February 4, 2020

362 words 2 mins read

Enabling traditional vision on specialized deep learning hardware

Enabling traditional vision on specialized deep learning hardware

In recent years, weve seen a shift from traditional vision algorithms to deep neural network algorithms. While many companies expect to move to deep learning for some or all of their algorithms, they may have a significant investment in classical vision. Paul Brasnett explains how to express and adapt a classical vision algorithm to become a trainable DNN.

Talk Title Enabling traditional vision on specialized deep learning hardware
Speakers Paul Brasnett (Imagination Technologies )
Conference Artificial Intelligence Conference
Conf Tag Put AI to Work
Location London, United Kingdom
Date October 9-11, 2018
URL Talk Page
Slides Talk Slides
Video

Over the past several decades, the industry has invested heavily in traditional vision algorithms. However, in the past few years, there’s been a major shift from traditional vision algorithms to deep neural network algorithms. While many companies expect to move to deep learning for some or all of their algorithms, they may already have a significant investment in classical vision and likely want to continue to use classical vision algorithms for specific tasks. After all, they don’t want to lose their investment. When specifying silicon for next-generation products, companies must consider whether to prioritize the hardware needed to support the traditional vision algorithms that they have today or that needed for deep neural network algorithms in the future. There is an increased cost to optimize for both. However, if some of the traditional algorithms can run on deep learning optimized hardware, more area could be allocated for the deep learning hardware. Paul Brasnett discusses how a company can maximize its existing investment in traditional algorithms while building toward deep learning algorithms. Paul examines the similarities between classical and deep vision and explains how to express and adapt a classical vision algorithm to become a trainable DNN. Along the way, Paul demonstrates how traditional computer vision algorithms (e.g., histograms and morphological operators) can be mapped on to deep learning primitives and explores a case study on mapping a feature point descriptor called Brisk. This enables traditional vision on specialized deep learning hardware, ultimately providing a low-risk path for developers transitioning from traditional vision algorithms to DNN-based approaches and enabling them to maximize their investments.

comments powered by Disqus