Long short-term memory and Learning-to-learn in networks of spiking
The brain carries out demanding computations and learning processes with
recurrent networks of spiking neurons (RSNNs). But computing and learning
capabilities of currently available RSNN models have remained poor,
especially in comparison with the performance of recurrent networks of
artificial neurons, such as Long Short-Term Memory (LSTM) networks. in this
article, we investigate whether deep learning can improve RSNN performance.
We applied backpropagation through time (BPTT), augmented by biologically
inspired heuristics for synaptic rewiring, to RSNNs whose inherent time
constants were enriched through simple models for adapting spiking neurons.
We found that the resulting RSNNs approximate, for the first time, the
computational power of LSTM networks on two common benchmark tasks.
Furthermore, our results show that recent successes with applications of
Learning-to-Learn (L2L) to LSTM networks can be ported to RSNNs. This opens
the door to the investigation of L2L in data-based models for neural networks
of the brain, whose activitiy can - unlike that of LSTM networks - be
compared directly with recordings from neurons in the brain. In particular,
L2L shows that RSNNs can learn large families of non-linear transformations
from very few examples, using previously unknown network learning mechanisms.
Furthermore, meta-reinforcement learning (meta-RL) shows that LSNNs can learn
and execute complex exploration and exploitation strategies.
Reference: G. Bellec, D. Salaj, A. Subramoney, R. Legenstein, and
Long short-term memory and learning-to-learn in networks of spiking neurons.
32nd Conference on Neural Information Processing Systems (NIPS 2018),
Montreal, Canada, 2018.