Session Info

Continual learning, the ability to integrate new information while preserving previously acquired knowledge, remains a significant challenge in current AI models. While humans and other mammals effortlessly exhibit this capability, replicating it in artificial systems has proven to be a daunting task. This talk will explore how this challenge can be addressed by applying insights from neuroscience to designing neuromorphic continual learning systems.

Inspired by the intricate workings of the brain, we investigate mechanisms such as metaplastic- ity and synaptic consolidation, neurogenesis, and context-based neuromodulation to overcome the challenges of catastrophic interference and adaptation in artificial systems. Realizing this form of learning enables a number of advancements, including the ability to solve complex sequential tasks with local learning, to operate within low memory bounds, to compute in spiking domain, and to deploy in energy- constrained environments.

For the first time in literature, we have showcased continual learning of sequential tasks with both spiking algorithms and neuromorphic ASICs. We present spiking algorithms that solve sequential classification in real-world deployment scenarios. A custom-designed Neuromorphic accelerator that learns continually in real time with imbalanced data will be highlighted. To effectively evaluate the performance of this new class of continual learning systems, we also introduce a comprehensive set of new metrics and benchmarks.