SA/MA: Spiking Neural Networks (SNN) on Memristor Crossbar-based Neuromorphic Hardware

Type: The project can start immediately as a Semester/Master Project, or (preferred) as a Master Thesis. The Institute of Neuroinformatics (INI) is a joint institute of the University of Zurich and ETH Zurich. INI carries out experimental, theoretical and applied research with the aim of discovering key principles by which the brain is built and works and using this knowledge in practical Artificial Intelligence (AI) applications where possible. The Neuromorphic Cognitive Systems (NCS) group at INI has been developing neuromorphic platforms that enable event-driven computing using algorithms and primitives inspired by our brain.

Motivation


Memristive devices have emerged as non-volatile memory devices with low energy consumption, low latency, and nano-scale area which are 3D-integratable on CMOS technology. This makes them a promising candidate for the next generation of memory elements in electronic devices. In existing multi-core neuromorphic chips, the connections between spiking neurons is stored in on-chip SRAM memory. The state of the memory dictates how the events are routed in the network. These memories require a significant overhead in terms of space and can take up to more than 50% of the on-chip real-estate area [1]. A hybrid memristive-CMOS neuromorphic processor architecture Mosaic [2] has been proposed re- cently in which the bulky SRAM cells used for routing spike events are replaced with non-volatile memristors, and programming them to define different SNN connectivity patterns. This neuromorphic architecture has potential to scale up the system. However, the architecture does not include the arbiter blocks that is usual in the design of the event-based routing architecture. Therefore, there is the possibility that events/spikes collide (i.e. arrive at the same time to a node). The goal of this project is to study the effect of the number of inputs and outputs of a neuron (fan-in and fan-out), and also the sparsity of the network on the probability of collision based on SNN simulations.

Task Description


The student will train Spiking Neural Networks on multiple tasks (e.g. heart beat anomaly detection, and/or Neuromorphic MNIST (NMNIST) and/or speech recognition), and will study the effect of spike collision of different neurons in terms of accuracy and sparsity of the network, along with the fan-in of the neurons.

References

[1] Saber Moradi, Ning Qiao, Fabio Stefanini, and Giacomo Indiveri. 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, 2018.
[2] The neuromorphic mosaic: re-configurable in-memory small-world graphs. 2021.

Requirements

Basic python skill is expected. Previous knowledge about SNN or artificial neural network (ANN), neuromorphic engineering is recommended.

Contact

Please contact the people below with a CV, short motivation and background (<0.5 page). If you have any questions about the project, do not hesitate to contact us.
Junren Chen: junren (at) ini.uzh.ch,
Dr. Melika Payvand: melika (at) ini.uzh.ch
Prof. Giacomo Indiveri: giacomo (at) ini.uzh.ch

© 2022 Institut für Neuroinformatik