Adham Ghazali, co-founder and CEO of Imagry, presents the "Edge Intelligence: Visual Reinforcement Learning for Mobile Devices" tutorial at the May 2017 Embedded Vision Summit.
Real-life visual data encompasses a tremendous amount of information and presents a huge challenge for the design and development of a perceptual engine. Smart machines equipped with visual understanding technology will always be presented and challenged with new data. In this talk, Ghazali presents algorithm methods to allow learning in the end device to enable it to understand new data.
The main challenges to learning in the end device stem from the lack of computing power, lack of access to sufficient data samples and the need for involvement of human experts. Imagry addresses these challenges by combining a binary weights representation of deep neural networks and reinforcement learning. In particular, Ghazali explores and introduces a self-expanding cost function and the incorporation of external memory to enable DNNs to adapt to new data.