Adaptive and Efficient Deep Neural Networks for Edge Computing

Abstract

Existing deep networks are mostly static – they can only run at a specific resource budget. In real applications,
the resource constraints (e.g., computation and memory) are always changing with many conditions
such as battery and workload. To meet the dynamic resource constraints, one needs to deploy several
trained networks with different scales and switch among them, which consumes much more memory
and is not efficient and scalable. This talk will discuss our recent work of addressing the important problem
of efficient and adaptive inference under dynamic resource constraints. We propose a new training
framework to mutually learn from network width and input resolution for adaptive accuracy-efficiency
trade-offs. At runtime, our method dynamically selects the optimal resource-accuracy trade-off to fit the
model’s resource demand to the system’s available runtime resources. This talk will conclude with a discussion
of some future prospects for developing adaptive neural networks.

For more info, please follow this link.