Robert Gütig: Margin learning in spiking neurons

Abstract

 Learning novel sensory features from few examples is a remarkable ability of humans and other animals. For example, we can recognize unfamiliar faces or words after seeing or hearing them only a few times, even in different contexts and noise levels. Previous work has shown that spiking neural networks can learn to detect unknown features in unsegmented input streams using multi-spike tempotron learning. However, this method requires many training patterns and the learned solutions can be sensitive to noise.

 In this work, we use multi-spike tempotron learning to implement margin learning in spiking neurons. Specifically, we introduce regularization terms that enable leaky-integrate-and-fire neurons to learn to detect recurring features using orders of magnitude less training data and converge to robust solutions. We test the novel learning rule on unsegmented spoken digit sequences contained in the TIDIGITS speech data set and find a twofold improvement in detection probability over the original learning algorithm. Our work shows how neurons can learn to detect embedded features from a limited number of unsegmented samples, provides fundamental bounds for the noise robustness of the leaky integrate-and-fire model and ties mathematically principled gradient-based optimization to biologically plausible learning in spiking neurons.  

About the speaker and his research

Hosted by Christian Leipold