Convolutional neural networks for efficient object detection on ultra low-power platforms

At the University of Bologna, the Microelectronics Research Group has been working on smart data analytics on ultra-low-power sensors for the past few years. This smart analysis is in many cases based on convolutional neural networks as the fundamental tool to extract features and information out of various raw data streams. Applying these techniques on the acquisition device itself can help reducing data transfer and storage but requires neural network models with small memory footprint and a really constrained computation workload. This work proposes a software architecture and advanced quantization techniques to obtain image classification models with high accuracy, small size and low memory footprint that can properly work on a low-power device. The design is specifically tailored to support the low-resolution environment available in the PULP platform, which includes a hardware convolution engine to efficiently compute convolution operations required by neural network models