Deep Learning with embedded skepticism about inputs and data. Useful for imperfect sensor data, and data from human interactions and queries. Leading to safe decisions in context. Required for many kinds of applications.
Algorithm Helps AI Systems Dodge 'Adversarial' Inputs
MIT News, Jennifer Chu, March 8, 2021
Massachusetts Institute of Technology (MIT) researchers have developed a deep learning algorithm designed to help machines navigate real-world environments by incorporating a level of skepticism of received measurements and inputs. The team mated a reinforcement-learning algorithm with a deep neural network, each used separately to train computers in playing games like Go and chess, to support the Certified Adversarial Robustness for Deep Reinforcement Learning (CARRL) approach. CARRL outperformed standard machine learning techniques in tests using simulated collision-avoidance and the videogame Pong, even when confronted with adversarial inputs. MIT's Michael Everett said, "Our approach helps to account for [imperfect sensor measurements] and make a safe decision. In any safety-critical domain, this is an important approach to be thinking about."
No comments:
Post a Comment