Stephen Grossberg Introduces Adaptive Resonance Theory

All the way back in 1976 in a brutally fiendish AI learning winter, Stephen Grossberg asked a question still mostly unsolved today: how a machine can keep learning without forgetting what it already knows. Models like ChatGPT stop learning once training ends, and teaching them new things can even damage old skills.

Grossberg's solution was feedback and his theory was adaptive resonance theory (ART), which he introduced in two 1976 papers in Biological Cybernetics. ART understood the power of feedback: a learned top-down expectation is matched against each bottom-up input. ART's memory changed only when the two agreed closely enough or when something completely new appeared.

In Grossberg and Gail Carpenter’s words from 2003, “ART matching leads either to a resonant state, which persists long enough to permit learning, or to a parallel memory search.” Carpenter and Grossberg later built ART networks such as ART 1 but the system never became mainstream.

While Grossberg and later Carpenter were not necessarily solving these problems, they were at least working on the right ones. It was hard to say that about most AI research of this time of symbolic AI's heyday.