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.

Pooling strategies in V1 can account for the functional and structural diversity across species.

Neurons in the primary visual cortex are selective to orientation with various degrees of selectivity to the spatial phase, from high selectivity in simple cells to low selectivity in complex cells. Various computational models have suggested a possible link between the presence of phase invariant cells and the existence of orientation maps in higher mammals' V1. These models, however, do not explain the emergence of complex cells in animals that do not show orientation maps. In this study, we build a theoretical model based on a convolutional network called Sparse Deep Predictive Coding (SDPC) and show that a single computational mechanism, pooling, allows the SDPC model to account for the emergence in V1 of complex cells with or without that of orientation maps, as observed in distinct species of mammals. In particular, we observed that pooling in the feature space is directly related to the orientation map formation while pooling in the retinotopic space is responsible for the emergence of a complex cells population. Introducing different forms of pooling in a predictive model of early visual processing as implemented in SDPC can therefore be viewed as a theoretical framework that explains the diversity of structural and functional phenomena observed in V1.

Learning heterogeneous delays in a layer of spiking neurons for fast motion detection.

The precise timing of spikes emitted by neurons plays a crucial role in shaping the response of efferent biological neurons. This temporal dimension of neural activity holds significant importance in understanding information processing in neurobiology, especially for the performance of neuromorphic hardware, such as event-based cameras. Nonetheless, many artificial neural models disregard this critical temporal dimension of neural activity. In this study, we present a model designed to efficiently detect temporal spiking motifs using a layer of spiking neurons equipped with heterogeneous synaptic delays. Our model capitalizes on the diverse synaptic delays present on the dendritic tree, enabling specific arrangements of temporally precise synaptic inputs to synchronize upon reaching the basal dendritic tree. We formalize this process as a time-invariant logistic regression, which can be trained using labeled data. To demonstrate its practical efficacy, we apply the model to naturalistic videos transformed into event streams, simulating the output of the biological retina or event-based cameras. To evaluate the robustness of the model in detecting visual motion, we conduct experiments by selectively pruning weights and demonstrate that the model remains efficient even under significantly reduced workloads. In conclusion, by providing a comprehensive, event-driven computational building block, the incorporation of heterogeneous delays has the potential to greatly improve the performance of future spiking neural network algorithms, particularly in the context of neuromorphic chips.

A dual foveal-peripheral visual processing model implements efficient saccade selection.

We develop a visuomotor model that implements visual search as a focal accuracy-seeking policy, with the target's position and category drawn independently from a common generative process. Consistently with the anatomical separation between the ventral versus dorsal pathways, the model is composed of two pathways that respectively infer what to see and where to look. The "What" network is a classical deep learning classifier that only processes a small region around the center of fixation, providing a "foveal" accuracy. In contrast, the "Where" network processes the full visual field in a biomimetic fashion, using a log-polar retinotopic encoding, which is preserved up to the action selection level. In our model, the foveal accuracy is used as a monitoring signal to train the "Where" network, much like in the "actor/critic" framework. After training, the "Where" network provides an "accuracy map" that serves to guide the eye toward peripheral objects. Finally, the comparison of both networks' accuracies amounts to either selecting a saccade or keeping the eye focused at the center to identify the target. We test this setup on a simple task of finding a digit in a large, cluttered image. Our simulation results demonstrate the effectiveness of this approach, increasing by one order of magnitude the radius of the visual field toward which the agent can detect and recognize a target, either through a single saccade or with multiple ones. Importantly, our log-polar treatment of the visual information exploits the strong compression rate performed at the sensory level, providing ways to implement visual search in a sublinear fashion, in contrast with mainstream computer vision.

Bayesian Modeling of Motion Perception Using Dynamical Stochastic Textures.

A common practice to account for psychophysical biases in vision is to frame them as consequences of a dynamic process relying on optimal inference with respect to a generative model. The study presented here details the complete formulation of such a generative model intended to probe visual motion perception with a dynamic texture model. It is derived in a set of axiomatic steps constrained by biological plausibility. We extend previous contributions by detailing three equivalent formulations of this texture model. First, the composite dynamic textures are constructed by the random aggregation of warped patterns, which can be viewed as three-dimensional gaussian fields. Second, these textures are cast as solutions to a stochastic partial differential equation (sPDE). This essential step enables real-time, on-the-fly texture synthesis using time-discretized autoregressive processes. It also allows for the derivation of a local motion-energy model, which corresponds to the log likelihood of the probability density. The log likelihoods are essential for the construction of a Bayesian inference framework. We use the dynamic texture model to psychophysically probe speed perception in humans using zoom-like changes in the spatial frequency content of the stimulus. The human data replicate previous findings showing perceived speed to be positively biased by spatial frequency increments. A Bayesian observer who combines a gaussian likelihood centered at the true speed and a spatial frequency dependent width with a "slow-speed prior" successfully accounts for the perceptual bias. More precisely, the bias arises from a decrease in the observer's likelihood width estimated from the experiments as the spatial frequency increases. Such a trend is compatible with the trend of the dynamic texture likelihood width.

The Flash-Lag Effect as a Motion-Based Predictive Shift.

Due to its inherent neural delays, the visual system has an outdated access to sensory information about the current position of moving objects. In contrast, living organisms are remarkably able to track and intercept moving objects under a large range of challenging environmental conditions. Physiological, behavioral and psychophysical evidences strongly suggest that position coding is extrapolated using an explicit and reliable representation of object's motion but it is still unclear how these two representations interact. For instance, the so-called flash-lag effect supports the idea of a differential processing of position between moving and static objects. Although elucidating such mechanisms is crucial in our understanding of the dynamics of visual processing, a theory is still missing to explain the different facets of this visual illusion. Here, we reconsider several of the key aspects of the flash-lag effect in order to explore the role of motion upon neural coding of objects' position. First, we formalize the problem using a Bayesian modeling framework which includes a graded representation of the degree of belief about visual motion. We introduce a motion-based prediction model as a candidate explanation for the perception of coherent motion. By including the knowledge of a fixed delay, we can model the dynamics of sensory information integration by extrapolating the information acquired at previous instants in time. Next, we simulate the optimal estimation of object position with and without delay compensation and compared it with human perception under a broad range of different psychophysical conditions. Our computational study suggests that the explicit, probabilistic representation of velocity information is crucial in explaining position coding, and therefore the flash-lag effect. We discuss these theoretical results in light of the putative corrective mechanisms that can be used to cancel out the detrimental effects of neural delays and illuminate the more general question of the dynamical representation at the present time of spatial information in the visual pathways.

Reinforcement effects in anticipatory smooth eye movements.

When predictive information about target motion is available, anticipatory smooth pursuit eye movements (aSPEM) are consistently generated before target appearance, thereby reducing the typical sensorimotor delay between target motion onset and foveation. By manipulating the probability for target motion direction, we were able to bias the direction and mean velocity of aSPEM. This suggests that motion-direction expectancy has a strong effect on the initiation of anticipatory movements. To further understand the nature of anticipatory smooth eye movements, we investigated different effects of reinforcement on aSPEM. In a first experiment, the reinforcement was contingent to a particular anticipatory behavior. A monetary reward was associated to a criterion-matching anticipatory velocity as estimated online during the gap before target motion onset. Our results showed a small but significant effect of behavior-contingent monetary reward on aSPEM. In a second experiment, the proportion of rewarded trials was manipulated across motion directions (right vs. left) independently from participants' behavior. Our results indicate that a bias in expected reward does not systematically affect anticipatory eye movements. Overall, these findings strengthen the notion that anticipatory eye movements can be considered as an operant behavior (similar to visually guided ones), whereas the expectancy for a noncontingent reward cannot efficiently bias them.

Testing the odds of inherent vs. observed overdispersion in neural spike counts.

The repeated presentation of an identical visual stimulus in the receptive field of a neuron may evoke different spiking patterns at each trial. Probabilistic methods are essential to understand the functional role of this variance within the neural activity. In that case, a Poisson process is the most common model of trial-to-trial variability. For a Poisson process, the variance of the spike count is constrained to be equal to the mean, irrespective of the duration of measurements. Numerous studies have shown that this relationship does not generally hold. Specifically, a majority of electrophysiological recordings show an "overdispersion" effect: responses that exhibit more intertrial variability than expected from a Poisson process alone. A model that is particularly well suited to quantify overdispersion is the Negative-Binomial distribution model. This model is well-studied and widely used but has only recently been applied to neuroscience. In this article, we address three main issues. First, we describe how the Negative-Binomial distribution provides a model apt to account for overdispersed spike counts. Second, we quantify the significance of this model for any neurophysiological data by proposing a statistical test, which quantifies the odds that overdispersion could be due to the limited number of repetitions (trials). We apply this test to three neurophysiological data sets along the visual pathway. Finally, we compare the performance of this model to the Poisson model on a population decoding task. We show that the decoding accuracy is improved when accounting for overdispersion, especially under the hypothesis of tuned overdispersion.

Active inference, eye movements and oculomotor delays.

This paper considers the problem of sensorimotor delays in the optimal control of (smooth) eye movements under uncertainty. Specifically, we consider delays in the visuo-oculomotor loop and their implications for active inference. Active inference uses a generalisation of Kalman filtering to provide Bayes optimal estimates of hidden states and action in generalised coordinates of motion. Representing hidden states in generalised coordinates provides a simple way of compensating for both sensory and oculomotor delays. The efficacy of this scheme is illustrated using neuronal simulations of pursuit initiation responses, with and without compensation. We then consider an extension of the generative model to simulate smooth pursuit eye movements-in which the visuo-oculomotor system believes both the target and its centre of gaze are attracted to a (hidden) point moving in the visual field. Finally, the generative model is equipped with a hierarchical structure, so that it can recognise and remember unseen (occluded) trajectories and emit anticipatory responses. These simulations speak to a straightforward and neurobiologically plausible solution to the generic problem of integrating information from different sources with different temporal delays and the particular difficulties encountered when a system-like the oculomotor system-tries to control its environment with delayed signals.

A robust event-driven approach to always-on object recognition.

We propose a neuromimetic architecture capable of always-on pattern recognition, i.e. at any time during processing. To achieve this, we have extended an existing event-based algorithm (Lagorce et al., 2017), which introduced novel spatio-temporal features as a Hierarchy Of Time-Surfaces (HOTS). Built from asynchronous events captured by a neuromorphic camera, these time surfaces allow to encode the local dynamics of a visual scene and to create an efficient event-based pattern recognition architecture. Inspired by neuroscience, we have extended this method to improve its performance. First, we add a homeostatic gain control on the activity of neurons to improve the learning of spatio-temporal patterns (Grimaldi et al., 2021). We also provide a new mathematical formalism that allows an analogy to be drawn between the HOTS algorithm and Spiking Neural Networks (SNN). Following this analogy, we transform the offline pattern categorization method into an online and event-driven layer. This classifier uses the spiking output of the network to define new time surfaces and we then perform the online classification with a neuromimetic implementation of a multinomial logistic regression. These improvements not only consistently increase the performance of the network, but also bring this event-driven pattern recognition algorithm fully online. The results have been validated on different datasets: Poker-DVS (Serrano-Gotarredona and Linares-Barranco, 2015), N-MNIST (Orchard, Jayawant et al., 2015) and DVS Gesture (Amir et al., 2017). This demonstrates the efficiency of this bio-realistic SNN for ultra-fast object recognition through an event-by-event categorization process.

Ultrafast Image Categorization in Biology and Neural Models.

Humans are able to categorize images very efficiently, in particular to detect the presence of an animal very quickly. Recently, deep learning algorithms based on convolutional neural networks (CNNs) have achieved higher than human accuracy for a wide range of visual categorization tasks. However, the tasks on which these artificial networks are typically trained and evaluated tend to be highly specialized and do not generalize well, e.g., accuracy drops after image rotation. In this respect, biological visual systems are more flexible and efficient than artificial systems for more general tasks, such as recognizing an animal. To further the comparison between biological and artificial neural networks, we re-trained the standard VGG 16 CNN on two independent tasks that are ecologically relevant to humans: detecting the presence of an animal or an artifact. We show that re-training the network achieves a human-like level of performance, comparable to that reported in psychophysical tasks. In addition, we show that the categorization is better when the outputs of the models are combined. Indeed, animals (e.g., lions) tend to be less present in photographs that contain artifacts (e.g., buildings). Furthermore, these re-trained models were able to reproduce some unexpected behavioral observations from human psychophysics, such as robustness to rotation (e.g., an upside-down or tilted image) or to a grayscale transformation. Finally, we quantified the number of CNN layers required to achieve such performance and showed that good accuracy for ultrafast image categorization can be achieved with only a few layers, challenging the belief that image recognition requires deep sequential analysis of visual objects. We hope to extend this framework to biomimetic deep neural architectures designed for ecological tasks, but also to guide future model-based psychophysical experiments that would deepen our understanding of biological vision.

Cortical recurrence supports resilience to sensory variance in the primary visual cortex.

Our daily endeavors occur in a complex visual environment, whose intrinsic variability challenges the way we integrate information to make decisions. By processing myriads of parallel sensory inputs, our brain is theoretically able to compute the variance of its environment, a cue known to guide our behavior. Yet, the neurobiological and computational basis of such variance computations are still poorly understood. Here, we quantify the dynamics of sensory variance modulations of cat primary visual cortex neurons. We report two archetypal neuronal responses, one of which is resilient to changes in variance and co-encodes the sensory feature and its variance, improving the population encoding of orientation. The existence of these variance-specific responses can be accounted for by a model of intracortical recurrent connectivity. We thus propose that local recurrent circuits process uncertainty as a generic computation, advancing our understanding of how the brain handles naturalistic inputs.

Motion clouds: model-based stimulus synthesis of natural-like random textures for the study of motion perception.

Choosing an appropriate set of stimuli is essential to characterize the response of a sensory system to a particular functional dimension, such as the eye movement following the motion of a visual scene. Here, we describe a framework to generate random texture movies with controlled information content, i.e., Motion Clouds. These stimuli are defined using a generative model that is based on controlled experimental parametrization. We show that Motion Clouds correspond to dense mixing of localized moving gratings with random positions. Their global envelope is similar to natural-like stimulation with an approximate full-field translation corresponding to a retinal slip. We describe the construction of these stimuli mathematically and propose an open-source Python-based implementation. Examples of the use of this framework are shown. We also propose extensions to other modalities such as color vision, touch, and audition.

Motion-based prediction is sufficient to solve the aperture problem.

In low-level sensory systems, it is still unclear how the noisy information collected locally by neurons may give rise to a coherent global percept. This is well demonstrated for the detection of motion in the aperture problem: as luminance of an elongated line is symmetrical along its axis, tangential velocity is ambiguous when measured locally. Here, we develop the hypothesis that motion-based predictive coding is sufficient to infer global motion. Our implementation is based on a context-dependent diffusion of a probabilistic representation of motion. We observe in simulations a progressive solution to the aperture problem similar to physiology and behavior. We demonstrate that this solution is the result of two underlying mechanisms. First, we demonstrate the formation of a tracking behavior favoring temporally coherent features independent of their texture. Second, we observe that incoherent features are explained away, while coherent information diffuses progressively to the global scale. Most previous models included ad hoc mechanisms such as end-stopped cells or a selection layer to track specific luminance-based features as necessary conditions to solve the aperture problem. Here, we have proved that motion-based predictive coding, as it is implemented in this functional model, is sufficient to solve the aperture problem. This solution may give insights into the role of prediction underlying a large class of sensory computations.

A behavioral receptive field for ocular following in monkeys: Spatial summation and its spatial frequency tuning.

In human and non-human primates, reflexive tracking eye movements can be initiated at very short latency in response to a rapid shift of the image. Previous studies in humans have shown that only a part of the central visual field is optimal for driving ocular following responses. Herein, we have investigated spatial summation of motion information across a wide range of spatial frequencies and speeds of drifting gratings by recording short-latency ocular following responses in macaque monkeys. We show that optimal stimulus size for driving ocular responses cover a small (<20° diameter), central part of the visual field that shrinks with higher spatial frequency. This signature of linear motion integration remains invariant with speed and temporal frequency. For low and medium spatial frequencies, we found a strong suppressive influence from surround motion, evidenced by a decrease of response amplitude for stimulus sizes larger than optimal. Such suppression disappears with gratings at high frequencies. The contribution of peripheral motion was investigated by presenting grating annuli of increasing eccentricity. We observed an exponential decay of response amplitude with grating eccentricity, the decrease being faster for higher spatial frequencies. Weaker surround suppression can thus be explained by sparser eccentric inputs at high frequencies. A Difference-of-Gaussians model best renders the antagonistic contributions of peripheral and central motions. Its best-fit parameters coincide with several, well-known spatial properties of area MT neuronal populations. These results describe the mechanism by which central motion information is automatically integrated in a context-dependent manner to drive ocular responses.Significance statementOcular following is driven by visual motion at ultra-short latency in both humans and monkeys. Its dynamics reflect the properties of low-level motion integration. Here, we show that a strong center-surround suppression mechanism modulates initial eye velocity. Its spatial properties are dependent upon visual inputs' spatial frequency but are insensitive to either its temporal frequency or speed. These properties are best described with a Difference-of-Gaussian model of spatial integration. The model parameters reflect many spatial characteristics of motion sensitive neuronal populations in monkey area MT. Our results further outline the computational properties of the behavioral receptive field underpinning automatic, context-dependent motion integration.

Precise Spiking Motifs in Neurobiological and Neuromorphic Data.

Why do neurons communicate through spikes? By definition, spikes are all-or-none neural events which occur at continuous times. In other words, spikes are on one side binary, existing or not without further details, and on the other, can occur at any asynchronous time, without the need for a centralized clock. This stands in stark contrast to the analog representation of values and the discretized timing classically used in digital processing and at the base of modern-day neural networks. As neural systems almost systematically use this so-called event-based representation in the living world, a better understanding of this phenomenon remains a fundamental challenge in neurobiology in order to better interpret the profusion of recorded data. With the growing need for intelligent embedded systems, it also emerges as a new computing paradigm to enable the efficient operation of a new class of sensors and event-based computers, called neuromorphic, which could enable significant gains in computation time and energy consumption-a major societal issue in the era of the digital economy and global warming. In this review paper, we provide evidence from biology, theory and engineering that the precise timing of spikes plays a crucial role in our understanding of the efficiency of neural networks.

Functional consequences of correlated excitatory and inhibitory conductances in cortical networks.

Neurons in the neocortex receive a large number of excitatory and inhibitory synaptic inputs. Excitation and inhibition dynamically balance each other, with inhibition lagging excitation by only few milliseconds. To characterize the functional consequences of such correlated excitation and inhibition, we studied models in which this correlation structure is induced by feedforward inhibition (FFI). Simple circuits show that an effective FFI changes the integrative behavior of neurons such that only synchronous inputs can elicit spikes, causing the responses to be sparse and precise. Further, effective FFI increases the selectivity for propagation of synchrony through a feedforward network, thereby increasing the stability to background activity. Last, we show that recurrent random networks with effective inhibition are more likely to exhibit dynamical network activity states as have been observed in vivo. Thus, when a feedforward signal path is embedded in such recurrent network, the stabilizing effect of effective inhibition creates an suitable substrate for signal propagation. In conclusion, correlated excitation and inhibition support the notion that synchronous spiking may be important for cortical processing.

Role of homeostasis in learning sparse representations.

Neurons in the input layer of primary visual cortex in primates develop edge-like receptive fields. One approach to understanding the emergence of this response is to state that neural activity has to efficiently represent sensory data with respect to the statistics of natural scenes. Furthermore, it is believed that such an efficient coding is achieved using a competition across neurons so as to generate a sparse representation, that is, where a relatively small number of neurons are simultaneously active. Indeed, different models of sparse coding, coupled with Hebbian learning and homeostasis, have been proposed that successfully match the observed emergent response. However, the specific role of homeostasis in learning such sparse representations is still largely unknown. By quantitatively assessing the efficiency of the neural representation during learning, we derive a cooperative homeostasis mechanism that optimally tunes the competition between neurons within the sparse coding algorithm. We apply this homeostasis while learning small patches taken from natural images and compare its efficiency with state-of-the-art algorithms. Results show that while different sparse coding algorithms give similar coding results, the homeostasis provides an optimal balance for the representation of natural images within the population of neurons. Competition in sparse coding is optimized when it is fair. By contributing to optimizing statistical competition across neurons, homeostasis is crucial in providing a more efficient solution to the emergence of independent components.

An Adaptive Homeostatic Algorithm for the Unsupervised Learning of Visual Features.

The formation of structure in the visual system, that is, of the connections between cells within neural populations, is by and large an unsupervised learning process. In the primary visual cortex of mammals, for example, one can observe during development the formation of cells selective to localized, oriented features, which results in the development of a representation in area V1 of images' edges. This can be modeled using a sparse Hebbian learning algorithms which alternate a coding step to encode the information with a learning step to find the proper encoder. A major difficulty of such algorithms is the joint problem of finding a good representation while knowing immature encoders, and to learn good encoders with a nonoptimal representation. To solve this problem, this work introduces a new regulation process between learning and coding which is motivated by the homeostasis processes observed in biology. Such an optimal homeostasis rule is implemented by including an adaptation mechanism based on nonlinear functions that balance the antagonistic processes that occur at the coding and learning time scales. It is compatible with a neuromimetic architecture and allows for a more efficient emergence of localized filters sensitive to orientation. In addition, this homeostasis rule is simplified by implementing a simple heuristic on the probability of activation of neurons. Compared to the optimal homeostasis rule, numerical simulations show that this heuristic allows to implement a faster unsupervised learning algorithm while retaining much of its effectiveness. These results demonstrate the potential application of such a strategy in machine learning and this is illustrated by showing the effect of homeostasis in the emergence of edge-like filters for a convolutional neural network.

Speed-Selectivity in Retinal Ganglion Cells is Sharpened by Broad Spatial Frequency, Naturalistic Stimuli.

Motion detection represents one of the critical tasks of the visual system and has motivated a large body of research. However, it remains unclear precisely why the response of retinal ganglion cells (RGCs) to simple artificial stimuli does not predict their response to complex, naturalistic stimuli. To explore this topic, we use Motion Clouds (MC), which are synthetic textures that preserve properties of natural images and are merely parameterized, in particular by modulating the spatiotemporal spectrum complexity of the stimulus by adjusting the frequency bandwidths. By stimulating the retina of the diurnal rodent, Octodon degus with MC we show that the RGCs respond to increasingly complex stimuli by narrowing their adjustment curves in response to movement. At the level of the population, complex stimuli produce a sparser code while preserving movement information; therefore, the stimuli are encoded more efficiently. Interestingly, these properties were observed throughout different populations of RGCs. Thus, our results reveal that the response at the level of RGCs is modulated by the naturalness of the stimulus - in particular for motion - which suggests that the tuning to the statistics of natural images already emerges at the level of the retina.