Reliably mapping e-prop gradients to memristor conductance updates under non-idealities

Establish reliable weight-update schemes that translate e-prop learning signals into conductance changes of Phase-Change Memory (PCM) devices in the presence of programming noise, read noise, drift, and limited bit precision.

Background

The thesis proposes using the e-prop local learning rule to train recurrent spiking neural networks with PCM-based synapses. PCM devices exhibit non-idealities such as stochastic write/read behavior and temporal drift, which complicate direct gradient-to-conductance mapping.

This open question highlights the need for robust update mechanisms that can faithfully implement gradient-based learning on non-ideal analog substrates.