Poster: Optimizing Keyword Spotting on Microcontrollers

Undergraduate Adithya Lakshminarayanan presented his summer research at the McGill SURE Poster Session today.

Professor Meyer and Adithya Lakshminarayanan with their SURE 2019 poster.


Neural networks have become increasingly popular, due to their promise of state-of-the-art accuracy, in fields such as image and speech recognition. However, with the proliferation of resource-constrained edge computing platforms, in particular embedded devices and microcontrollers, it is also crucial to optimize for the memory utilization, compute, and latency. One method proposed to optimize neural networks is to quantize weights and biases, i.e., reduce the precision of trainable parameters. We investigate the effects of using quantization on trade-offs between model accuracy and cost. In particular, we vary quantization format (8 bits, 16 bits, and 32 bits floating-point) and use design space exploration techniques to perform quantization-aware training using a speech dataset on CNN architectures. We note that the choice of quantization format is not a trivial one while attempting to optimize neural networks with respect to several objectives; a single format does not prove dominant at all regions of the pareto-optimal front approximated from the results we obtain.

Download the poster .

Learn more about the Produced by OmniGraffle 6.6.2 2019-07-31 17:33:34 +0000Canvas 1Layer 1W W W SURE Program at McGill.

Follow Adithya on Twitter.