SA/MA: Event-driven network structural plasticity on neuromorphic chip
The Institute of Neuroinformatics (INI) is a joint institute of the University of Zurich and the Swiss Federal Institute of Zurich (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 applications where possible. The Neuromorphic Cognitive Systems (NCS) group at INI has been developing neuromorphic processing platforms that enable event-driven computing using algorithms and primitives inspired by biology.
Neuromorphic hardware tends to pose limits on the connectivity of deep networks that one can run on them. One of the constraints is the number of available synapse circuits per neuron (the neuron fab-in), which include routing memory elements to store the network connectivity map. As these circuits occupy precious area on the silicon surface, the fan-in is naturally limited by their number and layout dimensions. In biology, neural systems have simuilar constraints: they have a need to reduce connectivity (by wiring) to a minimum. This results in a network topology that can be represented by a sparse adjacency matrix. To sustain this sparsity in a constantly changing environment, while maintaining high performance, the network continuously adapts and rearranges its connections, by a process called structural plasticity. Structural plasticity is necessary to learn efficient network architectures that retain computationally-relevant connections in hardware (both bio- and electronic-) system that have finite fan-in resources.
Local activity-driven synaptic pruning can solve the global problem of optimizing a network architecture. However, all structural plasticity architectures currently implemented on neuromorphic hardware use on-chip or off-chip processors [1,2] which requires global clocks or even access to lookup tables in memory to calculate some complex functions. This is not energy efficient and is not compatible with non-von Neumann neuromorphic chips.
 R. George, G. Indiveri, and S. Vassanelli, “Activity dependent structural plasticity in neuromorphic systems,” in2017 IEEEBiomedical Circuits and Systems Conference (BioCAS), pp. 1–4, 2017.
 Y. Yan, D. Kappel, F. Neumärker, J. Partzsch, B. Vogginger, S. Höppner, S. Furber, W. Maass, R. Legenstein, and C. Mayr, “Ef-ficient reward-based structural plasticity on a spinnaker 2 prototype,”IEEE Transactions on Biomedical Circuits and Systems,vol. 13, no. 3, pp. 579–591, 2019.
 G. Bellec, D. Kappel, W. Maass, and R. Legenstein, “Deep rewiring: Training very sparse deep networks,” 2017.
In this project, you will study bio-inspired methods for structural plasticity and network pruning . You will compare the pros and cons of the structural plasticity architecture that has already been implemented in the neuromorphic system with the new ones proposed in the literature (and with those existing in biology). The goal is to design a novel structural plasticity model suitable for a new generation of event driven and on-chip implementations. To validate your model, you will use a closed-loop system comprising a PC, an FPGA board, and our full-custom SNN neuromorphic chips.
You should be familiar with Python, Verilog, FPGA. Basic understanding of Spiking Neural Networks.
Zhe Su (zhesu (at) ini.uzh.ch); Prof. Giacomo Indiveri (giacomo (at) ini.uzh.ch),