Low-dimensional mutliplexed representations in large recurrent neural networks: illustrations in supervised and unsupervised learning.

Recurrent Neural Networks (RNN) are powerful tools to learn computational tasks from data. How the tasks to be performed are simultaneously encoded and the input/output data are represented in a high-dimensional network is an important question. I will consider two related problems, both inspired from computational neuroscience: (1) how multiple low-dimensional maps (environments) can be embedded in a RNN, (2) how multiplexed integrations of velocity signals can be carried out in the RNN to update positions in those maps. I will discuss the nature of the representations found in artificial RNN, and compare them to experimental recordings in the mammal brain.