Galiya M. Markova, Sergey I. Bartsev

Institute of biophysics Siberian Branch of RAS, School of fundamental biology and biotechnology, Siberian federal university

The paper presents results on modeling reflection, understood in a broad sense as the presence of an internal representation of the external world in an active agent that influences its behavior. The ability of the simplest neural network model objects of homogeneous and heterogeneous (modular) structure to solve tasks requiring the presence and use of stable internal representations of external stimuli is revealed. It is determined that these representations are decodable, i.e. based on the current type of neural activity pattern of a neural network, it is possible to determine which specific stimulus or time series of stimuli is currently being processed in it. The authors' initial assumptions made on the basis of general considerations regarding the effectiveness of neural networks of various structures in reflection tasks and the corresponding results are presented. In particular, the following effects are shown: 1) positions in the odd-even game are asymmetric under the condition of limited computational capabilities of the players; 2) formally similar tasks on reflection (the odd-even game and responding to fixed time series of stimuli according to the rules of this game) differ in the requirements for players; 3) decodable patterns of neural activity present not only in neural networks trained to respond to stimuli, but also in networks with random weight coefficients; 4) the accuracy of decoding the neural activity of recurrent neural networks with temporal heterogeneity exceeds the accuracy of the response of these networks when processing series of stimuli; 5) patterns of neural activity in homogeneous recurrent neural networks are more difficult to decode than in heterogeneous networks of comparable size. These effects illustrate the rich internal and behavioral dynamics of the simplest recurrent neural networks, which, on the one hand, is promising for research and practical purposes, and on the other hand, complicates the prediction and interpretation of their behavior.

recurrent neural networks, reflection, reflexive games, neural activity decoding, representation of external stimuli