I study machine learning and the brain.
Computation in "deep learning" neural networks superficially resembles computation in the brain. Can we use low-level details about computation in the brain to improve deep networks? In particular, what happens when deep networks use sparse connectivity, sparse activation, and quantized weights and activations? Can we achieve significant improvements in efficiency, enabling today’s hardware to run larger networks and thus become more capable? Does sparsity itself provide benefits other than efficiency? Do new classes of hardware become the new optimum with this type of network? If we co-evolve the networks and the hardware, where do we end up?
This is my main project at Numenta. (Fun fact: We live-stream most of our research meetings.)