Transfer Deep RNN on Neuromorphic Hardware
It is a truth universally acknowledged (well, at least by another workgroup), that the neuromorphic community is in need of efficient algorithms for spatiotemporal processing, which can guide the behaviour of autonomous robots in complex environments, AND spiking recurrent neural networks (sRNN) are suitable candidates for such efficient processing, however, it remains a challenge to stabilize their dynamics.
A recent paper (He et. al, 2019, http://minds.jacobs-university.de/uploads/papers/3158_Heetal19.pdf) showed that it is possible to transfer the dynamical properties of a well-performing neural network which has been optimized on a digital computer, onto neuromorphic hardware (Dynapse) by a computational scheme called Reservoir Transfer. Empirical results from an ECG signal monitoring task showed that the transferred reservoir with ternary weights is able to not only integrate information over a time span longer than the timescale of individual neurons but also function as an information processing medium with performance close to a standard, high precision, deterministic, non-spiking ESN.
In this workgroup, we will extend the previous work to transfer a BPTT-trained recurrent network instead of a reservoir to the Dynapse Hardware. Specifically, we will take a deep (vanilla) RNN trained on simple signal processing tasks (such as temporal MNIST digit recognition task) and transfer it into a BRIAN-simulated spiking network. If promising results can be achieved in simulation, we will then use the CortexControl interface to transfer it on Dynapse.
Timetable
| Day | Time | Location |
|---|---|---|
| Wed, 24.04.2019 | 16:30 - 17:00 | Sala Panorama |
| Thu, 25.04.2019 | 14:30 - 15:00 | Sala Panorama |
| Fri, 26.04.2019 | 15:30 - 16:00 | Disco |