Follow Us On

NengoEdge and the Legendre Memory Unit: State-of-the-Art Low-Power AI for Time Series Inference

October 14, 202112:20 - 13:00AI Hardware - AI Expo 2021

Moderator

David Wyatt

David Wyatt

CTO, President, PixelDisplay
  • October 14, 2021
  • 12:20 PM - 1:00 PM

Speaker

Peter Suma

Peter Suma

Chairman & Co-CEO, Applied Brain Research Inc.
Chris Eliasmith

Chris Eliasmith

Co-founder and Co-CEO, Applied Brain Research Inc.

Abstract

ABR has recently developed a new kind of recurrent neural network that is provably optimal at compressing information over time called the Legendre Memory Unit (LMU). We have used the LMU to set new state-of-the-art records for a variety of time-series problems, with minimal compute. In this talk we introduce the LMU, and discuss how it achieves state-of-the-art performance on industry standard keyword spotting datasets. We also show that the LMU outperforms and outscales transformers on large NLP problems.  We demonstrate how our easy-to-use cloud offering, NengoEdge, allows anyone to use the LMU on their time series problems, optimizing directly for a variety of edge hardware targets. Looking ahead, we discuss our LMU AI chip that is in development and will deliver low-power, low-cost full-speech and NLP inference at the edge.