Plan seminarów pod patronatem Sekcji Nauk Obliczeniowych i Bioinformatyki KI PAN


Włodzisław Duch

Brain-inspired cognitive computing

Komisja Informatyki i Automatyki, Oddział PAN w Poznaniu, WFAiIS UMK w Toruniu i WTIE UTP w Bydogszczy

Abstract: Neural networks are based on simple inspirations dating back to threshold neurons of McCulloch and Pitts, and pandemonium architecture of Selfridge, that later was developed into multilayer perceptrons (MLPs) and deep architecture models. However, backpropagation and other common forms of learning are hard to justify from neurobiological point of view. In this talk I would like to show that understanding brain networks give us more inspirations that have not been fully explored. Simple neurons have one internal parameter (activation threshold) and fixed synaptic connections. Cortical columns contain microcircuits that provide modules with many internal parameters that can be modeled by complex transfer functions. Larger cortical structures are capable of sophisticated processing of signals, showing flexible functional interactions, activated by priming, controlled by working memory, not restricted to fixed structural connections. This leads to attractor neurodynamics that engage many cortical ensembles, solving novel combinatorial problems. Finally at the whole brain level different tasks are solved by 'the society of mind' (as Marvin Minsky has called it), temporal binding of modules containing a lot of knowledge, without fixed connections. Our recent fMRI experiments show how learning leads to dynamical reconfiguration of functional brain networks. Societies of brains, or multi-agent systems, interacting on a symbolic level can solve problems at even higher level. Many algorithms derived from this line of thinking will be mentioned, including our work on transfer functions, heterogeneous systems, almost random projections, feature space mapping model, similarity based systems, prototype-based rules, k-separability and projection pursuit networks, support feature machines and meta-learning schemes.

For papers and links on this topic see:

Privacy policy

Template by: CssTemplateHeaven