Knowledge Based Neural Network Compression: Quality-Aware Model Abstractions.

In the state-of-the-practice IoT platforms complex decisions based on sensor information are made in a centralized data center. Each sensor sends its information over thereafter a decision is send to actuators. In certain applications the latency imposed by this communication can lead to problems. In real time applications it is crucial for the decision to be taken immediately. For this complex decisions should be made on the edge devices themselves. This is what the research track on resource and context aware AI is about. In this we want to develop inference edge systems that dynamically reconfigure to adapt to changing environments and resources constraints. This work if focused on compressing AI processing blocks, specifically neural networks. In this work we want to extend on the current state-of-the-art methods on neural network compression by incorporating a knowledge-based pruning method. By knowledge based we mean we want to prune a neural network in a context aware manner. A certain application context will impose requirements of the outputs of the network. For example, on a highway is the detection of pedestrians less important than cars. Based on these requirements we want to selectively prune a network by locating knowledge concepts related to the outputs. By selectively pruning them we expect to achieve higher compression ratios compared to the state-of-the-art for context specific networks.

Keywords:ARTIFICIAL INTELLIGENCE, INTERNET OF THINGS, EXPLAINABLE ARTIFICIAL INTELLIGENCE, NEURAL NETWORKS

Disciplines:Adaptive agents and intelligent robotics, Machine learning and decision making, Computer vision, Pattern recognition and neural networks