The project Action-perception loop and predictive codding in 1d binary world is started.
Cell assemblies paper is published on arXive. It's our first officially published paper. In this paper, we address excitation-inhibition balance in winners-take-all neural networks.
@Danylo Ulianych started working in the Institute of Neuroscience and Medicine at Jülich Forschungszentrum, Germany. Institute's research topics cover biological neural networks, in-vivo experiments, and electrophysiological analyses. KyivAIGroup meetings took a break for two years. Danylo was to come back in Mar 2021.
@Osaulenko Viacheslav and @Danylo Ulianych shared theory and practical knowledge in the second workshop https://www.facebook.com/events/2086228955040058/ "Math properties of brain codes". The workshop became ****a compilation of our past work.
@Danylo Ulianych held a workshop https://www.facebook.com/events/2086228955040058/ on Information Theory in Machine Learning and Neuroscience.
@Osaulenko Viacheslav gave a lecture https://www.facebook.com/events/295457637946608/ entitled "2018: where we stand in brain science".
Viacheslav Dudar joined the team.
Previously engaged in the Hierarchical Temporal Memory framework developed in Numenta, we stopped the experiments with this framework. The conclusions we draw - inability to predict the future events N steps ahead without recursively running the algorithm N times (Temporal Integration), encoding the next character in a sequence solely in the context of the previous characters ("BANANA" and "ANANA" will result in completely different encodings), the lack of topological connections (crucial for images) - are also stated in Numeta's HTM online resources. We opted for a different platform to avoid these drawbacks.
The first results are assembled in the paper Active Vision [with paper]. This paper has never been published - we erroneously thought that we shouldn't publish papers with negative results.