Scientists from France have built a mathematical theory of study of living organisms in the processing of sensory information. Their theory is based on recurrent neural networks, extends the classical theory of learning to the case of spatial correlations. Work published in the journal Physical Review Letters.
How sensory information is encoded and processed by neurons in the brain is one of the most interesting problems in computational neuroscience. In many areas of the brain neuronal activity is highly dependent on some touch option: for example, the activity of place cells depends on the head orientation and position of the animal in space. In recent decades artificial neural network with a continuous attractor became an attractive model for explaining the operation of the neurons of the brain in processing spatial information from the outside. Such models can explain how a large neural system can encode information from low-dimensional sensory space and continuously update it over time in accordance with the input signals.
Physics Aldo Batista (Aldo Battista) and Monasson Remy (Remi Monasson) of the Higher Normal School of Paris used a recurrent neural network (RNN) to explain how the attractors can arise from multidimensional dynamics and showed that the RNN is able to store large amounts of spatial information with high resolution.