Building and deploying applications to mixed-signal neuromorphic processors has historically been a difficult task. These processors often have a high degree of complexity with a low degree of parameterisation, meaning that parameters are highly constrained across a network.
Coupled with uncontrolled variation in parameters and neuron dynamics caused by fabrication non-idealities, deploying robust applications to these processors is challenging.
Uğurcan Cakal has built the most advanced and easiest to use toolchain for building and deploying SNN applications to mixed-signal NM processors.
https://doi.org/10.1088/2634-4386/ad2ec3Using differentiable computing, coupled with a hardware-aware neuron model and a model of parameter variation during training, we build applications which are robust against mismatch. For deployment we have an optimisation-based solution for parameter quantisation and mapping.
We made use of JAX-based JIT compilation at training time, which enabled us to train models on GPU, TPU or CPU. On several platforms this accelerated training by more than 3000 times, reducing training time from days to minutes.
Our approach for building and deploying SNN applications to mixed-signal processors is highly extensible, and can be adapted to additional processor families! Of course all our code is available open-source, and we encourage developers to experiment:
https://github.com/synsense/rockpool
GitHub - synsense/rockpool: A machine learning library for spiking neural networks. Supports training with both torch and jax pipelines, and deployment to neuromorphic hardware.
A machine learning library for spiking neural networks. Supports training with both torch and jax pipelines, and deployment to neuromorphic hardware. - synsense/rockpool
Dylan Richard Muir
