First author and postdoctoral researcher Gido van de Ven and principal investigator Andreas Tolias at Baylor, with Hava Siegelmann at UMass Amherst, write in Nature Communications that they have developed a new method to protect – “surprisingly efficiently” – deep neural networks from “catastrophic forgetting” – upon learning new lessons, the networks forget what they had learned before.
Siegelmann and colleagues point out that deep neural networks are the main drivers behind recent AI advances, but progress is held back by this forgetting.
They write, “One solution would be to store previously encountered examples and revisit them when learning something new. Although such ‘replay’ or ‘rehearsal’ solves catastrophic forgetting,” they add, “constantly retraining on all previously learned tasks is highly inefficient and the amount of data that would have to be stored becomes unmanageable quickly.”
Unlike AI neural networks, humans are able to continuously accumulate information throughout their life, building on earlier lessons. An important mechanism in the brain believed to protect memories against forgetting is the replay of neuronal activity patterns representing those memories, they explain.
Siegelmann says the team’s major insight is in “recognizing that replay in the brain does not store data.” Rather, “the brain generates representations of memories at a high, more abstract level with no need to generate detailed memories.” Inspired by this, she and colleagues created an artificial brain-like replay, in which no data is stored. Instead, like the brain, the network generates high-level representations of what it has seen before.
The “abstract generative brain replay” proved extremely efficient, and the team showed that replaying just a few generated representations is sufficient to remember older memories while learning new ones. Generative replay not only prevents catastrophic forgetting and provides a new, more streamlined path for system learning, it allows the system to generalize learning from one situation to another, they state.
For example, “if our network with generative replay first learns to separate cats from dogs, and then to separate bears from foxes, it will also be able to tell cats from foxes without having ever seen a cat and a fox at the same time,” says Van de Ven.
He and colleagues write, “We propose a new, brain-inspired variant of replay in which internal or hidden representations are replayed that are generated by the network’s own, context-modulated feedback connections. Our method achieves state-of-the-art performance on challenging continual learning benchmarks without storing data, and it provides a novel model for abstract level replay in the brain.”
Van de Ven says, “Our method makes several interesting predictions about the way replay might contribute to memory consolidation in the brain. We are already running an experiment to test some of these predictions.”