Abstract
Spikes are the currency in central nervous systems for information transmission and processing. They are also believed to play an essential role in low-power consumption of the biological systems, whose efficiency attracts increasing attentions to the field of neuromorphic computing. However, efficient processing and learning of discrete spikes still remain a challenging problem. In this article, we make our contributions toward this direction. A simplified spiking neuron model is first introduced with the effects of both synaptic input and firing output on the membrane potential being modeled with an impulse function. An event-driven scheme is then presented to further improve the processing efficiency. Based on the neuron model, we propose two new multispike learning rules which demonstrate better performance over other baselines on various tasks, including association, classification, and feature detection. In addition to efficiency, our learning rules demonstrate high robustness against the strong noise of different types. They can also be generalized to different spike coding schemes for the classification task, and notably, the single neuron is capable of solving multicategory classifications with our learning rules. In the feature detection task, we re-examine the ability of unsupervised spike-timing-dependent plasticity with its limitations being presented, and find a new phenomenon of losing selectivity. In contrast, our proposed learning rules can reliably solve the task over a wide range of conditions without specific constraints being applied. Moreover, our rules cannot only detect features but also discriminate them. The improved performance of our methods would contribute to neuromorphic computing as a preferable choice.
Introduction
While deep learning has achieved massive success, it relies on power-hungry GPUs and the "von Neumann" architecture, which separates processing from memory. In contrast, the human brain is extraordinarily efficient, communicating via discrete electrical pulses called spikes.
Neuromorphic computing aims to mimic this efficiency. However, training Spiking Neural Networks (SNNs) is notoriously difficult because spikes are non-differentiable, making standard backpropagation challenging to apply. Existing multi-spike learning rules are often computationally heavy or limited to specific coding schemes. This research bridges the gap between biological plausibility and computational efficiency, making SNNs more viable for low-power, real-world applications.
The Simplified Neuron Model
The authors propose a simplified current-based Leaky Integrate-and-Fire (LIF) model. In this model, both synaptic inputs and firing resets are modeled using a simple unit impulse function. This allows the model to be converted into a Spike Response Model (SRM) form that is much easier to analyze mathematically.
To further improve efficiency, an event-driven scheme is adopted. Instead of checking the neuron's state at every discrete time step, the system only performs computations when a spike actually occurs. This drastically reduces the number of operations required for simulation and training.
Efficient Learning Rules (EML & EMLC)
Two novel supervised learning rules are introduced to accelerate training:
- EML (Efficient Multi-Spike Learning): This rule is based on the "Spike-Threshold-Surface" (STS). It simplifies the gradient calculation by proving that the impact of previous output spikes on the derivative of the firing threshold can be neglected without losing accuracy.
- EMLC (EML-Current): A further simplification that ignores the threshold surface entirely. It adjusts weights based only on the neuron's current state—using the maximum subthreshold voltage for potentiation (LTP) and the minimum reset potential for depression (LTD).
Key Results and Insights
Performance and Speed
The proposed rules demonstrate significant improvements in computational efficiency. EMLC is approximately 10x faster than the Multi-Spike Tempotron (MST) in terms of CPU execution time. In feature detection tasks, EML was roughly 2x faster than other state-of-the-art multi-spike rules like TDP.
Robustness and Generalization
The learning rules maintained 100% accuracy even when subjected to 100ms of temporal jitter or 40% random spike deletion. Furthermore, the rules proved effective across both time-based coding (precise timing) and rate-based coding (spike frequency).
Single-Neuron Multi-Classing
Traditionally, a network of neurons is required to classify multiple categories. This work demonstrates that by training a single neuron to fire a specific number of spikes for different categories (e.g., 5 spikes for Class A, 10 for Class B), a single unit can handle multi-class problems effectively.
Selectivity Loss in STDP
A notable finding is that unsupervised Spike-Timing-Dependent Plasticity (STDP) can lose its ability to detect features over long training periods. Because depression (LTD) eventually dominates, the neuron's weights can shrink until it stops firing entirely. The proposed supervised rules avoid this plateau, reliably solving feature detection tasks over a wide range of conditions.
Conclusion
This research provides a more efficient framework for processing and learning with spikes. By simplifying the neuron model and adopting event-driven computation, the EML and EMLC rules offer a high-performance alternative for neuromorphic computing. Future work will focus on extending these rules to multi-layer architectures and validating them on actual neuromorphic hardware like Intel's Loihi or IBM's TrueNorth.
Summary updated