This research project delves into the performance evaluation of lightweight neural networks when processing both event and frame-based data for digit classification. Specifically, it investigates the impact of quantization on the efficiency and effectiveness of these networks across two distinct datasets: MNIST (images) and N-MNIST (events).
- Evaluation of lightweight neural networks' performance with event and frame-based data.
- Examination of how quantization affects lightweight neural networks processing event and frame-based data.
- Investigate the impact of parameter reduction and increased network depth on the performance of quantified neural networks
Three different neural network architectures were employed throughout the experimentation process, with models tested both with and without quantization. The precision levels ranged from 8-bit down to 1-bit.
The findings of this research have been published. You can access the publication here.