Bachelor/Master Thesis: "SpikeSynth: A Robust Analog Printed Spiking Neuromorphic Circuit with Learnable Spike Generator"

Description:

With the rapid growth of emerging fields such as the Internet of Things and wearable devices, more requirements such as flexibility, low-cost and non-toxicity are posed. However, these advanced properties are often beyond the reach of conventional rigid silicon technology, which make printed electronics increasingly recognized as a key driver of these fields. With neuromorphic computing, printed neuromorphic circuits offer not only the aforementioned properties, but also their learning-based design process offers high optimization efficiency. This work mainly focuses on designing a robust spiking neural network at the software level (the yellow box) and explore an approach to make the components of spike-generator subcircuit optimizable. Thereafter it will be trained and tested on various benchmark datasets to see the classification accuracy and the power and energy-efficiency on the robust P-SNN.

The primary objectives of this thesis project will be:

  1. Development: Define a learnable SG model which can mimic the behavior of the spiking neuromorphic circuit.
  2. Evaluation: Design and conduct experiments to assess the effectiveness of the LSG, comparing its performance against conventional SNNs and other baseline models.
  3. Optimization: Investigate methods to optimize the training process and computational efficiency of SNNs.
  4. Application: Apply the developed method to real-world tasks where spiking neural network is crucial, such as medical diagnosis, brain-inspired computing etc.

 

Requirements:

  • Good background in machine learning and deep learning.
  • Familiarity with Spiking Neural Networks or interest to learn
  • Proficiency in Python and deep learning libraries such as PyTorch or TensorFlow
  • Good mathematical and statistical skills
  • Interest in novel research and pushing the boundaries of current machine learning techniques

 

Benefits:

  • Opportunity to contribute to the cutting edge of machine learning and circuit related research
  • Gain experience in developing and evaluating novel spiking neural network architectures
  • Potential to publish findings in a top-tier machine learning/EDA conference or journal

 

Interested candidates should contact Priyanjana Pal/Tara Gheshlaghi at priyanjana.pal@kit.edu, tara.gheshlaghi@kit.edu and Prof. Mehdi Tahoori at mehdi.tahoori@kit.edu with their CV, transcript, and a brief description of their interest in the project.

References:

  1. P. Pal et al., “Analog printed spiking neuromorphic circuit,” in IEEE DATE, 2024, p. 6 S
  2. H. Zhao et al., “Highly-bespoke robust printed neuromorphic circuits,” in 2023 DATE. IEEE, 2023, pp. 1–6
  3. H. Zhao, P. Pal, M. Hefenbrock, M. Beigl, and M. Tahoori, “Poweraware training for energy-efficient printed neuromorphic circuits,” in 42nd IEEE/ACM ICCAD, 2023

References on SNN:

  1. J. K. Eshraghian et al., “Training Spiking Neural Networks Using Lessons from Deep Learning,” arXiv preprint arXiv:2109.12894, 2021.
  2. C. Tang and J. Han, “Hardware Efficient Weight-Binarized Spiking Neural Networks,” in 2023,(DATE). IEEE, 2023, pp. 1–6.
  3. Vaswani et al., “Attention Is All You Need,” Advances in neural information processing systems, vol. 30, 2017.

Tutorials:

Cosyne 2022 Tutorial on Spiking Neural Networks - Part 1/2

Tutorials — snntorch documentation