Seminar, Hosted by NeurIPS 2019 · Real Neurons & Hidden Units
Eligibility Traces Provide a Data-Inspired Alternative to Backpropagation Through Time
Guillaume Bellec and Franz Scherr
Graz University of Technology
Abstract
Guillaume Bellec and Franz Scherr present eligibility propagation, or e-prop, a way to train recurrent networks by combining traces of local activity with learning signals. The gradient factorization yields a three-factor plasticity rule for spiking neuron models. On a speech-recognition benchmark, e-prop was competitive with backpropagation through time for both spiking recurrent networks and long short-term memory networks. The authors also examine how diverse learning signals and slow neural dynamics contribute to performance.
Topics
- Eligibility traces, search World Wide
- eligibility propagation
- e-prop
- backpropagation through time
- recurrent networks
- spiking neuron models
- three-factor plasticity rule, search World Wide
- speech-recognition benchmark, search World Wide
- Long short-term memory networks, broader topic: networks
- learning signals
- slow neural dynamics, broader topic: neural dynamics