Invariant Information Spike Sorting
Abstract
Spike sorting is the process of identifying and classifying voltage recordings from the brain or nervous system into discrete labelled waveform events. The core difficulty lies in unsupervised classification - one does not have definitive labels for signals, or how many unique labels there are per recording. Recent works within the field have converged on a prevalent architecture for approaching classification: feature extraction (FE) followed by a traditional clustering algorithm (CA). While there is unanimity in architecture, there is ambiguity as to what techniques to use for a given problem due to: inconsistency across datasets, arbitrary parameterization, abstract representation, and no standard dataset for sorting n independent waveforms. Given these issues, our goal was to challenge typical architecture with a deep learning based approach. It is worth noting that there have already been attempts at adding deep learning to spike sorting that act as extensions on prior methods. We propose to remove FE entirely by extending Invariant Information Clustering (IIC) - a method built for image classification - to spike sorting; thus creating Invariant Information Spike Sorting (IISS). IISS uses a physics inspired transform Φ, such as background noise addition, to create paired spike data [x, Φ(x)] where one is a plausible facsimile of the other. A neural network learns to predict identical δ distributions for paired spikes by maximizing mutual information between pair’s predicted classes. Clusters emerge after learning core semantics, ensuring inter-cluster variance surpasses Φ induced differences. To test our model against classical clustering approaches we developed 4 simulation waveform datasets that mimic real data taken from the peripheral nervous system. To steel-man against IISS we compared 18 possible pairings of FE+CA solutions given optimal parameterization - a highly improbable event - while using a single parameter setting for IISS. We find IISS comparable or superior across datasets. The parameter choices for IISS are intuitive and stable. The final design can be seen as a first draft with substantial scope for enhancements. Consequently IISS demonstrated capacity to supersede traditional methods and paves the way for more intuitive, robust, physics-grounded spike sorting.