Source-linked AI summary
RMP-SNN: Residual Membrane Potential Neuron for Enabling Deeper High-Accuracy and Low-Latency Spiking Neural Network
Bing Han, Gopalakrishnan Srinivasan, Kaushik Roy
TL;DR
Converted SNNs can lose accuracy relative to their source ANNs and require many inference time-steps, with hard-reset neurons discarding residual membrane potential. The paper replaces them with soft-reset RMP neurons and reports near loss-less conversion plus higher accuracy with 2–8 times fewer time-steps across architectures and datasets.
Problem
ANN-SNN conversion can lose accuracy and require sizable inference time because hard-reset neurons discard membrane potential above threshold.
Method
The paper uses soft-reset Residual Membrane Potential neurons that retain above-threshold potential and applies threshold balancing for ANN-SNN conversion.
Results
RMP-SNN surpasses hard-reset converted SNN accuracy using 2–8 times fewer inference time-steps across architectures and datasets, including 93.63% top-1 on CIFAR-10 and 70.93% top-1 on CIFAR-100.
Takeaways & Limitations
RMP neurons enable near loss-less ANN-SNN conversion for deep VGG and residual architectures evaluated on CIFAR-10, CIFAR-100, and ImageNet.
Takeaways & Limitations
Lowering thresholds increases spiking activity and can adversely impact computational efficiency, while hard-reset conversion faces a latency–accuracy trade-off.
Abstract
from arXiv · showhide
Spiking Neural Networks (SNNs) have recently attracted significant research interest as the third generation of artificial neural networks that can enable low-power event-driven data analytics. The best performing SNNs for image recognition tasks are obtained by converting a trained Analog Neural Network (ANN), consisting of Rectified Linear Units (ReLU), to SNN composed of integrate-and-fire neurons with "proper" firing thresholds. The converted SNNs typically incur loss in accuracy compared to that provided by the original ANN and require sizable number of inference time-steps to achieve the best accuracy. We find that performance degradation in the converted SNN stems from using "hard reset" spiking neuron that is driven to fixed reset potential once its membrane potential exceeds the firing threshold, leading to information loss during SNN inference. We propose ANN-SNN conversion using "soft reset" spiking neuron model, referred to as Residual Membrane Potential (RMP) spiking neuron, which retains the "residual" membrane potential above threshold at the firing instants. We demonstrate near loss-less ANN-SNN conversion using RMP neurons for VGG-16, ResNet-20, and ResNet-34 SNNs on challenging datasets including CIFAR-10 (93.63% top-1), CIFAR-100 (70.93% top-1), and ImageNet (73.09% top-1 accuracy). Our results also show that RMP-SNN surpasses the best inference accuracy provided by the converted SNN with "hard reset" spiking neurons using 2-8 times fewer inference time-steps across network architectures and datasets.
1. Introduction
SNNs offer sparse, spike-based computation for power-efficient data analytics, but their training remains less mature than ANN training. ANN-SNN conversion is the strongest-performing approach, and this paper proposes RMP neurons to reduce conversion loss and inference latency.
- SNNs exploit sparse spike-based computation and communication to improve computational efficiency on specialized neuromorphic hardware.
- STDP-based learning methods have remained restricted to shallow SNNs of ≤5 layers and achieve considerably lower accuracy than ANNs.
- ANN-SNN conversion maps trained ReLU-based ANNs to integrate-and-fire SNNs with layer-wise firing thresholds chosen to align spiking rates with analog activations.
- RMP neurons retain residual membrane potential above threshold after firing instead of hard-resetting, alleviating information loss during ANN-SNN conversion.
- The paper evaluates RMP-based VGG-16 and residual SNNs on CIFAR-10, CIFAR-100, and ImageNet, targeting near loss-less conversion and fewer inference time-steps.
2. Related Work
Prior work primarily converts ReLU-based ANNs into IF-neuron SNNs, using threshold-balancing schemes to align firing rates with ANN activations. These methods scale to complex tasks, but efficient conversion depends critically on threshold initialization.
- ANN-SNN conversion maps backpropagation-trained ReLU ANNs to IF-neuron SNNs for deep image-recognition and natural-language-processing tasks.
- Efficient conversion requires layer-wise threshold initialization so SNN spiking rates remain proportional to corresponding ReLU activations.
- Model-based threshold balancing uses ANN weights, whereas data-based schemes use training data and weights, with later methods incorporating SNN spiking statistics.
- A notable alternative trains ANNs with rate-based soft-LIF nonlinearities and uses Leaky-Integrate-and-Fire neurons during inference.
3. ANN-SNN Conversion
ANN-SNN conversion maps ReLU activations to IF firing rates through threshold selection, but hard resets discard residual membrane potential and distort linearity. RMP soft reset preserves that residual potential to improve the conversion's accuracy-latency trade-off.
- ANN-SNN conversion maps ReLU activations to IF spiking rates by selecting layer-specific firing thresholds.Thresholds may be initialized using model-based or data-based schemes, and input images can be encoded as Poisson spike trains.
- IF neurons integrate weighted spike inputs and fire when membrane potential exceeds the firing threshold, then hard-reset to zero.The membrane-potential dynamics use transferred ANN weights and incoming spike trains.
- Hard reset reduces output firing rates because residual potential above threshold is discarded at spike times.For weighted inputs totaling 3Vth across three time-steps, the example produces only two spikes instead of three.
- Hard-reset neurons are approximately linear only when average input amplitude is much lower than threshold, forcing a trade-off between conversion accuracy and inference latency.Higher thresholds reduce input activity and conversion loss but increase latency; lower thresholds increase activity and nonlinear distortion.
- RMP neurons use soft reset by subtracting the threshold, retaining residual potential and producing a linear input-output response over a wider input range.With the same 3Vth example, the RMP neuron produces three spikes by carrying residual potential into subsequent time-steps.
4. Residual Membrane Potential SNN
RMP neurons use soft reset to retain residual membrane potential, extending the linear operating range and reducing information loss during ANN-SNN conversion. Thresholds are initialized layer-wise from weighted-input statistics to balance accuracy, latency, and spiking activity.
- RMP neuron: Soft reset subtracts the firing threshold from Vm instead of resetting it to a fixed potential, retaining residual potential at spike times.This preserves information above threshold during inference.
- Operating range: RMP neurons provide an extended linear input-output operating range for firing thresholds compared with hard-reset IF neurons.The desired output-rate range is [fin, 1), supporting sufficient activity without excessive spiking.
- Accuracy, latency, and activity: Hard-reset IF neurons have fout ≤ fin because residual potential is ignored, inherently producing higher inference latency than RMP neurons.Lowering thresholds can reduce latency but increases spiking activity, adversely affecting computational efficiency; scaled thresholds increased activity by approximately 1-2%.
- Threshold balancing: Layer-wise thresholds are initialized sequentially from maximum weighted input sums measured using Poisson spike inputs, with scaling factor α applied to Vmax_in.Thresholds are transferred and frozen layer by layer as subsequent thresholds are estimated.
- Threshold balancing: 91.36% CIFAR-10 accuracy was achieved by ResNet-20 SNN using Vmax_in, compared with 91.47% for its corresponding ANN.Scaling the threshold by up to 0.6× converged to the same accuracy.
- Accuracy, latency, and activity: Threshold scaling improves accuracy-latency trade-offs until excessively low thresholds enter a nonlinear excessive-spiking regime and reduce accuracy.The authors recommend α ≤0.6 in this example and validate the initialization across architectures and datasets.
5. Results
RMP-SNN achieves near lossless ANN-to-SNN conversion across VGG-16 and ResNet experiments while reducing inference latency and maintaining low spike rates. Compared with IF baselines, reduced-threshold RMP-SNNs reach comparable accuracy with 2–8 times fewer time-steps.
- VGG-16 results: 93.63% top-1 accuracy matches the trained ANN for VGG-16 RMP-SNN on CIFAR-10 at 2048 time-steps.The IF baseline reaches 93.50% at the same latency.
- VGG-16 results: 70.93% top-1 accuracy is achieved by VGG-16 RMP-SNN on CIFAR-100 at 2048 time-steps, versus 70.77% for IF.A reduced-threshold RMP-SNN reaches 68.34% using 256 time-steps, twice as fast as the baseline using about 512.
- VGG-16 results: 73.09% accuracy is achieved by VGG-16 RMP-SNN on ImageNet at 4096 time-steps, versus 69.96% for IF.The reduced-threshold model reaches 68.93% using 512 time-steps, 4.5 times faster than IF using over 2300.
- ResNet results: 91.36% top-1 accuracy is achieved by ResNet-20 RMP-SNN on CIFAR-10 at 2048 time-steps, versus 90.45% for IF.The reduced-threshold RMP-SNN exceeds 85% accuracy in 64 time-steps, eight times faster than IF using 512.
- ResNet results: 67.82% top-1 accuracy is achieved by ResNet-20 RMP-SNN on CIFAR-100 at 2048 time-steps, versus 64.09% for IF.The reduced-threshold model reaches 64.06% using 256 time-steps, eight times faster than IF using about 2048.
- ResNet results: 69.89% accuracy is achieved by ResNet-34 RMP-SNN on ImageNet at 4096 time-steps, versus 65.47% for IF.The reduced-threshold model reaches 60.08% using 512 time-steps, seven times faster than IF using more than 3500.
6. Conclusion and Discussion
The paper proposes an ANN-to-SNN conversion technique using RMP neurons and threshold balancing, achieving strong accuracy and low conversion loss across deep architectures and datasets.
- RMP spiking neurons retain residual membrane potential after firing, better mimicking ReLU behavior and alleviating information loss during conversion.
- Threshold balancing alleviates spike rate vanishing and improves the latency and scalability of RMP-SNNs to very deep architectures.
- The authors implement VGG and residual networks using the proposed conversion-based training on CIFAR-10, CIFAR-100, and ImageNet.
- RMP-SNNs achieve the best accuracies and lowest conversion loss among state-of-the-art methods across all tested architectures and datasets.