Neuromorphic Human Activity Recognition through LIF-based neurons

Human Activity Recognition (HAR) is the classification task focused on daily-life activities. It relies on time-dependent signals coming from body monitoring, and it can be performed in different domains ranging from healthcare to sport and working conditions. Consequently, HAR is inherently well suited for on-edge deployment of tailored services, whose effectiveness must be evaluated taking into account energy efficiency as an additional, major figure of merit. Bringing HAR into the neuromorphic domain can lead to significant improvements in this regard, especially if dedicated hardware is used. An example of this latter is the Intel’s Loihi chip, which implements a discrete Current-Based Leaky Integrate-and-Fire (CuBa-LIF) neuron model also referred to as second-order LIF model. In view of on-edge applications of such neuromorphic hardware, it can be useful to investigate the suitability of a fully LIF-based Spiking Neural Network (SNN) in terms of both accuracy performance and energy consumption. Additionally, by working with such network, it is possible to investigate the possibility of targeting HAR by means of SNNs relying on the most biologically plausible neuron model among the less computational expensive ones.