Final Report Summary - NEUCOD (Neural coding, specification, design and test of message passing neural machines) The Neucod project was initiated on the basis of the observation that the brain offers remarkable properties of error correction. Whether it is about recognizing partially erased or distorted images, quickly understanding a word in which some letters have been mixed or recalling a precise fact from some vague clues, strongly redundant encoding is necessary to ensure such robustness. This redundant code is indispensable to combat the very noisy behaviour of neurons (unguaranteed availability of neurotransmitter or receptor in synaptic contacts at spike arrivals, ionic channel current fluctuation, etc.) which then cannot be reliable by themselves. Assembly coding, more specifically neural clique, is the model of coding that has permanently been at the centre of our different studies.In the first part of the project, neural cliques and their sequential extensions (tournaments) have been studied in deep detail: their ability to store information (proved quasi-optimal), their tolerance to noise, the different possible recovering algorithms in presence of erasures and errors, their duplication and aggregation, their electronic implementation (that we showed were much more efficient with analog circuits), etc. The properties of neural cliques make them good candidates for the explanation of biological long-term memory. Cliques being digital objects, this would mean that our cerebral memory is digital. Therefore, many questions, which are usual in telecommunications, have arisen about compression rates, coding rate, modulation, network organization, etc and are still under investigation.In collaboration with experts in electroencephalography (EEG), we have carried out several experiments trying to show evidence of the existence of cliques in the cortex. Using novel algorithms able to deal with the inverse problem for thousands of sources, at the millisecond scale, we succeeded in visualizing metacliques (that is, cliques whose vertices belong to different cortical regions and which are supposed to bind local cliques). We were also able to bring to the fore reproducible transfers of activity between cortical regions which allowed us to launch a new project called "neural communication" (now in progress). By simulation, we have shown that assembly coding is quite appropriate for robust information transfer despite the neuronal noise.In the second part of Neucod, the problematic of numerical (statistic) to symbolic conversion was considered. Indeed, whereas a lot of work in Artificial Intelligence (AI) is currently being done on learning algorithms, notably thanks to the success of so-called deep learning, few studies are devoted to the way information acquired by sensorial modules is transferred to the sparse mental (digital) world. In particular, the question of memory organization (and of the underlying principles of multimodality) seemed fundamental to us: what is needed to store locally and what is the level of abstraction of these local patterns? The case of visual attention was considered as a concrete challenging question. Also, to help us provide realistic preliminary answers, we first considered EEG recording and analysis of image recognition (steady stimuli) and are currently doing the same with sound (sequential stimuli). Theoretical models were also proposed to describe the aggregation process in associative regions of the brain (hubs).In the wide pluridisciplinary field of neurosciences, the Neucod project was positioned at the mesoscopic level, dealing with information processing, architecture and circuits in a rather bottom-up approach, from simple binary neuronal components to networks and functions. This research discipline, which we call Informational Neuroscience, is quite recent but is gaining influence on two main directions: brain-inspired AI and cognitive psychology. On the former point, one important application with significant payoff in both personal and professional life will be about natural language processing (automatic translation, sense extraction, data mining, automatic summarization, etc). This will certainly represent a singularity in many professions, especially research. That is why we have launched the design project of the first ever full-connexionnist multimodal language processing machine. On the latter point, understanding the brain at the informational level will help detection and treatment of mental disorders, for instance by the means of cortical network imagery and neuromarkers. These mid-term objectives will likely make use of the work accomplished during the Neucod project.
