AI system that continually learns
Artificial intelligence (AI) systems are driving technology advancements in many commercial and government applications such as speech recognition and autonomous robotics. However, current systems are not able to handle new scenarios that these are not trained on. AI systems today can repeatedly make the same mistakes. Even with retraining, the systems are prone to catastrophic forgetting when a new item disrupts previously-learned knowledge.
To address these limitations, SRI International is developing a next-generation AI system that can learn continuously and apply that learning to become better and more reliable at performing new tasks. The work is being done under Defense Advanced Research Projects Agency’s (DARPA’s) Lifelong Learning Machines (L2M) programme.
SRI researchers are training AI agents using real-time strategy games such as StarCraft II. Using deep reinforcement learning methods, AI agents are trained with surprises injected into the game (for example, terrain and unit capability change). Using this method, key metrics for life-long learning metrics such as adaptation, robustness and safety can be measured.
Overlapping tasks in latent space (blue/orange) makes Lifelong RL training prone to forgetting (a), while our generative memory and task encoding separate the latent space to ensure efficient memory use to avoid forgetting (b) (Credit: SRI International)