The current neural networks revolution in AI that has led to advanced solutions for tasks ranging from image classification to natural language processing has been fed by significant advances computing power and the availability of nearly endless training data. However, the outlook for these underlying advantages increasingly looks bleak. The ever-increasing scale of neural network models have ballooned the energy and monetary costs of AI research and the unlimited availability of data looks to be increasingly vulnerable to concerns about privacy and security as well as significant costs of data collection and management.
For these reasons there is a growing need for novel data-efficient AI solutions that can achieve cognitive capabilities while leveraging low-power hardware such as neuromorphic systems. This talk will describe Sandia’s research in brain-inspired algorithms that not only enable current AI methods to be deployed on low power spiking neuromorphic hardware but also help reach towards novel cognitive capabilities that have continued to elude the current generation of AI methods. Dr. Aimone will describe two Sandia-developed tools for porting algorithms to neuromorphic hardware (Whetstone and Fugu) and he will illustrate how seemingly non-cognitive algorithms can be efficiently deployed in spiking hardware. Finally, Dr. Aimone will describe Sandia’s recent progress in formalizing neuroscience knowledge of hippocampus subregions into algorithms suitable for a brain-like memory formation.