Session Info

Algorithms Keynote: Lifelong Learning at the Edge: Algorithms and Foundations

Lifelong Learning at the Edge: Algorithms and Foundations

Lifelong Learning is at the cutting edge of artificial intelligence – encompassing computational methods that allow systems to learn in runtime applying this knowledge to new, unanticipated situations. Until recently, this sort of computation had been found exclusively in nature; thus, Lifelong Learning looks to nature for its underlying principles and mechanisms applying them to this new technology.

Learning is an energy-hungry process and since computational machines are considered to either learn or compute, one might reasonably assume Lifelong learning would be limited to cloud and off-line computing only. But we demonstrate that these constraints are specifically associated with the backpropagation algorithm and are not applicable to our new learning method. We introduce signal propagation that combines learning into the inference, enabling it to add classes on the go with no predefined limit. We note also that Lifelong Learning, like classical learning, is limited in its ability to utilize temporal patterns in its computation. We introduce mechanisms that make the AI temporally aware, enabling unparalleled capacity, strong temporal accuracy, the ability to work effectively with limited data, and significantly, supports great savings in size and power usage.

Our computational foundations need to be overhauled to enable this new type of computation to evolve. While Turing computation has worked extremely well, it is unable to describe machines that change regularly, adapt in real-time to their environment and apply lessons learned to new, previously unseen circumstances. Having introduced his logical universal model, Alan Turing spent most of his research time in an effort to find a new model that would more closely learn as the brain does. Super-Turing computation, which accurately describes biological learning, may be the answer to Turing’s quest; in the meantime, Super-Turing has become the foundational theory of Lifelong Learning, leading toward a stronger, more robust AI.


Prof. Hava Siegelmann

University of Massachusetts Amherst