Speaker: Yanis Bahroun
Topic: NeurIPS submission: A neurally plausible minor subspace analyzer balances feedforward excitation and inhibition.
Most neural circuits in the brain include inhibitory neurons in various configurations such as feedforward inhibition (FFI). Despite much research, the specific computational function of FFI remains elusive. Here, we suggest that FFI may participate in extracting the lowest-variance subspace, or minor subspace analysis (MSA), from the signals communicated by upstream neurons. Starting from a novel similarity-preserving objective function, we derive an online algorithm for MSA that naturally maps onto a neural network (NN) with biologically plausible learning rules. This NN comprises FFI and multi-compartment neurons. The resulting algorithm also outperforms existing online MSA algorithms in terms of computational complexity, thus addressing the critical challenge of building both biologically plausible and computationally efficient learning algorithms.
Join from PC, Mac, Linux, iOS or Android: https://simonsfoundation.zoom.us/j/732630308?pwd=U29ZTmFzZzB4WEg4cnpCZitqYlZqUT09