Learning intrinsic neural adaptation to enhance sRNN computational power
The neuromorphic community is in need of efficient algorithms for spatiotemporal processing, which can guide the behavior of autonomous robots in complex environments. Spiking recurrent neural networks (sRNN) are suitable candidates for such efficient processing, however, it remains a challenge to stabilize their dynamics. Recent theoretical work suggests that having adapting neurons within a sRNN, due to their longer time constants, helps with this stabilization and allows the encoding of richer and longer spatiotemporal patterns in the network (Bellec et. al, 2018; Nicola and Clopath, 2017). Bellec and colleagues also show that by adding adaptation, the computational performance of sRNN matches that of LSTMs.
Interestingly, we have recently described a form of intrinsic plasticity in which neurons tune their spiking pattern -from more regular to adapting and also burst spiking patterns- after a few minutes of structured subthreshold input stimulation (Soldado-Magraner et al., 2018). In the previously mentioned sRNN theoretical work (Bellec et. al, 2018; Nicola and Clopath, 2017), the adaptation variable of the neurons within the network is manually adjusted (and fixed) for the different computational tasks the network is set to perform. Inspired by our biological observations, we would like instead to learn this parameter, with the hypothesis that it will help stabilize sRNN dynamics and align them to the temporal statistics of the different computational tasks.
We will explore different supervised algorithms in order to learn, for example, the adaptation time constant or the adapting coupling variable of the AdEx neurons within a sRNN. If we succeed on the training, we will team up with the CortexControl group to implement the trained network into the Dynapse. Possible unsupervised implementations of the learning will also be discussed during the course of the workshop.
Literature can be found on the Wiki.
Timetable
| Day | Time | Location |
|---|---|---|
| Wed, 24.04.2019 | 21:00 - 21:30 | Panorama room |
| Fri, 26.04.2019 | 14:00 - 15:00 | Panorama room |
| Sat, 27.04.2019 | 14:00 - 16:00 | Panorama room |
| Tue, 30.04.2019 | 14:00 - 15:00 | Panorama room |
| Wed, 01.05.2019 | 14:00 - 15:00 | Panorama rooom |
| Thu, 02.05.2019 | 14:00 - 15:00 | Sala Panorama |