Narrowing the ANN–SNN Gap for Continuous 1D Temporal Signal Classification with Multi-Scale Temporal Encoding and Sparsity-Regularized Transform Encoding

Spiking neural networks (SNNs) promise energy-efficient inference, yet on static vision benchmarks, they often trail matched ANNs under short simulation horizons. Under a matched-backbone and matched-budget protocol without extra tricks, we find that this ANN-SNN accuracy gap is consistently smaller on representative continuous 1D temporal signal classification benchmarks than on image benchmarks. We attribute this to a mechanism-level mismatch: leaky integration naturally implements causal evidence accumulation over time for native temporal signals, while static images typically require amplitude-to-spike encoding, whose finite-window estimation error becomes non-negligible at short horizons. Guided by this view, we propose a plug-and-play framework that combines Multi-Scale Temporal Encoding (MTE) and Sparsity-Regularized Transform Encoding (STE). MTE replaces naive repetition with multi-scale streams and allocates scale-aligned multi-bit integer spikes to increase per-step information density, and STE replaces a controllable fraction of LIF units with a transform-encoding neuron trained using auxiliary reconstruction and sparsity regularization, with a synthesis branch used only during training. Across diverse 1D datasets and backbone families, MTE$\times$STE consistently improves the accuracy-efficiency trade-off over standard SNN baselines and matches or occasionally surpasses ANN counterparts.

Many real-world signals, such as brain activity, heartbeats, muscle movements, speech, radio signals, and machine vibrations, change over time. This paper studies a type of brain-inspired neural network called a spiking neural network, which processes information using short bursts of activity and has the potential to be more efficient than standard neural networks. A common concern is that spiking neural networks often lose accuracy compared with standard neural networks, especially when they are allowed to run for only a few time steps. We find that this accuracy gap is much smaller for real time-varying one-dimensional signals than for static images. This is because spiking networks naturally keep short-term memory over time, which fits temporal signals better than images that must first be converted into artificial time sequences. Based on this observation, we propose a new method called MTE×STE. It gives the spiking network several time-scale views of the input signal and encourages the network to keep useful activity while suppressing redundant activity. Across ten signal classification benchmarks and several model types, our method improves the balance between accuracy and estimated computational efficiency compared with standard spiking baselines, and in some cases matches or surpasses standard neural networks. The efficiency results are based on commonly used operation-level estimates, not direct hardware power measurements.