Sanja Karilanova: Bridging Spiking Neural Networks and Deep State-Space Models

Spiking neural networks (SNNs) are biologically inspired, event-driven models that are suitable for processing temporal data and offer low-latency, energy-efficient computation when implemented on neuromorphic hardware. 

Similar to SNNs, state-space models (SSMs) utilize the concepts of state and recurrence. Recently, deep SSMs have emerged as a promising model family for obtaining competitive performance for a wide range of tasks on long temporal sequences. 

In this talk, we highlight the bridge between SNNs and SSMs; and use this relationship to investigate several temporal modeling and classification problems. In particular, we leverage the SSM perspective to develop zero-shot adaptation methods for SNNs under temporal-resolution mismatch between training data and test data. 

This corresponds to the scenarios where the temporal resolution of target data during deployment is not the same with that of the pre-deployment source data used for training. Furthermore, building upon core SNN concepts, we introduce spiking and state reset mechanisms into deep SSMs, and demonstrate the potential performance gains provided by these biologically inspired non-linearities. 

Overall, our results on various tasks, including an audio keyword spotting task, an event-based vision task and a sequential pattern recognition task, demonstrate that promising performance gains for temporal modeling emerge from careful integration of concepts from SNN and SSM frameworks.

Sanja Karilanova is a final PhD student in the Department of Electrical Engineering, Division of Signals and Systems at Uppsala University. Her doctoral research lies at the intersection of machine learning and neuromorphic computing, with a particular focus on spiking neural networks and deep state-space models. She holds a Master’s degree in Mathematics with Honours from the University of Manchester, where she completed a four-year master degree with integrated bachelor degree.