Fast Event Classification

Overview

Dynamic Vision Sensors (DVS) output asynchronous streams of events instead of conventional image frames, enabling low-latency and sparse visual processing. This project explores event-based classification using convolutional spiking neural networks deployed under neuromorphic hardware constraints, with particular focus on temporal efficiency and compatibility with Intel Loihi.

The system processes event streams from the DAVIS240c event camera and evaluates low-resolution, low-latency representations suitable for neuromorphic deployment.

Event Representation

The input representation is based on raw DVS events temporally downsampled from the camera's microsecond resolution to Loihi-compatible millisecond-scale temporal bins.

To reduce computational and hardware cost, the original DAVIS240c spatial resolution is downsampled to 32×32 and 64×64. Event streams are converted into sparse spike-based tensors while preserving temporal structure and event polarity information. The representation is designed to maintain low-latency responsiveness while remaining compatible with constrained neuromorphic hardware resources.

Classification Architecture

The classification model consists of a two-layer convolutional spiking neural network implemented in PyTorch.

The architecture processes temporally discretised event streams directly as spike activity, enabling sparse computation, temporal feature extraction, and event-driven inference. The network is designed with deployment-oriented constraints in mind, including compatibility with Loihi-supported neuron and connectivity configurations.

Results

Placeholder for accuracy, latency, and energy metrics. Add benchmark comparisons against frame-based baselines here.