Simon Brodeur

Network-selectivity and stimulus-discrimination in the primary visual cortex: cell-assembly dynamics.

Visual neurons coordinate their responses in relation to the stimulus; however, the complex interplay between a stimulus and the functional dynamics of an assembly still eludes neuroscientists. To this aim, we recorded cell assemblies from multi‐electrodes in the primary visual cortex of anaesthetized cats in response to randomly presented sine‐wave drifting gratings whose orientation tilted in 22.5° steps. Cross‐correlograms revealed the functional connections at all the tested orientations. We show that a cell‐assembly discriminates between orientations by recruiting a ‘salient’ functional network at every presented orientation, wherein the connections and their strengths (peak‐probabilities in the cross‐correlogram) change from one orientation to another. Within these assemblies, closely tuned neurons exhibited increased connectivity and connection‐strengths compared with differently tuned neurons. Minimal connectivity between untuned neurons suggests the significance of neuronal selectivity in assemblies. This study reflects upon the dynamics of functional connectivity, and brings to the fore the importance of a ‘signature’ functional network in an assembly that is strictly related to a specific stimulus. It appears that an assembly is the major ‘functional unit’ of information processing in cortical circuits, rather than the individual neurons.

Regulation toward self-organized criticality in a recurrent spiking neural reservoir

Generating stable yet performant spiking neural reservoirs for classification applications is still an open issue. This is due to the extremely non-linear dynamics of recurrent spiking neural networks. In this perspective, a local and unsupervised learning rule that tunes the reservoir toward self-organized criticality is proposed, and applied to networks of leaky integrate-and-fire neurons with random and small-world topologies. Longer sustained activity for both topologies was elicited after learning compared to spectral radius normalization (global rescaling scheme). The ability to control the desired regime of the reservoir was shown and quick convergence toward it was observed for speech signals. Proposed regulation method can be applied online and leads to reservoirs more strongly adapted to the task at hand.

CREATE: Multimodal Dataset for Unsupervised Learning, Generative Modeling and Prediction of Sensory Data from a Mobile Robot in Indoor Environments.

The CREATE database is composed of 14 hours of multimodal recordings from a mobile robotic platform based on the iRobot Create. The various sensors cover vision, audition, motors and proprioception. The dataset has been designed in the context of a mobile robot that can learn multimodal representations of its environment, thanks to its ability to navigate the environment. This ability can also be used to learn the dependencies and relationships between the different modalities of the robot (e.g. vision, audition), as they reflect both the external environment and the internal state of the robot. The provided multimodal dataset is expected to have multiple usages, such as multimodal unsupervised object learning, multimodal prediction and egomotion/causality detection.

Proposal of a Generative Model of Event-based Representations for Grounded Language Understanding

Grounding is the problem of correspondence between the symbolic concepts of language and the physical environment. The research direction that we propose to tackle language acquisition and grounding is based on multimodal event-based representations and probabilistic generative modeling. First, we establish a new multimodal dataset recorded from a mobile robot and describe how such multimodal signals can be efficiently encoded into compact, event-based representations using sparse coding. We highlight how they could be better suited to ground concepts. We then describe a generative probabilistic model based on those event-based representations. We discuss possible applications of this probabilistic framework in the context of a cognitive agent, such as detecting novelty at the inputs or reasoning by building internal simulations of the environment. While this work is still in progress, this could open new perspectives on how representational learning can play a key role in the ability to map structures of the multimodal scene to language.

Optimality of inference in hierarchical coding for distributed object-based representations

Hierarchical approaches for representation learning have the ability to encode relevant features at multiple scales or levels of abstraction. However, most hierarchical approaches exploit only the last level in the hierarchy, or provide a multiscale representation that holds a significant amount of redundancy. We argue that removing redundancy across the multiple levels of abstraction is important for an efficient representation of compositionality in object-based representations. With the perspective of feature learning as a data compression operation, we propose a new greedy inference algorithm for hierarchical sparse coding. Convolutional matching pursuit with a L0-norm constraint was used to encode the input signal into compact and non-redundant codes distributed across levels of the hierarchy. Simple and complex synthetic datasets of temporal signals were created to evaluate the encoding efficiency and compare with the theoretical lower bounds on the information rate for those signals. Empirical evidence have shown that the algorithm is able to infer near-optimal codes for simple signals. However, it failed for complex signals with strong overlapping between objects. We explain the inefficiency of convolutional matching pursuit that occurred in such case. This brings new insights about the NP-hard optimization problem related to using L0-norm constraint in inferring optimally compact and distributed object-based representations.

Auditory object feature maps with a hierarchical network of independent components

Auditory Object representation in the brain is still a controversial question [1,2]. Kumar et al. [3] discuss the hierarchical organization for auditory object perception and observe that the Planum Temporale (PT) area of the cortex encodes invariant representations of the spectral envelops of sounds. Many other studies find maps of representations elsewhere in the brain (Cochlear Nucleus, Inferior Nucleus, etc.). Sparse representations with minimum overlap could be considered, according to Barlow [4]. Griffiths and Warren [5] propose that auditory object representations might be segmented or segregated in the Planum Temporal (PT) by increasing the independence between the neural activities. We therefore explore the potential of a hierarchical neural assembly - with the use of a computer simulation - whose layers increase the feature independence during training, to represent auditory object parts. It is observed that learned features are organized into non-overlapping maps (Figure ​(Figure1)1) and that redundancy of the representation is in fact reduced. Learning was done on three categories of sounds having distinct acoustical statistics: speech, music and natural sounds. We observed that the learned feature maps are very different from one sound category to another and might be, to some extend, comparable to receptive fields measured in the brain. We discuss of their potential similarity with receptive fields measured in the Inferior Colliculus of the Guinea Pig and how they might be part of a representation of auditory objects in the brain. Figure 1 Learned representations illustrated for speech and natural sounds. FastICA training is first done on patches of 80ms x 32 cochlear channels of the envelopes coming from a 128 channels cochleagram (Level L0). Then, patches of 160ms x 64 cochlear channels ...

Robust Hierarchical and Sparse Representation of Natural Sounds in High-Dimensional Space

Based on general findings from the field of neuroscience and their algorithmic implementations using signal processing, information theory and machine learning techniques, this paper highlights the advantages of modelling a signal in a sparse and high-dimensional feature space. The emphasis is put on the hierarchical organisation, very high dimensionality and sparseness aspects of auditory information, that allow unsupervised learning of meaningful auditory objects from simple linear projections. When the dictionaries are learned using independent component analysis (ICA), it is shown that specific spectro-temporal modulation patterns are learned to optimally represent speech, noise and tonal components. In a noisy isolated-word speech recognition task, sparse and high-dimensional features have shown greater robustness to noise compared to a standard system based on a dense low-dimensional feature space. This brings new ways of thinking in the field of recognition and classification of acoustic signals.