Training Neuromorphic Devices with a novel feedback control algorithm - Applications for online learning with spikes

This project addresses the challenge of training multi-layered Spiking Neural Networks (SNNs) online on Neuromorphic devices. For this purpose, you will explore the implementation of a novel spike-based feedback-control algorithm on mixed-signals Neuromorphic devices (the DYNAP-SE [1]) and test its performance on different tasks.

The research will be conducted at the Institute of Neuroinformatics (INI, Y55, Irchel Campus ETH/UZH), in the Neuromorphic Cognitive Systems (NCS) group of Prof. Dr. Giacomo Indiveri. The work will be supervised by Dr. Matteo Saponati and Dr. Chiara De Luca, with the technical supervision of Dr. Dmitrii Zendrikov.

 

Motivation

Neuromorphic engineers have successfully applied SNNs to sensory signal processing and bio-signal classification [2]. However, neuromorphic devices are typically used to implement shallow networks, as there is no well-established method yet for training deep SNNs, using local learning rules directly in hardware. Further, an in-depth understanding of the learning properties of biological neural networks in complex neural circuits is missing. In essence, building trainable hierarchical (deep) SNNs is still a hard computational and engineering problem. To address these challenges, you will use an innovative learning framework for SNNs that combines spike-based local learning [3,4] and control feedback signals [5]. The algorithm operates locally and does not require backpropagating gradients, allowing online learning in SNNs and neuromorphic devices.

 

Approach

Taking advantage of the current technologies available in the NCS group, we will implement the feedback-control algorithm on the DYNAP-SE chip [1].

1. 1. Getting hands-on experience with the DYNAP-SE chip, and how to implement custom learning rules. (1 months)

2. 2. (Training setup) Define a set of binary classification problems and implement the inputs and the target spike trains with the DynapSE1. (0.5 months)

3. 3. (Training phase) Train SNNs on the DYNAP-SE on this family of binary classification tasks, evaluation of performances, and troubleshooting. (2 months)

4. 4. (Validation phase) Investigate the robustness of the results to variations in model parameters and device mismatches. Find which range of device parameters are best suited to implement the feedback-control algorithm on the DYNAP-SE. (1.5 months)

 

Your profile

• - Interest in Computational Neuroscience (Neuroinformatics), Neuroscience, and Neuromorphic Engineering.

• - Basic programming skills (Python strongly preferred).

• - Basic knowledge of Artificial Neural Networks (ANNs) and/or Spiking Neural Networks (SNNs).

• - Basic understanding of neural dynamics and dynamical systems.

• - Interest in testing real-world (electronic) spiking neural networks in real-world (noisy/variable) conditions.

 

References

[1] Moradi, S., Qiao, N., Stefanini, F., & Indiveri, G. (2017). A scalable multicore architecture with heterogeneous memory structures for dynamic neuromorphic asynchronous processors (DYNAPs). IEEE transactions on biomedical circuits and systems, 12(1), 106-122.
[2] Mitra, S., Fusi, S., & Indiveri, G. (2008). Real-time classification of complex patterns using spike-based learning in neuromorphic VLSI. IEEE transactions on biomedical circuits and systems, 3(1), 32-42.
[3] Chicca, E., Stefanini, F., Bartolozzi, C., & Indiveri, G., Proceedings of the IEEE, 102(9), 1367–1388 (2014)
[4] Moradi, S., Qiao, N., Stefanini, F., & Indiveri, G., IEEE Transactions on Biomedical Circuits and Systems, 12(1), 106–122 (2018)
[5] Meulemans, A., Tristany Farinha, M., Garcia Ordonez, J., Vilimelis Aceituno, P., Sacramento, J., & Grewe, B. F., Advances in Neural Information Processing Systems, 34, 4674–4687 (2021)