[Paper] Direct-to-Event Spiking Neural Network Transfer

Source: arXiv - 2605.07207v1

Overview

The paper tackles a practical bottleneck in spiking neural networks (SNNs): models that are trained with direct‑coding (continuous surrogate activations) are easy to optimise but end up consuming far more energy than true event‑based SNNs when deployed on neuromorphic chips. The authors ask: Can we take a pretrained direct‑coded SNN and convert it into an efficient event‑driven version without sacrificing accuracy? Their systematic study is the first to address this “transfer” problem and proposes concrete techniques to bridge the gap.

Key Contributions

• First systematic analysis of the challenges involved in converting direct‑coded SNNs to event‑based inference.
• Quantitative benchmark showing the energy‑efficiency gap between direct‑coded and event‑based SNNs across several standard vision datasets.
• Transfer pipeline comprising (1) surrogate‑to‑spike calibration, (2) adaptive threshold scaling, (3) spike‑rate regularisation, and (4) post‑conversion fine‑tuning.
• Empirical validation that the pipeline recovers >95 % of the original accuracy while cutting estimated energy consumption by 3–7× on typical neuromorphic hardware.
• Open‑source implementation and a reusable “pretrained‑SNN database” to encourage community adoption.

Methodology

1. Baseline Direct‑Coded Training – The authors start from SNNs trained with standard surrogate‑gradient back‑propagation (e.g., using a piecewise‑linear surrogate for the spiking non‑linearity).
2. Problem Decomposition – They identify three sources of inefficiency when switching to event‑based inference:
   • Mismatch between surrogate activations and actual spike counts.
   • Fixed firing thresholds that are sub‑optimal for the learned weight distribution.
   • Over‑active neurons that generate unnecessary spikes.
3. Conversion Steps
   • Surrogate‑to‑Spike Calibration: Fit a monotonic mapping from surrogate activation values to expected spike counts per timestep, using a small calibration dataset.
   • Adaptive Threshold Scaling: Rescale each neuron’s firing threshold based on the calibrated mapping to preserve the original firing probability distribution.
   • Spike‑Rate Regularisation: Apply a lightweight L1‑penalty on spike counts during a short post‑conversion fine‑tuning phase to prune redundant spikes.
   • Fine‑Tuning with Event‑Based Loss: Run a few epochs of back‑propagation where the forward pass is fully event‑driven, allowing the network to adjust weights to the new spiking dynamics.
4. Energy Estimation – Energy consumption is approximated using the standard neuromorphic cost model (≈ 1 pJ per spike, negligible static cost), enabling a direct comparison between the original and converted models.

Results & Findings

• Accuracy loss is typically <1 % after conversion, well within the variance of the original training.
• Energy reduction ranges from 3× to 7×, confirming that the event‑based representation regains the low‑power advantage of SNNs.
• Ablation studies show that each component of the pipeline contributes: threshold scaling alone recovers ~60 % of the energy gain, while adding spike‑rate regularisation pushes it over 80 %.

Practical Implications

• Deployable SNNs: Developers can now train SNNs with the mature, gradient‑friendly direct‑coding workflow and later convert them for ultra‑low‑power inference on neuromorphic chips (e.g., Intel Loihi, IBM TrueNorth, or emerging edge ASICs).
• Model Reuse: Existing repositories of pretrained direct‑coded SNNs become instantly valuable for hardware‑constrained products—no need to retrain from scratch.
• Toolchain Integration: The conversion steps are lightweight (a few minutes of calibration and a handful of fine‑tuning epochs) and can be wrapped into existing PyTorch or TensorFlow pipelines, enabling CI/CD‑style deployment for edge AI.
• Energy‑aware Design: System architects can now factor in a predictable energy budget when choosing SNNs versus conventional ANNs, especially for battery‑operated IoT devices, wearables, or autonomous drones.

Limitations & Future Work

• Calibration Data Dependence: The surrogate‑to‑spike mapping requires a small labeled dataset; the method’s robustness to distribution shift remains to be tested.
• Hardware Specificity: Energy estimates are based on generic spike‑cost models; real‑world hardware quirks (e.g., communication overhead, memory access patterns) could affect the actual savings.
• Scalability to Very Deep Networks: Experiments were limited to networks up to ~30 layers; extending the pipeline to transformer‑style SNNs or ultra‑deep CNNs may need additional tricks.
• Future Directions the authors suggest exploring online conversion (dynamic threshold adaptation at inference time) and extending the framework to spatio‑temporal tasks such as event‑camera video processing.

Authors

• Nhan Trong Luu
• Duong Trung Luu
• Pham Ngoc Nam
• Truong Cong Thang

Paper Information

• arXiv ID: 2605.07207v1
• Categories: cs.NE
• Published: May 8, 2026
• PDF: Download PDF