17 July 2020 in Nature Communications
Recurrently connected networks of spiking neurons underlie the astounding information processing capabilities of the brain. Yet in spite of extensive research, how they can learn through synaptic plasticity to carry out complex network computations remains unclear. We argue that two pieces of this puzzle were provided by experimental data from neuroscience. A mathematical result tells us how these pieces need to be combined to enable biologically plausible online network learning through gradient descent, in particular deep reinforcement learning. This learning method–called e-prop–approaches the performance of backpropagation through time (BPTT), the best-known method for training recurrent neural networks in machine learning. In addition, it suggests a method for powerful on-chip learning in energy-efficient spike-based hardware for artificial intelligence.
19 August 2019 in bioRxiv
Recurrently connected networks of spiking neurons underlie the astounding information processing capabilities of the brain. But in spite of extensive research, it has remained open how they can learn through synaptic plasticity to carry out complex network computations. We argue that two pieces of this puzzle were provided by experimental data from neuroscience. A new mathematical insight tells us how these pieces need to be combined to enable biologically plausible online network learning through gradient descent, in particular deep reinforcement learning. This new learning method – called e-prop – approaches the performance of BPTT (backpropagation through time), the best known method for training recurrent neural networks in machine learning. In addition, it suggests a method for powerful on-chip learning in novel energy-efficient spike-based hardware for AI.
25 December 2018 in arXiv
Recurrent networks of spiking neurons (RSNNs) underlie the astounding computing and learning capabilities of the brain. But computing and learning capabilities of RSNN models have remained poor, at least in comparison with artificial neural networks (ANNs). We address two possible reasons for that. One is that RSNNs in the brain are not randomly connected or designed according to simple rules, and they do not start learning as a tabula rasa network. Rather, RSNNs in the brain were optimized for their tasks through evolution, development, and prior experience. Details of these optimization processes are largely unknown. But their functional contribution can be approximated through powerful optimization methods, such as backpropagation through time (BPTT). A second major mismatch between RSNNs in the brain and models is that the latter only show a small fraction of the dynamics of neurons and synapses in the brain. We include neurons in our RSNN model that reproduce one prominent dynamical process of biological neurons that takes place at the behaviourally relevant time scale of seconds: neuronal adaptation. We denote these networks as LSNNs because of their Long short-term memory. The inclusion of adapting neurons drastically increases the computing and learning capability of RSNNs if they are trained and configured by deep learning (BPTT combined with a rewiring algorithm that optimizes the network architecture). In fact, the computational performance of these RSNNs approaches for the first time that of LSTM networks. In addition RSNNs with adapting neurons can acquire abstract knowledge from prior learning in a Learning-to-Learn (L2L) scheme, and transfer that knowledge in order to learn new but related tasks from very few examples. We demonstrate this for supervised learning and reinforcement learning.
27 January 2021 in bioRxiv
Spike-based neural network models have so far not been able to reproduce the capability of the brain to learn from very few, often even from just a single example. We show that this deficiency of models disappears if one allows synaptic weights to store priors and other information that optimize the learning process, while using the network state to quickly absorb information from new examples. For that, it suffices to include biologically realistic neurons with spike frequency adaptation in the neural network model, and to optimize the learning process through meta-learning. We demonstrate this on a variety of tasks, including fast learning and deletion of attractors, adaptation of motor control to changes in the body, and solving the Morris water maze task – a paradigm for fast learning of navigation to a new goal. Significance Statement It has often been conjectured that STDP or other rules for synaptic plasticity can only explain some of the learning capabilities of brains. In particular, learning a new task from few trials is likely to engage additional mechanisms. Results from machine learning show that artificial neural networks can learn from few trials by storing information about them in their network state, rather than encoding them in synaptic weights. But these machine learning methods require neural networks with biologically unrealistic LSTM (Long Short Term Memory) units. We show that biologically quite realistic models for neural networks of the brain can exhibit similar capabilities. In particular, these networks are able to store priors that enable learning from very few examples.