Spike-based neural computation and learning in the time domain: Applications to auditory and visual processing
Final Report Abstract
Understanding the role of action potential timing in neural information processing and learning has challenged brain researchers for over half a century. Despite advances in sensory system electrophysiology that have suggested that the precise timing of spikes can carry substantial information about eliciting stimuli, little understanding has been gained with respect to how downstream processing stages can decode such spike-timing based neuronal representations. In this research project we used our recently proposed spike-based synaptic learning rule, the tempotron, to study the function of spiking neural networks engaged in biologically realistic visual and auditory processing tasks. We progressed along three lines of research that were aimed to jointly bridge the gap between basic biophysical intracellular processes and higher-level perceptual functions. We successfully addressed several longstanding problems in computational and systems neuroscience: • We showed how conductance-based rescaling of the effective integration time constant of sensory cells could lead to time-warp invariant neuronal processing. By applying this mechanism to a model of neuronal speech recognition we showed that its operation is biologically feasible and that already small neuronal networks can match the performance of sopisticated state-of-the-art speech recognition systems. • We derived a supervised spike-based learning rule for synapses that express short-term plasticity. Using this model we resolved the reset problem by aligning the dynamics of synapses within afferent populations and showed that synaptic dynamics can allow neurons to learn and utilize synaptic representations of time. • Using ensembles of real retinal spike trains we showed that neuronal decoders that utilize spike-timing based sensory information outperform firing-rate based networks on several image classification tasks. Importantly, this mode of neuronal processing can operate with very few spikes and thus explain the astonishing speed of high-level vision. In addition, spike-timing based visual representations allowed us to explain important perceptual invariances to contrast and phase, which are considered difficult computational problems in established rate coding frameworks and require multi-layer architectures. • We succeeded in generalizing the binary tempotron rule such that it allows neurons to learn an arbitrary number of output spikes. With this novel synaptic learning rule we could show that simple neuronal models can readily learn to process sensory objects even if they are embedded in streams of random background noise and no temporal information is made available during learning.
Publications
-
(2009). Temporal processing with plastic short term synaptic dynamics. Computational and Systems Neuroscience, Salt Lake City, USA
Gütig R, Sompolinsky H & Tsodyks M
-
(2009). Time-warp-invariant neuronal processing. PLoS Biology 7, e1000141-e1000141
Gütig R & Sompolinsky H
-
(2012). The multi-class tempotron: a neuron model for processing of sensory streams. Computational and Systems Neuroscience, Salt Lake City, USA
Gütig R
