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.

Active Fovea-Based Vision Through Computationally-Effective Model-Based Prediction.

What motivates an action in the absence of a definite reward? Taking the case of visuomotor control, we consider a minimal control problem that is how select the next saccade, in a sequence of discrete eye movements, when the final objective is to better interpret the current visual scene. The visual scene is modeled here as a partially-observed environment, with a generative model explaining how the visual data is shaped by action. This allows to interpret different action selection metrics proposed in the literature, including the Salience, the Infomax and the Variational Free Energy, under a single information theoretic construct, namely the view-based Information Gain. Pursuing this analytic track, two original action selection metrics named the Information Gain Lower Bound (IGLB) and the Information Gain Upper Bound (IGUB) are then proposed. Showing either a conservative or an optimistic bias regarding the Information Gain, they strongly simplify its calculation. An original fovea-based visual scene decoding setup is then proposed, with numerical experiments highlighting different facets of artificial fovea-based vision. A first and principal result is that state-of-the-art recognition rates are obtained with fovea-based saccadic exploration, using less than 10% of the original image's data. Those satisfactory results illustrate the advantage of mixing predictive control with accurate state-of-the-art predictors, namely a deep neural network. A second result is the sub-optimality of some classical action-selection metrics widely used in the literature, that is not manifest with finely-tuned inference models, but becomes patent when coarse or faulty models are used. Last, a computationally-effective predictive model is developed using the IGLB objective, with pre-processed visual scan-path read-out from memory, bypassing computationally-demanding predictive calculations. This last simplified setting is shown effective in our case, showing both a competing accuracy and a good robustness to model flaws.

Multistability in Large Scale Models of Brain Activity.

Noise driven exploration of a brain network's dynamic repertoire has been hypothesized to be causally involved in cognitive function, aging and neurodegeneration. The dynamic repertoire crucially depends on the network's capacity to store patterns, as well as their stability. Here we systematically explore the capacity of networks derived from human connectomes to store attractor states, as well as various network mechanisms to control the brain's dynamic repertoire. Using a deterministic graded response Hopfield model with connectome-based interactions, we reconstruct the system's attractor space through a uniform sampling of the initial conditions. Large fixed-point attractor sets are obtained in the low temperature condition, with a bigger number of attractors than ever reported so far. Different variants of the initial model, including (i) a uniform activation threshold or (ii) a global negative feedback, produce a similarly robust multistability in a limited parameter range. A numerical analysis of the distribution of the attractors identifies spatially-segregated components, with a centro-medial core and several well-delineated regional patches. Those different modes share similarity with the fMRI independent components observed in the "resting state" condition. We demonstrate non-stationary behavior in noise-driven generalizations of the models, with different meta-stable attractors visited along the same time course. Only the model with a global dynamic density control is found to display robust and long-lasting non-stationarity with no tendency toward either overactivity or extinction. The best fit with empirical signals is observed at the edge of multistability, a parameter region that also corresponds to the highest entropy of the attractors.

Computational neuroscience, from multiple levels to multi-level.

Despite the long and fruitful history of neuroscience, a global, multi-level description of cardinal brain functions is still far from reach. Using analytical or numerical approaches, Computational Neuroscience aims at the emergence of such common principles by using concepts from Dynamical Systems and Information Theory. The aim of this Special Issue of the Journal of Physiology (Paris) is to reflect the latest advances in this field which has been presented during the NeuroComp08 conference that took place in October 2008 in Marseille (France). By highlighting a selection of works presented at the conference, we wish to illustrate the intrinsic diversity of this field of research but also the need of an unification effort that is becoming more and more necessary to understand the brain in its full complexity, from multiple levels of description to a multi-level understanding.

Resonant spatiotemporal learning in large random recurrent networks.

Taking a global analogy with the structure of perceptual biological systems, we present a system composed of two layers of real-valued sigmoidal neurons. The primary layer receives stimulating spatiotemporal signals, and the secondary layer is a fully connected random recurrent network. This secondary layer spontaneously displays complex chaotic dynamics. All connections have a constant time delay. We use for our experiments a Hebbian (covariance) learning rule. This rule slowly modifies the weights under the influence of a periodic stimulus. The effect of learning is twofold: (i) it simplifies the secondary-layer dynamics, which eventually stabilizes to a periodic orbit; and (ii) it connects the secondary layer to the primary layer, and realizes a feedback from the secondary to the primary layer. This feedback signal is added to the incoming signal, and matches it (i.e., the secondary layer performs a one-step prediction of the forthcoming stimulus). After learning, a resonant behavior can be observed: the system resonates with familiar stimuli, which activates a feedback signal. In particular, this resonance allows the recognition and retrieval of partial signals, and dynamic maintenance of the memory of past stimuli. This resonance is highly sensitive to the temporal relationships and to the periodicity of the presented stimuli. When we present stimuli which do not match in time or space, the feedback remains silent. The number of different stimuli for which resonant behavior can be learned is analyzed. As with Hopfield networks, the capacity is proportional to the size of the second, recurrent layer. Moreover, the high capacity displayed allows the implementation of our model on real-time systems interacting with their environment. Such an implementation is reported in the case of a simple behavior-based recognition task on a mobile robot. Finally, we present some functional analogies with biological systems in terms of autonomy and dynamic binding, and present some hypotheses on the computational role of feedback connections.

Self-organization and dynamics reduction in recurrent networks: stimulus presentation and learning.

Freeman's investigations on the olfactory bulb of the rabbit showed that its signal dynamics was chaotic, and that recognition of a learned stimulus is linked to a dimension reduction of the dynamics attractor. In this paper we address the question whether this behavior is specific of this particular architecture, or if it is a general property. We study the dynamics of a non-convergent recurrent model-the random recurrent neural networks. In that model a mean-field theory can be used to analyze the autonomous dynamics. We extend this approach with various observations on significant changes in the dynamical regime when sending static random stimuli. Then we propose a Hebb-like learning rule, viewed as a self-organization dynamical process inducing specific reactivity to one random stimulus. We numerically show the dynamics reduction during learning and recognition processes and analyze it in terms of dynamical repartition of local neural activity.