Research map / Architectures and efficiency / Quantization
NeuralFuse: Learning to Recover the Accuracy of Access-Limited Neural Network Inference in Low-Voltage Regimes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuralFuse, a learned input-transformation module (encoder-decoder generator) trained with an expectation-over-perturbed-models optimizer to recover accuracy of quantized DNNs under SRAM low-voltage bit errors, without retraining the base model.
Problem
DNN accuracy degrades under random SRAM bit flips caused by aggressive supply-voltage reduction for energy savings, and existing mitigations require hardware changes or error-aware retraining that is infeasible for access-limited models (cloud APIs, non-configurable hardware).
Approach
A trainable, input-aware additive perturbation generator placed in front of a frozen deployed model, trained with EOPM (expectation over sampled bit-error perturbed models) so transformed inputs stay correct under bit errors and at nominal voltage; for restricted access, the generator is trained on a white-box surrogate and transferred.
Key results
- At 1% bit-error rate on CIFAR-10 ResNet18, ConvL NeuralFuse raises perturbed accuracy from 38.9% to 87.8% (about 49-point recovery).
- NeuralFuse reduces SRAM access energy by up to 24% (ResNet50 with ConvS/UNetS reaches about 29.9%; ResNet18 19-29%) at 1% BER.
- Transfer from ResNet18/VGG19 surrogates to other target models in restricted access recovers up to ~48 points (e.g., VGG19 with UNetL to VGG19 at 1% BER: 84.3% PA).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add NeuralFuse input transformation | ResNet18 · CIFAR-10, 1% BER | accuracy: 38.9 ± 12.4 -> 87.8 ± 1.7 | Helped |
| Add NeuralFuse input transformation | ResNet18 · CIFAR-10, 0.5% BER | accuracy: 70.1 ± 11.6 -> 90.0 ± 0.2 | Helped |
| Use larger generators rather than smaller generators | NeuralFuse · CIFAR-10 and GTSRB | perturbed accuracy | Helped |
| Set the number of EOPM perturbed models per iteration to a larger value (ablation) | NeuralFuse ConvL and ConvS · CIFAR-10, ResNet18, 1% BER | perturbed accuracy and stability | Helped |
| Set lambda to a smaller value (ablation) | NeuralFuse · CIFAR-10, ResNet18 and VGG19 | clean accuracy: 92.2; 90.2 | Helped |
| Set lambda to a smaller value (ablation) | NeuralFuse · CIFAR-10, ResNet18 and VGG19 | perturbed accuracy: 43.6 ± ~13; 36.8 ± ~12 | Hurt |
| Replace input-aware transformation with a universal input perturbation (ablation) | NeuralFuse on ResNet18, ResNet50, and VGG19 · CIFAR-10, 1% BER | perturbed accuracy: 38.9 ± 12.4 -> 37.9 ± ~11; 26.1 ± ~~9.4 -> 21.0 ± 5.9; 36.0 ± 12.0 -> 35.3 ± ~12 | Hurt |
| Train the generator with a larger BER and transfer it to a lower-BER target | NeuralFuse using VGG19 surrogate and UNetL · CIFAR-10, VGG19 and VGG11 at 0.5% BER | accuracy: 85.86% (as against 84.99% for original VGG19); 84.81% (as against 82.42% for original VGG11) | Helped |
| Add NeuralFuse to VGG11 with ConvL or DeConvL | VGG11 · CIFAR-10, 1% BER | MAC-based energy saving: -21.8; -11.5 | Hurt |
| Reduce model weight precision to 3 bits (ablation) | GTSRB pre-trained ResNet18 with NeuralFuse · GTSRB, low voltage with bit errors | accuracy recovery | Hurt |
| Add ensemble training of white-box surrogate base models | NeuralFuse in restricted-access settings · CIFAR-10 transfer | transferability | Helped |
Surprising findings
- Universal input perturbation (a non-input-aware baseline) performs much worse than NeuralFuse, and for some models it even lowers accuracy (e.g., ResNet50 at 1% BER: RP -5.1).
- For VGG11, large generators ConvL and DeConvL increase total energy (MAC-based saving -21.8% and -11.5%).
Limitations
- Latency overhead of the added generator (about 2-5 ms per inference at batch size 1) is an acknowledged trade-off.
- Energy savings are estimated via SCALE-SIM and Cadence SRAM simulations rather than measured hardware; some large generators increase energy for VGG11.
- Proof-of-concept limited to CNN image classifiers with simulated random bit errors; not evaluated on transformers or language models.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 GTSRB ImageNet-10
Benchmarks: CIFAR-10 CIFAR-100 GTSRB ImageNet-10
Metrics: clean accuracy perturbed accuracy recovery percentage energy saving (%) MAC-based energy saving latency
Baselines: universal input perturbation (UIP) base model without NeuralFuse
Builds on: expectation over transformation (EOT) input-aware backdoor trigger generator UNet adversarial training
Closest papers by meaning
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- Are Conventional SNNs Really Efficient? A Perspective from Network Quantization (CVPR 2024 · similarity 0.84)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.84)
- Sorbet: A Neuromorphic Hardware-Compatible Transformer-Based Spiking Language Model (ICML 2025 · similarity 0.84)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.84)
- LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.