A Neuromorphic Framework for Event-Based DNNs using Minifloats
Supervisors:
Description:
Researching a novel neuromorphic framework that performs proven deep learning algorithms (to solve real world problems) using event-based signal processing. Biological cortex, which intrinsically computes with discrete-time events by using spiking neurons, exhibit complex yet stereotypical network architectures that support rich dynamics. Developing the computing models for our neuromorphic framework using low precision formats. Recent 8-bit foating-point (minifoat) representations used by DNNs have achieved marginal equivalent accuracy to FP32 foating point precision over dikerent tasks and datasets while providing orders of magnitude reduction in silicon area and power consumption. Minifoats are ideal candidates for neuromorphic computing. This framework will have a set of minifoat neuromorphic computing models that include a versatile three compartment spiking neuron model. This neuromorphic framework will use event-based DNNs that only compute changes, which is significantly more effcient than updating all the layers each time. By combining the event-based processing and the minifoat together, this framework will be capable of achieving marginal differences in accuracy compared to the corresponding frame based DNNs with full precision foating point number over different tasks and datasets while with orders of magnitudes reduction in computation resouces in terms of memory footprint and foating point operations per second (FLOPS).
