… thanks to high-order associations

The Unified Radical Associationism proposed by Arnaud Rey points to the question of how complex cognition can be understood in terms of associative learning at the level of synapses. But are associations sufficient forms of knowledge to account for complex processing of information? Complex cognition such as language comprehension or selection of behavior is reported to rely on pairwise transitional probability that can be learned thanks to Hebbian-like learning rules. Cortical network architectures embedding matrices of pairwise associations have proven successful to account for a rich phenomenology of cognitive processes, in line with the Unified Radical Associationism proposed by Arnaud Rey, which states that cognition can be explained by the functioning of large networks of knowledge structured by associations, themselves learned by a Hebbian-like rule. However, some forms of knowledge involve not only pairs but also patterns of more than two stimuli and/or actions. Behavioral responses paired to a given stimulus also depend on other stimuli, motivations and goals. Such context-dependent activation requires learning of high-order n-wise relations between more than two items. This involves second-order transitional probability that can be learned by multi-layered deep architectures and also by less structured recurrent cortical networks that use a multisynaptic learning rule. This rule is biologically realistic and compatible with Hebbian learning principles. It allows a context to dynamically select a subset of pairs that can activate each other among all possible pairs. It then make possible to explain how (somewhat) complex knowledge can be learned and processed, thus supporting the Unified Radical Associationism proposed by Arnaud Rey: “Associations are all we need”? thanks to high-order associations.