Fast and energy-efficient neuromorphic deep learning with first-spike times

The evolution of neuromorphic computing has aimed to emulate biological neural systems' efficiency and parallelism. Early models like SpikeProp and Tempotron introduced spike-based learning but faced limitations in biological realism and hardware implementation. Recent advances, such as IBM's TrueNorth and BrainScaleS, leverage event-driven processing and spike timing, especially TTFS coding, which encodes information in the timing of the first spike. Despite these developments, training deep, biologically plausible spiking networks remains challenging due to non-differentiability of spike events and hardware variability. Prior approaches often relied on approximations or non-leaky models, limiting their biological fidelity and scalability. This context motivates the current work, which seeks to derive exact analytical gradients for leak-type neurons, enabling efficient supervised learning directly based on spike timing.

The core challenge is enabling error backpropagation in networks of leaky integrate-and-fire neurons, which exhibit non-linear, non-differentiable spike generation dynamics. Traditional methods struggle with the discontinuities introduced by leak and threshold crossing, hindering deep supervised learning. Additionally, hardware constraints such as limited parameter resolution, device mismatch, and noise complicate training. Addressing these issues is crucial for deploying scalable, energy-efficient neuromorphic systems capable of complex tasks, bridging the gap between biological plausibility and practical AI applications.

The primary innovation is the derivation of exact formulas for the first spike time in LIF neurons with specific time constant ratios, leveraging the Lambert W function. This enables the computation of precise, differentiable gradients for spike times, facilitating error backpropagation. The approach supports multi-layer networks and is compatible with hardware constraints, unlike previous methods relying on approximations or non-biological models. It introduces a unified framework for supervised learning directly on spike timing, maintaining biological realism while enabling deep learning. The validation on BrainScaleS-2 hardware demonstrates its practical potential, marking a significant advance in neuromorphic deep learning.

• �� Model the membrane potential dynamics of LIF neurons with CuBa synapses, incorporating leak conductance and synaptic currents.
• �� Derive analytical expressions for the first spike time T based on specific time constant ratios (τm=τs and τm=2τs), using the Lambert W function.
• �� Formulate the gradient of spike time T with respect to synaptic weights and presynaptic spike times, ensuring differentiability.
• �� Develop a backpropagation algorithm that propagates errors through spike times, enabling supervised learning.
• �� Implement the training procedure on datasets like MNIST and Yin-Yang, tuning hyperparameters for optimal performance.
• �� Emulate the entire framework on BrainScaleS-2 hardware, measuring classification accuracy, latency, and energy consumption.

The experiments involved training deep networks on MNIST and Yin-Yang datasets, with input features converted into spike times. Hyperparameters such as learning rate and time constant ratios were optimized for best accuracy. Hardware robustness was tested by introducing parameter noise and fixed-pattern variations, assessing impact on classification performance. Performance metrics included accuracy, training epochs, inference latency, and energy consumption. Comparative analysis with existing spike-time coding methods highlighted the advantages of the proposed exact gradient approach. Ablation studies examined the influence of different time constants and network depths, confirming the method's scalability and robustness.

The method achieved 97.1% accuracy on MNIST with 150 epochs, outperforming previous spike-time coding approaches. Hardware emulation demonstrated inference within 10μs, with energy per classification at 8.4μJ. On the Yin-Yang dataset, accuracy reached 95.9%, with clear class separation in spike timing. Hardware robustness tests showed minimal performance degradation under parameter noise and device mismatch, confirming the approach's practicality. These results establish the method as a promising candidate for real-time neuromorphic AI, combining high accuracy, speed, and energy efficiency.

The framework is suitable for low-power, high-speed neuromorphic systems in applications like autonomous robots, real-time sensory processing, and edge AI devices. Its robustness to hardware imperfections makes it practical for deployment in diverse neuromorphic platforms. The ability to train deep networks directly on spike timing enables complex pattern recognition tasks, facilitating advances in brain-inspired computing and energy-efficient AI hardware. Future integration with adaptive learning and multi-modal data processing could further expand its utility.

Current implementation relies on specific time constant ratios, which may limit flexibility in diverse hardware environments. The approach's performance under extreme noise or hardware variability requires further validation. Scalability to very large networks remains to be demonstrated, and the computational complexity of gradient calculations could pose challenges. Additionally, the method primarily addresses supervised learning; extending to unsupervised or reinforcement learning paradigms is an open challenge. Further hardware optimization is needed to fully realize its potential in real-world applications.