Sam Hames

[1903.03129] SLIDE : In Defense of Smart Algorithms over Hardware Acceleration for Large-Scale Deep Learning Systems

An interesting approach to a deep learning problem. Instead of computing everything as matrix multiplication (which generally requires a GPU for throughput), turn it into a sparse lookup table and use a conventional CPU.

I'm not sure I understand the paper well enough to comment on the methodology, but fast inference and training on conventional CPUs would be very exciting - building and running GPU based stacks is fiddly and time consuming whereas the CPU is there and just works. CPUs are also great for scaling down!