Computer Vision News - April 2019

experience previous tasks .” Aswin further elaborates: “ One key technique we use is generative modeling to learn and model experiences related to a task. The generative memory supports recall of the most relevant memory to your current situation. The second big technique is reinforcement learning, specifically multi-task reinforcement learning. It lets us transfer these memories to an actionable representation of what the agent has to do in a new, unseen task. ” Currently, the project is currently at an early research stage, but eventually, it could have many commercial applications for creating a technology that would not require retraining. Lifelong learning has several applications especially now that deep learning applies to all kinds of devices and applications. As Aswin explains: “ Our lifelong learning technique extends beyond StarCraft. StarCraft is a very hard problem that gives a faithful representation of the capability that we are developing. ” Typically, scientists train a deep learning model and apply it to a device such as the iPhone, Siri, Alexa, or an autonomous car with the hopes that the model will capture all of the situations that will arise in the device. Lifelong learning strives to continue the learning process even after deploying the device in order to handle any surprises. Jesse adds, “ I’d say that lifelong learning is part of a broader goal of making learning systems have more human-like capabilities, especially when it comes to making sense of complex information, learning over time, building a knowledge base. I’d say that the way that this fits into broader machine learning is that so far, machine learning mostly focuses on solving particular tasks, but when we have tools like lifelong learning, we start to think about much more human-like capabilities in terms of building up skills, knowledge, etc. over a lifetime. ” Pauline Luc 27 Computer Vision News Lifelong Learning Machines - SRI Application Feedback of the Month The team at RSIP has been great to work with, in supplementing some of our internal computer vision engineering efforts and meeting aggressive deadlines! Modar Alaoui Founder & CEO, eyeris.ai