Exploiting Heterogeneity in Neuromorphic Systems for Robust Analog Computing

Contact: Prof. Giacomo Indiveri, giacomo (at) ini.uzh.ch
Dr. Chiara De Luca, chiara.deluca (at) ini.uzh.ch

Neuromorphic computing represents a cutting-edge approach to designing computational systems by mimicking the architecture and functionality of biological neurons. One of the persistent challenges in fabricating neuromorphic devices is the cross-device response variability, which is often seen as a limitation. However, biological neurons and synapses are intrinsically heterogeneous, exhibiting a wide spectrum of responses that enhance robustness and adaptability. Inspired by this, recent computational study[1] demonstrated that neural networks composed of heterogeneous neurons—without the need for plasticity—significantly outperform their homogeneous counterparts, particularly in their reliability across a range of temporal tasks.

This finding suggests that embracing the inherent diversity of analog devices, rather than suppressing it, can unlock significant performance benefits. The implications extend well beyond theory, providing a compelling opportunity to harness this diversity for real-world neuromorphic applications.

 

Project Objectives:

1. 1. Analog Implementation of Heterogeneous Neuron Networks:
   The first objective is to translate the computational algorithm we developed for networks with static heterogeneous neurons, onto analog hardware (Dynap-SE)[2]. This will provide the foundation for real-world neuromorphic systems that exploit the variability of analog components, paralleling the natural diversity seen in biological neurons.

2. 2. Extension to Networks with Plastic Synapses:
   Building on the static network model, the next step is to integrate synaptic plasticity into the system. In this phase, synaptic weights will evolve based on local, unsupervised learning rules driven by neuronal activity at the synapse terminals, simulating the adaptive nature of biological systems.

3. 3. Adaptation to Non-Stationary Inputs:
   Finally, we will extend the framework to handle non-stationary inputs, investigating how heterogeneous networks with plastic synapses can dynamically adapt to changing input distributions, further increasing the applicability and resilience of neuromorphic systems in complex, real-world scenarios.

 

Student Expectations and Opportunities: The ideal candidate for this project will have a solid foundation in Python, experience with Git version control, and proficiency with Unix-based systems. A background in computational mathematics, physics, or a related field will be highly advantageous.

The student will benefit from access to neuromorphic chips at INI as well as the modular Python framework that implements the first two objectives on a CPU, developed at the University of Göttingen. Additionally, the student will receive direct mentorship from the author of the original algorithm, as well as technical instruction for working with the chip and its simulator, ensuring a well-supported and collaborative research environment.

 

Additional Conditions:

• - Familiarity with neural networks or neuromorphic computing is a plus, though not strictly required.

• - Enthusiasm for interdisciplinary work at the intersection of physics, biology, and computer science is highly encouraged.

 

Suggested bibliography

[1] Golmohammadi et al 2024, https://arxiv.org/abs/2412.05126

[2] Zendrikov et al, 2023 10.1088/2634-4386/ace64c