Student project: Bio-inspired near-sensor feature extraction with spiking neural networks

At IMEC (Thesis)

This project aims to implement and validate a fast and efficient bio-inspired learning+inference algorithm that can be deployed near the sensor front-end.

 

What you will do
The project will be part of a collaboration with CNRS Toulouse. Demonstrations and experiments with real-time sensor data will take place in the context of a European project, DAIS.

 

By contrast to standard ANNs and DNNs, neurobiology-inspired (neuromorphic) neural networks function in a (semi)-asynchronous, event-driven manner, optimally exploiting spatio-temporal sparsity. This allows them to operate with a minimal power budget while efficiently utilizing processing resources and memory IO.

 

Initially, the training part of the algorithm will be implemented in software. In contrast, the inference part will be implemented directly in HDL/RTL code on an FPGA, aiming at a prototype with suitable interfacing to sensors and debugging from an SoC. The system will then be evaluated against streaming sensor data (e.g., radar and/or vision sensor). Evaluation will involve task performance and processing latency in simple pattern recognition tasks using inputs from a sensor, as well as energy consumption per input-output. Depending on the project duration, at a 2nd stage, the system should be show-cased against live sensor feeds, and at a 3rd stage, the training part should also be moved at /near the sensor for in-situ learning.

 

Tasks:

  • Literature review on neuromorphic computing/engineering.
  • Develop a good understanding of the learning algorithm (experimentation in software simulation).
  • Implementation of the algorithm in HDL together with a testbench.
  • IP block design of interfaces for sensor inputs and embedded processor (for debugging/monitoring).
  • Unit-testing, validation experiments, and performance measurements.
  • Thesis/report writing.

What we do for you
Under the IoT sensor department of Imec-NL, the Neuromorphic computing & engineering group is developing energy and resource-efficient machine learning solutions for application processing on edge and sensor devices. The project will be co-supervised by researchers at Imec-NL and Prof. Simon Thorpe at CNRS-Toulouse. It will be carried out primarily at Imec-NL (with an optional visit if required at CNRS -- also subject to covid19 restrictions).
ImecNL will provide a monthly allowance and a desk to the students.

 

Who you are

  • Enrolled in a Master/Ph.D. program (non-European students are only eligible if they study in the Netherlands).
  • You are available for a period of 12 months.
  • Provable competence/experience in digital design with Verilog/VHDL.
  • Experience with a digital design workflow for FPGA (Xilinx tools).
  • Knowledge of algorithm design.
  • Good working knowledge of python and/or C/C++.
  • Some exposure to Neuromorphic computing/engineering or Machine learning with neural networks is a plus.
  • Good knowledge of spoken and written English.

Interested
Click on 'apply' to submit your application. You will then be redirected to e-recruiting.

Back