Close

Presentation

kCC-Net for Compression of Biomedical Image Segmentation Networks
Time
Location
Event Type
Special Session (Research Track)
Virtual Programs
Hosted in Virtual Platform
Topics
Design
Machine Learning/AI
DescriptionDeep learning has achieved great success in medical image segmentation, mostly based on the state-of-the-art fully convolutional networks (FCN). However, the resulting networks typically have millions of weight values and lack inference efficiency, especially when they are to be placed in embedded hardware platforms for accelerated local inference. As an example of application, there is a strong clinical need in analyzing skin images for various diseases using hand-held devices such as smart phones. In this talk, we will present a novel hardware-oriented deeply compression scheme for FCNs towards embedded semantic segmentation. Different from known compression techniques for convolutional neural networks (CNNs) which simply prune weights/neurons/connections based on their amplitudes, our framework combines the domain knowledge of medical imaging and the embedded hardware platform structure, and achieves more than 1,000x memory reduction and up to 17x inference speedup without much sacrifice in accuracy.