coding network's representations in each sensory modality. In particular, a sparse coding network with synaptically local plasticity rules learns different sparse features from speech data than are found by more conventional sparse coding algorithms, but the learned features are qualitatively the same for these models when trained on natural images.

A hierarchical sparse coding model predicts acoustic feature encoding in both auditory midbrain and cortex

The auditory pathway consists of multiple stages, from the cochlear nucleus to the auditory cortex. Neurons acting at different stages have different functions and exhibit different response properties. It is unclear whether these stages share a common encoding mechanism. We trained an unsupervised deep learning model consisting of alternating sparse coding and max pooling layers on cochleogram-filtered human speech. Evaluation of the response properties revealed that computing units in lower layers exhibited spectro-temporal receptive fields (STRFs) similar to those of inferior colliculus neurons measured in physiological experiments, including properties such as sound onset and termination, checkerboard pattern, and spectral motion. Units in upper layers tended to be tuned to phonetic features such as plosivity and nasality, resembling the results of field recording in human auditory cortex. Variation of the sparseness level of the units in each higher layer revealed a positive correlation between the sparseness level and the strength of phonetic feature encoding. The activities of the units in the top layer, but not other layers, correlated with the dynamics of the first two formants (F1, F2) of all phonemes, indicating the encoding of phoneme dynamics in these units. These results suggest that the principles of sparse coding and max pooling may be universal in the human auditory pathway.

When speech enters the ear, it is subjected to a series of processing stages prior to arriving at the auditory cortex. Neurons acting at different processing stages have different response properties. For example, at the auditory midbrain, a neuron may specifically detect the onsets of a frequency component in the speech, whereas in the auditory cortex, a neuron may specifically detect phonetic features. The encoding mechanisms underlying these neuronal functions remain unclear. To address this issue, we designed a hierarchical sparse coding model, inspired by the sparse activity of neurons in the sensory system, to learn features in speech signals. We found that the computing units in different layers exhibited hierarchical extraction of speech sound features, similar to those of neurons in the auditory midbrain and auditory cortex, although the computational principles in these layers were the same. The results suggest that sparse coding and max pooling represent universal computational principles throughout the auditory pathway.

Neural coding: A single neuron's perspective

What any sensory neuron knows about the world is one of the cardinal questions in Neuroscience. Information from the sensory periphery travels across synaptically coupled neurons as each neuron encodes information by varying the rate and timing of its action potentials (spikes). Spatiotemporally correlated changes in this spiking regimen across neuronal populations are the neural basis of sensory representations. In the somatosensory cortex, however, spiking of individual (or pairs of) cortical neurons is only minimally informative about the world. Recent studies showed that one solution neurons implement to counteract this information loss is adapting their rate of information transfer to the ongoing synaptic activity by changing the membrane potential at which spike is generated. Here we first introduce the principles of information flow from the sensory periphery to the primary sensory cortex in a model sensory (whisker) system, and subsequently discuss how the adaptive spike threshold gates the intracellular information transfer from the somatic post-synaptic potential to action potentials, controlling the information content of communication across somatosensory cortical neurons.

Gradual Development of Visual Texture-Selective Properties Between Macaque Areas V2 and V4

Complex shape and texture representations are known to be constructed from V1 along the ventral visual pathway through areas V2 and V4, but the underlying mechanism remains elusive. Recent study suggests that, for processing of textures, a collection of higher-order image statistics computed by combining V1-like filter responses serves as possible representations of textures both in V2 and V4. Here, to gain a clue for how these image statistics are processed in the extrastriate visual areas, we compared neuronal responses to textures in V2 and V4 of macaque monkeys. For individual neurons, we adaptively explored their preferred textures from among thousands of naturalistic textures and fitted the obtained responses using a combination of V1-like filter responses and higher-order statistics. We found that, while the selectivity for image statistics was largely comparable between V2 and V4, V4 showed slightly stronger sensitivity to the higher-order statistics than V2. Consistent with that finding, V4 responses were reduced to a greater extent than V2 responses when the monkeys were shown spectrally matched noise images that lacked higher-order statistics. We therefore suggest that there is a gradual development in representation of higher-order features along the ventral visual hierarchy.

Large-scale two-photon imaging revealed super-sparse population codes in the V1 superficial layer of awake monkeys

One general principle of sensory information processing is that the brain must optimize efficiency by reducing the number of neurons that process the same information. The sparseness of the sensory representations in a population of neurons reflects the efficiency of the neural code. Here, we employ large-scale two-photon calcium imaging to examine the responses of a large population of neurons within the superficial layers of area V1 with single-cell resolution, while simultaneously presenting a large set of natural visual stimuli, to provide the first direct measure of the population sparseness in awake primates. The results show that only 0.5% of neurons respond strongly to any given natural image - indicating a ten-fold increase in the inferred sparseness over previous measurements. These population activities are nevertheless necessary and sufficient to discriminate visual stimuli with high accuracy, suggesting that the neural code in the primary visual cortex is both super-sparse and highly efficient.

Assessing the Role of Inhibition in Stabilizing Neocortical Networks Requires Large-Scale Perturbation of the Inhibitory Population

Neurons within cortical microcircuits are interconnected with recurrent excitatory synaptic connections that are thought to amplify signals (Douglas and Martin, 2007), form selective subnetworks (Ko et al., 2011), and aid feature discrimination. Strong inhibition (Haider et al., 2013) counterbalances excitation, enabling sensory features to be sharpened and represented by sparse codes (Willmore et al., 2011). This balance between excitation and inhibition makes it difficult to assess the strength, or gain, of recurrent excitatory connections within cortical networks, which is key to understanding their operational regime and the computations that they perform. Networks that combine an unstable high-gain excitatory population with stabilizing inhibitory feedback are known as inhibition-stabilized networks (ISNs) (Tsodyks et al., 1997). Theoretical studies using reduced network models predict that ISNs produce paradoxical responses to perturbation, but experimental perturbations failed to find evidence for ISNs in cortex (Atallah et al., 2012). Here, we reexamined this question by investigating how cortical network models consisting of many neurons behave after perturbations and found that results obtained from reduced network models fail to predict responses to perturbations in more realistic networks. Our models predict that a large proportion of the inhibitory network must be perturbed to reliably detect an ISN regime robustly in cortex. We propose that wide-field optogenetic suppression of inhibition under promoters targeting a large fraction of inhibitory neurons may provide a perturbation of sufficient strength to reveal the operating regime of cortex. Our results suggest that detailed computational models of optogenetic perturbations are necessary to interpret the results of experimental paradigms.SIGNIFICANCE STATEMENT Many useful computational mechanisms proposed for cortex require local excitatory recurrence to be very strong, such that local inhibitory feedback is necessary to avoid epileptiform runaway activity (an "inhibition-stabilized network" or "ISN" regime). However, recent experimental results suggest that this regime may not exist in cortex. We simulated activity perturbations in cortical networks of increasing realism and found that, to detect ISN-like properties in cortex, large proportions of the inhibitory population must be perturbed. Current experimental methods for inhibitory perturbation are unlikely to satisfy this requirement, implying that existing experimental observations are inconclusive about the computational regime of cortex. Our results suggest that new experimental designs targeting a majority of inhibitory neurons may be able to resolve this question.

Deep Learning Predicts Correlation between a Functional Signature of Higher Visual Areas and Sparse Firing of Neurons

Visual information in the visual cortex is processed in a hierarchical manner. Recent studies show that higher visual areas, such as V2, V3, and V4, respond more vigorously to images with naturalistic higher-order statistics than to images lacking them. This property is a functional signature of higher areas, as it is much weaker or even absent in the primary visual cortex (V1). However, the mechanism underlying this signature remains elusive. We studied this problem using computational models. In several typical hierarchical visual models including the AlexNet, VggNet, and SHMAX, this signature was found to be prominent in higher layers but much weaker in lower layers. By changing both the model structure and experimental settings, we found that the signature strongly correlated with sparse firing of units in higher layers but not with any other factors, including model structure, training algorithm (supervised or unsupervised), receptive field size, and property of training stimuli. The results suggest an important role of sparse neuronal activity underlying this special feature of higher visual areas.

Temporal Information Entropy of the Blood-Oxygenation Level-Dependent Signals Increases in the Activated Human Primary Visual Cortex

Time-domain analysis of blood-oxygenation level-dependent (BOLD) signals allows the identification of clusters of voxels responding to photic stimulation in primary visual cortex (V1). However, the characterization of information encoding into temporal properties of the BOLD signals of an activated cluster is poorly investigated. Here, we used Shannon entropy to determine spatial and temporal information encoding in the BOLD signal within the most strongly activated area of the human visual cortex during a hemifield photic stimulation. We determined the distribution profile of BOLD signals during epochs at rest and under stimulation within small (19-121 voxels) clusters designed to include only voxels driven by the stimulus as highly and uniformly as possible. We found consistent and significant increases (2-4% on average) in temporal information entropy during activation in contralateral but not ipsilateral V1, which was mirrored by an expected loss of spatial information entropy. These opposite changes coexisted with increases in both spatial and temporal mutual information (i.e., dependence) in contralateral V1. Thus, we showed that the first cortical stage of visual processing is characterized by a specific spatiotemporal rearrangement of intracluster BOLD responses. Our results indicate that while in the space domain BOLD maps may be incapable of capturing the functional specialization of small neuronal populations due to relatively low spatial resolution, some information encoding may still be revealed in the temporal domain by an increase of temporal information entropy.

What do eye movements tell us about patients with neurological disorders? - An introduction to saccade recording in the clinical setting

Non-invasive and readily implemented in the clinical setting, eye movement studies have been conducted extensively not only in healthy human subjects but also in patients with neurological disorders. The purpose of saccade studies is to "read out" the pathophysiology underlying neurological disorders from the saccade records, referring to known primate physiology. In the current review, we provide an overview of studies in which we attempted to elucidate the patterns of saccade abnormalities in over 250 patients with neurological disorders, including cerebellar ataxia and brainstem pathology due to neurodegenerative disorders, and what they tell about the pathophysiology of patients with neurological disorders. We also discuss how interventions, such as deep brain stimulation, affect saccade performance and provide further insights into the workings of the oculomotor system in humans. Finally, we argue that it is important to understand the functional significance and behavioral correlate of saccade abnormalities in daily life, which could require eye tracking methodologies to be performed in settings similar to daily life.

Sparse coding generates curvature selectivity in V4 neurons

The cortical area V4 produces a representation of curvature as the intermediate-level representation of an object's shape. We investigated whether sparse coding is the principle driving the generation of the spatial properties of the receptive field in V4 that exhibit curvature selectivity. To investigate the role of sparseness in the construction of curvature representations, we applied component analysis with a sparseness constraint to the activity of model V2 neurons that were responding to shapes derived from natural images. Our simulation results showed that single basis functions with medium degrees of sparseness (0.7-0.8) produced curvature selectivity, and their population activity produced acute curvature bias. The results support the hypothesis that sparseness plays an essential role in the construction of curvature selectivity in V4.

Spatial Correlations in Natural Scenes Modulate Response Reliability in Mouse Visual Cortex

Intrinsic neuronal variability significantly limits information encoding in the primary visual cortex (V1). Certain stimuli can suppress this intertrial variability to increase the reliability of neuronal responses. In particular, responses to natural scenes, which have broadband spatiotemporal statistics, are more reliable than responses to stimuli such as gratings. However, very little is known about which stimulus statistics modulate reliable coding and how this occurs at the neural ensemble level. Here, we sought to elucidate the role that spatial correlations in natural scenes play in reliable coding. We developed a novel noise-masking method to systematically alter spatial correlations in natural movies, without altering their edge structure. Using high-speed two-photon calcium imaging in vivo, we found that responses in mouse V1 were much less reliable at both the single neuron and population level when spatial correlations were removed from the image. This change in reliability was due to a reorganization of between-neuron correlations. Strongly correlated neurons formed ensembles that reliably and accurately encoded visual stimuli, whereas reducing spatial correlations reduced the activation of these ensembles, leading to an unreliable code. Together with an ensemble-specific normalization model, these results suggest that the coordinated activation of specific subsets of neurons underlies the reliable coding of natural scenes.

Network Anisotropy Trumps Noise for Efficient Object Coding in Macaque Inferior Temporal Cortex

How neuronal ensembles compute information is actively studied in early visual cortex. Much less is known about how local ensembles function in inferior temporal (IT) cortex, the last stage of the ventral visual pathway that supports visual recognition. Previous reports suggested that nearby neurons carry information mostly independently, supporting efficient processing (Barlow, 1961). However, others postulate that noise covariation effects may depend on network anisotropy/homogeneity and on how the covariation relates to representation. Do slow trial-by-trial noise covariations increase or decrease IT's object coding capability, how does encoding capability relate to correlational structure (i.e., the spatial pattern of signal and noise redundancy/homogeneity across neurons), and does knowledge of correlational structure matter for decoding? We recorded simultaneously from ∼80 spiking neurons in ∼1 mm(3) of macaque IT under light neurolept anesthesia. Noise correlations were stronger for neurons with correlated tuning, and noise covariations reduced object encoding capability, including generalization across object pose and illumination. Knowledge of noise covariations did not lead to better decoding performance. However, knowledge of anisotropy/homogeneity improved encoding and decoding efficiency by reducing the number of neurons needed to reach a given performance level. Such correlated neurons were found mostly in supragranular and infragranular layers, supporting theories that link recurrent circuitry to manifold representation. These results suggest that redundancy benefits manifold learning of complex high-dimensional information and that subsets of neurons may be more immune to noise covariation than others.

SIGNIFICANCE STATEMENT

How noise affects neuronal population coding is poorly understood. By sampling densely from local populations supporting visual object recognition, we show that recurrent circuitry supports useful representations and that subsets of neurons may be more immune to noise covariation than others.

A Probabilistic Analysis of Sparse Coded Feature Pooling and Its Application for Image Retrieval

Feature coding and pooling as a key component of image retrieval have been widely studied over the past several years. Recently sparse coding with max-pooling is regarded as the state-of-the-art for image classification. However there is no comprehensive study concerning the application of sparse coding for image retrieval. In this paper, we first analyze the effects of different sampling strategies for image retrieval, then we discuss feature pooling strategies on image retrieval performance with a probabilistic explanation in the context of sparse coding framework, and propose a modified sum pooling procedure which can improve the retrieval accuracy significantly. Further we apply sparse coding method to aggregate multiple types of features for large-scale image retrieval. Extensive experiments on commonly-used evaluation datasets demonstrate that our final compact image representation improves the retrieval accuracy significantly.

Differential combinatorial coding of pheromones in two olfactory subsystems of the honey bee brain

Neural coding of pheromones has been intensively studied in insects with a particular focus on sex pheromones. These studies favored the view that pheromone compounds are processed within specific antennal lobe glomeruli following a specialized labeled-line system. However, pheromones play crucial roles in an insect's life beyond sexual attraction, and some species use many different pheromones making such a labeled-line organization unrealistic. A combinatorial coding scheme, in which each component activates a set of broadly tuned units, appears more adapted in this case. However, this idea has not been tested thoroughly. We focused here on the honey bee Apis mellifera, a social insect that relies on a wide range of pheromones to ensure colony cohesion. Interestingly, the honey bee olfactory system harbors two central parallel pathways, whose functions remain largely unknown. Using optophysiological recordings of projection neurons, we compared the responses of these two pathways to 27 known honey bee pheromonal compounds emitted by the brood, the workers, and the queen. We show that while queen mandibular pheromone is processed by l-ALT (lateral antennal lobe tract) neurons and brood pheromone is mainly processed by m-ALT (median antennal lobe tract) neurons, worker pheromones induce redundant activity in both pathways. Moreover, all tested pheromonal compounds induce combinatorial activity from several AL glomeruli. These findings support the combinatorial coding scheme and suggest that higher-order brain centers reading out these combinatorial activity patterns may eventually classify olfactory signals according to their biological meaning.

Mapping of functionally characterized cell classes onto canonical circuit operations in primate prefrontal cortex

Microcircuits are composed of multiple cell classes that likely serve unique circuit operations. But how cell classes map onto circuit functions is largely unknown, particularly for primate prefrontal cortex during actual goal-directed behavior. One difficulty in this quest is to reliably distinguish cell classes in extracellular recordings of action potentials. Here we surmount this issue and report that spike shape and neural firing variability provide reliable markers to segregate seven functional classes of prefrontal cells in macaques engaged in an attention task. We delineate an unbiased clustering protocol that identifies four broad spiking (BS) putative pyramidal cell classes and three narrow spiking (NS) putative inhibitory cell classes dissociated by how sparse, bursty, or regular they fire. We speculate that these functional classes map onto canonical circuit functions. First, two BS classes show sparse, bursty firing, and phase synchronize their spiking to 3-7 Hz (theta) and 12-20 Hz (beta) frequency bands of the local field potential (LFP). These properties make cells flexibly responsive to network activation at varying frequencies. Second, one NS and two BS cell classes show regular firing and higher rate with only marginal synchronization preference. These properties are akin to setting tonically the excitation and inhibition balance. Finally, two NS classes fired irregularly and synchronized to either theta or beta LFP fluctuations, tuning them potentially to frequency-specific subnetworks. These results suggest that a limited set of functional cell classes emerges in macaque prefrontal cortex (PFC) during attentional engagement to not only represent information, but to subserve basic circuit operations.

Correlated activity supports efficient cortical processing

Visual recognition is a computational challenge that is thought to occur via efficient coding. An important concept is sparseness, a measure of coding efficiency. The prevailing view is that sparseness supports efficiency by minimizing redundancy and correlations in spiking populations. Yet, we recently reported that "choristers", neurons that behave more similarly (have correlated stimulus preferences and spontaneous coincident spiking), carry more generalizable object information than uncorrelated neurons ("soloists") in macaque inferior temporal (IT) cortex. The rarity of choristers (as low as 6% of IT neurons) indicates that they were likely missed in previous studies. Here, we report that correlation strength is distinct from sparseness (choristers are not simply broadly tuned neurons), that choristers are located in non-granular output layers, and that correlated activity predicts human visual search efficiency. These counterintuitive results suggest that a redundant correlational structure supports efficient processing and behavior.

Frequency-band signatures of visual responses to naturalistic input in ferret primary visual cortex during free viewing

Neuronal firing responses in visual cortex reflect the statistics of visual input and emerge from the interaction with endogenous network dynamics. Artificial visual stimuli presented to animals in which the network dynamics were constrained by anesthetic agents or trained behavioral tasks have provided fundamental understanding of how individual neurons in primary visual cortex respond to input. In contrast, very little is known about the mesoscale network dynamics and their relationship to microscopic spiking activity in the awake animal during free viewing of naturalistic visual input. To address this gap in knowledge, we recorded local field potential (LFP) and multiunit activity (MUA) simultaneously in all layers of primary visual cortex (V1) of awake, freely viewing ferrets presented with naturalistic visual input (nature movie clips). We found that naturalistic visual stimuli modulated the entire oscillation spectrum; low frequency oscillations were mostly suppressed whereas higher frequency oscillations were enhanced. In average across all cortical layers, stimulus-induced change in delta and alpha power negatively correlated with the MUA responses, whereas sensory-evoked increases in gamma power positively correlated with MUA responses. The time-course of the band-limited power in these frequency bands provided evidence for a model in which naturalistic visual input switched V1 between two distinct, endogenously present activity states defined by the power of low (delta, alpha) and high (gamma) frequency oscillatory activity. Therefore, the two mesoscale activity states delineated in this study may define the degree of engagement of the circuit with the processing of sensory input.

Contextual modulation and stimulus selectivity in extrastriate cortex

Contextual modulation is observed throughout the visual system, using techniques ranging from single-neuron recordings to behavioral experiments. Its role in generating feature selectivity within the retina and primary visual cortex has been extensively described in the literature. Here, we describe how similar computations can also elaborate feature selectivity in the extrastriate areas of both the dorsal and ventral streams of the primate visual system. We discuss recent work that makes use of normalization models to test specific roles for contextual modulation in visual cortex function. We suggest that contextual modulation renders neuronal populations more selective for naturalistic stimuli. Specifically, we discuss contextual modulation's role in processing optic flow in areas MT and MST and for representing naturally occurring curvature and contours in areas V4 and IT. We also describe how the circuitry that supports contextual modulation is robust to variations in overall input levels. Finally, we describe how this theory relates to other hypothesized roles for contextual modulation.

Experience-dependent specialization of receptive field surround for selective coding of natural scenes

At eye opening, neurons in primary visual cortex (V1) are selective for stimulus features, but circuits continue to refine in an experience-dependent manner for some weeks thereafter. How these changes contribute to the coding of visual features embedded in complex natural scenes remains unknown. Here we show that normal visual experience after eye opening is required for V1 neurons to develop a sensitivity for the statistical structure of natural stimuli extending beyond the boundaries of their receptive fields (RFs), which leads to improvements in coding efficiency for full-field natural scenes (increased selectivity and information rate). These improvements are mediated by an experience-dependent increase in the effectiveness of natural surround stimuli to hyperpolarize the membrane potential specifically during RF-stimulus epochs triggering action potentials. We suggest that neural circuits underlying surround modulation are shaped by the statistical structure of visual input, which leads to more selective coding of features in natural scenes.

•

V1 firing is more selective to natural than phase-scrambled surround stimuli

•

Improved selectivity is caused by Vm hyperpolarisation prior to RF-spiking events

•

Natural surround sensitivity is experience dependent, absent in immature/deprived V1

•

Vision shapes circuits to improve V1 coding of natural scenes across RF and surround

Online stimulus optimization rapidly reveals multidimensional selectivity in auditory cortical neurons

Neurons in sensory brain regions shape our perception of the surrounding environment through two parallel operations: decomposition and integration. For example, auditory neurons decompose sounds by separately encoding their frequency, temporal modulation, intensity, and spatial location. Neurons also integrate across these various features to support a unified perceptual gestalt of an auditory object. At higher levels of a sensory pathway, neurons may select for a restricted region of feature space defined by the intersection of multiple, independent stimulus dimensions. To further characterize how auditory cortical neurons decompose and integrate multiple facets of an isolated sound, we developed an automated procedure that manipulated five fundamental acoustic properties in real time based on single-unit feedback in awake mice. Within several minutes, the online approach converged on regions of the multidimensional stimulus manifold that reliably drove neurons at significantly higher rates than predefined stimuli. Optimized stimuli were cross-validated against pure tone receptive fields and spectrotemporal receptive field estimates in the inferior colliculus and primary auditory cortex. We observed, from midbrain to cortex, increases in both level invariance and frequency selectivity, which may underlie equivalent sparseness of responses in the two areas. We found that onset and steady-state spike rates increased proportionately as the stimulus was tailored to the multidimensional receptive field. By separately evaluating the amount of leverage each sound feature exerted on the overall firing rate, these findings reveal interdependencies between stimulus features as well as hierarchical shifts in selectivity and invariance that may go unnoticed with traditional approaches.

Population code in mouse V1 facilitates readout of natural scenes through increased sparseness

Neural codes are believed to have adapted to the statistical properties of the natural environment. However, the principles that govern the organization of ensemble activity in the visual cortex during natural visual input are unknown. We recorded populations of up to 500 neurons in the mouse primary visual cortex and characterized the structure of their activity, comparing responses to natural movies with those to control stimuli. We found that higher order correlations in natural scenes induced a sparser code, in which information is encoded by reliable activation of a smaller set of neurons and can be read out more easily. This computationally advantageous encoding for natural scenes was state-dependent and apparent only in anesthetized and active awake animals, but not during quiet wakefulness. Our results argue for a functional benefit of sparsification that could be a general principle governing the structure of the population activity throughout cortical microcircuits.

Selectivity and sparseness in randomly connected balanced networks

Neurons in sensory cortex show stimulus selectivity and sparse population response, even in cases where no strong functionally specific structure in connectivity can be detected. This raises the question whether selectivity and sparseness can be generated and maintained in randomly connected networks. We consider a recurrent network of excitatory and inhibitory spiking neurons with random connectivity, driven by random projections from an input layer of stimulus selective neurons. In this architecture, the stimulus-to-stimulus and neuron-to-neuron modulation of total synaptic input is weak compared to the mean input. Surprisingly, we show that in the balanced state the network can still support high stimulus selectivity and sparse population response. In the balanced state, strong synapses amplify the variation in synaptic input and recurrent inhibition cancels the mean. Functional specificity in connectivity emerges due to the inhomogeneity caused by the generative statistical rule used to build the network. We further elucidate the mechanism behind and evaluate the effects of model parameters on population sparseness and stimulus selectivity. Network response to mixtures of stimuli is investigated. It is shown that a balanced state with unselective inhibition can be achieved with densely connected input to inhibitory population. Balanced networks exhibit the “paradoxical” effect: an increase in excitatory drive to inhibition leads to decreased inhibitory population firing rate. We compare and contrast selectivity and sparseness generated by the balanced network to randomly connected unbalanced networks. Finally, we discuss our results in light of experiments.

Neural codes formed by small and temporally precise populations in auditory cortex

The encoding of sensory information by populations of cortical neurons forms the basis for perception but remains poorly understood. To understand the constraints of cortical population coding we analyzed neural responses to natural sounds recorded in auditory cortex of primates (Macaca mulatta). We estimated stimulus information while varying the composition and size of the considered population. Consistent with previous reports we found that when choosing subpopulations randomly from the recorded ensemble, the average population information increases steadily with population size. This scaling was explained by a model assuming that each neuron carried equal amounts of information, and that any overlap between the information carried by each neuron arises purely from random sampling within the stimulus space. However, when studying subpopulations selected to optimize information for each given population size, the scaling of information was strikingly different: a small fraction of temporally precise cells carried the vast majority of information. This scaling could be explained by an extended model, assuming that the amount of information carried by individual neurons was highly nonuniform, with few neurons carrying large amounts of information. Importantly, these optimal populations can be determined by a single biophysical marker-the neuron's encoding time scale-allowing their detection and readout within biologically realistic circuits. These results show that extrapolations of population information based on random ensembles may overestimate the population size required for stimulus encoding, and that sensory cortical circuits may process information using small but highly informative ensembles.

Sparsity-regularized HMAX for visual recognition

About ten years ago, HMAX was proposed as a simple and biologically feasible model for object recognition, based on how the visual cortex processes information. However, the model does not encompass sparse firing, which is a hallmark of neurons at all stages of the visual pathway. The current paper presents an improved model, called sparse HMAX, which integrates sparse firing. This model is able to learn higher-level features of objects on unlabeled training images. Unlike most other deep learning models that explicitly address global structure of images in every layer, sparse HMAX addresses local to global structure gradually along the hierarchy by applying patch-based learning to the output of the previous layer. As a consequence, the learning method can be standard sparse coding (SSC) or independent component analysis (ICA), two techniques deeply rooted in neuroscience. What makes SSC and ICA applicable at higher levels is the introduction of linear higher-order statistical regularities by max pooling. After training, high-level units display sparse, invariant selectivity for particular individuals or for image categories like those observed in human inferior temporal cortex (ITC) and medial temporal lobe (MTL). Finally, on an image classification benchmark, sparse HMAX outperforms the original HMAX by a large margin, suggesting its great potential for computer vision.

Natural image sequences constrain dynamic receptive fields and imply a sparse code

In their natural environment, animals experience a complex and dynamic visual scenery. Under such natural stimulus conditions, neurons in the visual cortex employ a spatially and temporally sparse code. For the input scenario of natural still images, previous work demonstrated that unsupervised feature learning combined with the constraint of sparse coding can predict physiologically measured receptive fields of simple cells in the primary visual cortex. This convincingly indicated that the mammalian visual system is adapted to the natural spatial input statistics. Here, we extend this approach to the time domain in order to predict dynamic receptive fields that can account for both spatial and temporal sparse activation in biological neurons. We rely on temporal restricted Boltzmann machines and suggest a novel temporal autoencoding training procedure. When tested on a dynamic multi-variate benchmark dataset this method outperformed existing models of this class. Learning features on a large dataset of natural movies allowed us to model spatio-temporal receptive fields for single neurons. They resemble temporally smooth transformations of previously obtained static receptive fields and are thus consistent with existing theories. A neuronal spike response model demonstrates how the dynamic receptive field facilitates temporal and population sparseness. We discuss the potential mechanisms and benefits of a spatially and temporally sparse representation of natural visual input.

Are v1 simple cells optimized for visual occlusions? A comparative study

Simple cells in primary visual cortex were famously found to respond to low-level image components such as edges. Sparse coding and independent component analysis (ICA) emerged as the standard computational models for simple cell coding because they linked their receptive fields to the statistics of visual stimuli. However, a salient feature of image statistics, occlusions of image components, is not considered by these models. Here we ask if occlusions have an effect on the predicted shapes of simple cell receptive fields. We use a comparative approach to answer this question and investigate two models for simple cells: a standard linear model and an occlusive model. For both models we simultaneously estimate optimal receptive fields, sparsity and stimulus noise. The two models are identical except for their component superposition assumption. We find the image encoding and receptive fields predicted by the models to differ significantly. While both models predict many Gabor-like fields, the occlusive model predicts a much sparser encoding and high percentages of 'globular' receptive fields. This relatively new center-surround type of simple cell response is observed since reverse correlation is used in experimental studies. While high percentages of 'globular' fields can be obtained using specific choices of sparsity and overcompleteness in linear sparse coding, no or only low proportions are reported in the vast majority of studies on linear models (including all ICA models). Likewise, for the here investigated linear model and optimal sparsity, only low proportions of 'globular' fields are observed. In comparison, the occlusive model robustly infers high proportions and can match the experimentally observed high proportions of 'globular' fields well. Our computational study, therefore, suggests that 'globular' fields may be evidence for an optimal encoding of visual occlusions in primary visual cortex.

A three-layer model of natural image statistics

An important property of visual systems is to be simultaneously both selective to specific patterns found in the sensory input and invariant to possible variations. Selectivity and invariance (tolerance) are opposing requirements. It has been suggested that they could be joined by iterating a sequence of elementary selectivity and tolerance computations. It is, however, unknown what should be selected or tolerated at each level of the hierarchy. We approach this issue by learning the computations from natural images. We propose and estimate a probabilistic model of natural images that consists of three processing layers. Two natural image data sets are considered: image patches, and complete visual scenes downsampled to the size of small patches. For both data sets, we find that in the first two layers, simple and complex cell-like computations are performed. In the third layer, we mainly find selectivity to longer contours; for patch data, we further find some selectivity to texture, while for the downsampled complete scenes, some selectivity to curvature is observed.

Balanced increases in selectivity and tolerance produce constant sparseness along the ventral visual stream

Although popular accounts suggest that neurons along the ventral visual processing stream become increasingly selective for particular objects, this appears at odds with the fact that inferior temporal cortical (IT) neurons are broadly tuned. To explore this apparent contradiction, we compared processing in two ventral stream stages (visual cortical areas V4 and IT) in the rhesus macaque monkey. We confirmed that IT neurons are indeed more selective for conjunctions of visual features than V4 neurons and that this increase in feature conjunction selectivity is accompanied by an increase in tolerance ("invariance") to identity-preserving transformations (e.g., shifting, scaling) of those features. We report here that V4 and IT neurons are, on average, tightly matched in their tuning breadth for natural images ("sparseness") and that the average V4 or IT neuron will produce a robust firing rate response (>50% of its peak observed firing rate) to ∼10% of all natural images. We also observed that sparseness was positively correlated with conjunction selectivity and negatively correlated with tolerance within both V4 and IT, consistent with selectivity-building and invariance-building computations that offset one another to produce sparseness. Our results imply that the conjunction-selectivity-building and invariance-building computations necessary to support object recognition are implemented in a balanced manner to maintain sparseness at each stage of processing.

Associative fear learning enhances sparse network coding in primary sensory cortex

Several models of associative learning predict that stimulus processing changes during association formation. How associative learning reconfigures neural circuits in primary sensory cortex to "learn" associative attributes of a stimulus remains unknown. Using 2-photon in vivo calcium imaging to measure responses of networks of neurons in primary somatosensory cortex, we discovered that associative fear learning, in which whisker stimulation is paired with foot shock, enhances sparse population coding and robustness of the conditional stimulus, yet decreases total network activity. Fewer cortical neurons responded to stimulation of the trained whisker than in controls, yet their response strength was enhanced. These responses were not observed in mice exposed to a nonassociative learning procedure. Our results define how the cortical representation of a sensory stimulus is shaped by associative fear learning. These changes are proposed to enhance efficient sensory processing after associative learning.

Harnessing plasticity to understand learning and treat disease

A large body of evidence suggests that neural plasticity contributes to learning and disease. Recent studies suggest that cortical map plasticity is typically a transient phase that improves learning by increasing the pool of task-relevant responses. Here, I discuss a new perspective on neural plasticity and suggest how plasticity might be targeted to reset dysfunctional circuits. Specifically, a new model is proposed in which map expansion provides a form of replication with variation that supports a Darwinian mechanism to select the most behaviorally useful circuits. Precisely targeted neural plasticity provides a new avenue for the treatment of neurological and psychiatric disorders and is a powerful tool to test the neural mechanisms of learning and memory.

Precise feature based time scales and frequency decorrelation lead to a sparse auditory code

Sparse redundancy reducing codes have been proposed as efficient strategies for representing sensory stimuli. A prevailing hypothesis suggests that sensory representations shift from dense redundant codes in the periphery to selective sparse codes in cortex. We propose an alternative framework where sparseness and redundancy depend on sensory integration time scales and demonstrate that the central nucleus of the inferior colliculus (ICC) of cats encodes sound features by precise sparse spike trains. Direct comparisons with auditory cortical neurons demonstrate that ICC responses were sparse and uncorrelated as long as the spike train time scales were matched to the sensory integration time scales relevant to ICC neurons. Intriguingly, correlated spiking in the ICC was substantially lower than predicted by linear or nonlinear models and strictly observed for neurons with best frequencies within a "critical band," the hallmark of perceptual frequency resolution in mammals. This is consistent with a sparse asynchronous code throughout much of the ICC and a complementary correlation code within a critical band that may allow grouping of perceptually relevant cues.

Surround suppression and sparse coding in visual and barrel cortices

During natural vision the entire retina is stimulated. Likewise, during natural tactile behaviors, spatially extensive regions of the somatosensory surface are co-activated. The large spatial extent of naturalistic stimulation means that surround suppression, a phenomenon whose neural mechanisms remain a matter of debate, must arise during natural behavior. To identify common neural motifs that might instantiate surround suppression across modalities, we review models of surround suppression and compare the evidence supporting the competing ideas that surround suppression has either cortical or sub-cortical origins in visual and barrel cortex. In the visual system there is general agreement lateral inhibitory mechanisms contribute to surround suppression, but little direct experimental evidence that intracortical inhibition plays a major role. Two intracellular recording studies of V1, one using naturalistic stimuli (Haider et al., 2010), the other sinusoidal gratings (Ozeki et al., 2009), sought to identify the causes of reduced activity in V1 with increasing stimulus size, a hallmark of surround suppression. The former attributed this effect to increased inhibition, the latter to largely balanced withdrawal of excitation and inhibition. In rodent primary somatosensory barrel cortex, multi-whisker responses are generally weaker than single whisker responses, suggesting multi-whisker stimulation engages similar surround suppressive mechanisms. The origins of suppression in S1 remain elusive: studies have implicated brainstem lateral/internuclear interactions and both thalamic and cortical inhibition. Although the anatomical organization and instantiation of surround suppression in the visual and somatosensory systems differ, we consider the idea that one common function of surround suppression, in both modalities, is to remove the statistical redundancies associated with natural stimuli by increasing the sparseness or selectivity of sensory responses.