Computer Vision News - April 2019

devices. For a long time, he has worked on the challenging problem to makes neural networks more compact and faster using low-bit precision techniques. He also writes papers and codes as well as grant proposals. In addition, Aswin works on a more applied project called AEOSP for DARPA Communicating with Computers (CwC) , which functions as a visual storytelling platform. During the interactive process, the user types instructions and uses natural language. Then the system turns the story into a visual movie. Aswin and Jesse enthusiastically speak about their work on Lifelong Learning Machines or L2M , a program funded by DARPA MTO (Microelectronics Technology Office) and led by Hava Siegelmann . The program looks at machine learning in a fundamentally different way that addresses the many drawbacks of deep learning and assumptions surrounding machine learning. Currently, deep learning can solve one task but then forgets all of the previous tasks, a process referred to as catastrophic forgetting. The L2M project aims to reduce the forgetting problem and allow for forward transfer, which means taking knowledge from earlier tasks and using that knowledge to help solve later tasks or tasks that the program hasn’t learned yet. Because of the difficulty of lifelong learning and solving the forgetting problem in neural networks, researchers have put aside the problem for several decades. The team at SRI are tackling the high- dimensional problem using a strategy video game, by Blizzard Entertainment called StarCraft2 , as a simulation environment for lifelong learning. StarCraft requires players to switch between solving different tasks. First, the player has to understand how to play the game, a process more difficult than simple image classification. Then, the user must learn to control the game units. The game also involves multiple feature maps that is similar to computer vision and image processing professionals are accustomed to seeing and similar to what the human actually sees when playing the game. Although a Google DeepMind bot recently defeated Pauline Luc 25 Computer Vision News Lifelong Learning Machines - SRI Application