Exploiting spatio-temporal patterns with neuromorphic systems

New demands in artificial intelligence, the increase of data available, and the forecasts of reaching Moore’s law ceiling push algorithms towards the edge for low latency, low power, and highly intelligent devices. Neuromorphic systems mimic brain-like computations to capture the efficiency and adaptive behaviour exhibited in biological systems being promising candidates to lead the new generation of artificial systems. Sensory information is encoded with precise spatio-temporal patterns in the nervous system. Similarly, neuromorphic computation utilises event-based representations in its computational systems, inspiring its use for building artificial cognitive systems. In this work, we build Spiking Neural Networks (SNNs) in the tactile and auditory sensory modalities for classification tasks in the context of full event-based sensory systems. Compared with standard machine learning implementations in GPU, SNNs implemented in neuromorphic hardware output close to 500 times less energy consumption highlighting the strengths of the new hardware paradigm. Also, we explore how to encode spatio-temporal features from sensing devices emphasising the benefits of using full neuromorphic event-based sensors and systems. We further explore improvements in spiking networks based on the Leaky-Integrate-and-Fire (LIF) models. The proposed network’s architecture implements Time Difference Encoders (TDEs), a cell model based on brain-inspired computations. We find promising results with a 92% reduction of synaptic operations and highly interpretable network results against typical current-based recurrent LIF networks. The results of this work contribute towards building neuromorphic sensing systems from sensor devices, algorithms and circuit design to quantify the strengths of the new computational paradigm with promising capabilities for intelligent machines on edge inspired by their biological counterparts.