Studying quantization effects in RNNs for neuromorphic systems

alternate text

The need for low-power machine learning computing solutions has triggered significant interest in energy-efficient neuromorphic systems. In our previous work [1], we proposed an adaptive neuron model that can be abstracted as a low-pass filter. The model dramatically improved the inference performance of recurrent neural networks on a number of complex spatio-temporal learning tasks: the temporal addition task, the temporal copying task, and a spoken-phrase recognition task. The key value proposition of the model is its compatibility to in-memory neuromorphic processing systems. We estimate that it should be possible to achieve at least 500× higher energy-efficiency using this model on compatible neuromorphic chips in comparison to popular solutions. The current state of this work has been captured in [1].

Project goals

In this project, we will work on implementing an end-to-end spoken phrase detection task using the lpRNN model. The version described in [1] does not take into account quantization and noise effects in weights and activations in the lpRNN model. The objective of this project is to include these constraints on the model and to study the performance of the resulting system in various datasets.

Primary goal

  1. Study the effect of weight and activation quantization in the performance of the lpRNN cell on a spoken phrase detection task in a standard benchmark [2].

Secondary goals

  1. An FPGA implementation of the model.

  2. Integrating the system with the signal processing functions to condition the data arriving from a microphone.,

  3. Extend the results of the study to other datasets for tasks such as human activity recognition.

  4. Publish a machine learning paper describing the results of this work.

Necessary skillset

  1. Python with experience in deep learning libraries such as PyTorch/Tensorflow.

Application

Interested candidates should send an email to Dr. Manu V Nair - mnair (at) ini.uzh.ch

Bibliography

  1. Nair, Manu V., and Giacomo Indiveri. "Mapping high-performance RNNs to in-memory neuromorphic chips." arXiv preprint arXiv:1905.10692 (2019).

  2. Warden, Pete. "Speech commands: A dataset for limited-vocabulary speech recognition." arXiv preprint arXiv:1804.03209 (2018).