About this Research Topic
The human brain is a very complex system, comprised of trillions of interconnected synapses. Sensory Information is transferred between neurons by electrical pulses, i.e., spikes. Spiking neural network (SNN) is a brain-inspired computing model, using spatiotemporal dynamics to mimic human cortex mechanism using binary spike signals to communicate data between neurons. Among them, deep SNNs are capable of more accurately classifying complex input patterns.
However, deep SNNs have not quite reached the accuracy performance comparable to conventional real-valued artificial neural network (ANN) in many machine learning tasks. One solution to overcome this challenge is to train first a conventional ANN model offline and then transfer the learned weights into an equivalent deep SNN topology. But this precludes the possibility of on-chip learning or online learning, which are of great significance to adaptive neuromorphic systems. It is therefore desirable to devise learning rules directly applicable to the spike-domain for deep SNN learning. Nevertheless, training deep SNNs remains difficult due to the complex neural dynamics, discrete and non-differentiable spike trains, as well as proper error representation and back-propagation in the form of individual spikes or spike statistics.
In this research topic, we aim to explore high-accuracy deep spiking neural network models and learning algorithms. Specifically, in terms of information encoding method, neuron model, network topology, and, of particular interest, the direct spike-domain BP-based or BP-like deep learning algorithms. We will explore innovative techniques that can overcome the discontinues and nonlinear difficulties encountered in deep SNN training.
The scope of the research topic includes but are not limited to:
- Spike-domain BP-based learning algorithms of deep spiking neural networks.
- Spike-domain BP-like learning algorithms of deep spiking neural networks.
- Neural information encoding and representation mechanisms suitable for high-accuracy deep SNNs.
- New datasets for deep SNN evaluations.
- Real-world applications of deep SNNs.
- Analysis of deep SNN performance and inspirations from SNNs cast onto neuroscience research.
However, deep SNNs have not quite reached the accuracy performance comparable to conventional real-valued artificial neural network (ANN) in many machine learning tasks. One solution to overcome this challenge is to train first a conventional ANN model offline and then transfer the learned weights into an equivalent deep SNN topology. But this precludes the possibility of on-chip learning or online learning, which are of great significance to adaptive neuromorphic systems. It is therefore desirable to devise learning rules directly applicable to the spike-domain for deep SNN learning. Nevertheless, training deep SNNs remains difficult due to the complex neural dynamics, discrete and non-differentiable spike trains, as well as proper error representation and back-propagation in the form of individual spikes or spike statistics.
In this research topic, we aim to explore high-accuracy deep spiking neural network models and learning algorithms. Specifically, in terms of information encoding method, neuron model, network topology, and, of particular interest, the direct spike-domain BP-based or BP-like deep learning algorithms. We will explore innovative techniques that can overcome the discontinues and nonlinear difficulties encountered in deep SNN training.
The scope of the research topic includes but are not limited to:
- Spike-domain BP-based learning algorithms of deep spiking neural networks.
- Spike-domain BP-like learning algorithms of deep spiking neural networks.
- Neural information encoding and representation mechanisms suitable for high-accuracy deep SNNs.
- New datasets for deep SNN evaluations.
- Real-world applications of deep SNNs.
- Analysis of deep SNN performance and inspirations from SNNs cast onto neuroscience research.
Keywords: spiking neural network, neuromorphic computing, deep learning, spike-domain learning, back-propagation
Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
