FORGE: Forward-Only Test-Time Adaptation for Integer-Only Vision Models on Microcontrollers
Muhammad Rehan, Haider Ali, Muhammad Ali Munir, Moaz Amjad
cs.CV · cs.AR · cs.LG
Quick verdict
Forward-only recalibration revives TTA on BN-folded int8 MCUs; strong hardware validation, but eval is small-scale and its safety gate fails on many classes.
Summary
The paper addresses a deployment-specific blind spot in test-time adaptation (TTA): models on microcontrollers are quantized to int8 with batch normalization folded into convolutions, which removes both the gradients that TENT-style TTA requires and the BN statistics that forward-only recalibration requires. FORGE restores adaptability with a gradient-free per-channel correction at each former-BN site: it stores the clean per-channel output mean and standard deviation (β, |γ|) recorded at fold time, tracks an EMA of the shifted stream's per-channel statistics, and re-normalizes each channel onto its clean target. On CIFAR-10-C (ResNet-20, int8, severity 5) it recovers +20.9 points, versus +20.6 for BN-adapt on a BN-preserving model and +24.9 for the undeployable TENT. Adapting 3 of 21 layers chosen on held-out corruptions retains 93% of the benefit, and window-matched momentum (m=bs/640) sustains single-sample streaming. Results generalize to CIFAR-100-C (+14.5), MobileNetV2 on CIFAR-10-C (+20.5) and Tiny-ImageNet-C (+11.0). On an ESP32-S3 with bit-exact int8 verification, adaptation adds 8.3 mJ (6.8% of inference energy), 21.9 ms, and 12.2 KB of memory. Limitations: the mechanism is essentially AdaBN re-applied post-fold; evaluation covers only severity-5 corruptions on two small architectures, with Tiny-ImageNet-C on a single seed; the safety gate forfeits all recovery on Tiny-ImageNet-C and trades recovery for safety on CIFAR-100-C; hardware measurement covers only ResNet-20; no code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Folding BN into convolutions for integer deployment eliminates forward-only normalization adaptation entirely, dropping recovery from +20.1 to +0.0 points on CIFAR-10-C at unchanged corruption error (Fig. 2, Sec. 3.2).
- FORGE recovers +20.9 points on the deployed folded int8 model using only forward passes, matching forward-only BN-adapt (+20.6) and approaching undeployable TENT (+24.9) (Table 2).
- Adapting 3 of 21 layers, selected on held-out corruptions, retains 93% of the full-layer benefit, and the best subset exceeds adapting all layers (Sec. 4.2, Fig. 4).
- On an ESP32-S3 with bit-exact int8 verification, forward-only adaptation costs a measured 8.3 mJ (6.8% of inference energy), 21.9 ms, and 12.2 KB added memory (Sec. 4.7, Table 8).
- A batch-size-scaled momentum (m=bs/640) sustains single-sample streaming, and the method generalizes across CIFAR-100-C (+14.5), MobileNetV2 (+20.5) and Tiny-ImageNet-C (+11.0) (Secs. 4.3-4.4).
Strengths
The adaptation gap is cleanly isolated and demonstrated
Fig. 2 shows forward-only BN recalibration recovers +20.1 points on a BN-preserving int8 model but +0.0 after folding, with identical corruption error across bars — a sharp, memorable demonstration that folding, not shift, removes the remedy (Sec. 3.2).
Unusually rigorous on-device verification
The deployed C inference matches an integer reference bit-exactly (max |Δ| = 0.0, clean 92.6% vs 92.2% PyTorch), energy is integrated with a Nordic PPK2 at 100 kHz (Table 8: 121.7 mJ / 286.1 ms inference vs +8.3 mJ / 21.9 ms adapt), and memory is accounted from firmware (6.1 KB SRAM + 6.1 KB Flash, no activation-sized scratch).
Honest baseline framing and limitation reporting
Mechanistic appendix that explains, not just justifies
Flat hyperparameter and precision sensitivity
Weaknesses
Mechanism novelty is admittedly incremental
Evaluation breadth is thin for the generalization claims
The safety gate is fragile exactly where classes grow
The headline deployment regime (batch 1) lacks numeric results
Reproducibility gaps in the quantization pipeline
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
FORGE (TMLR 08/2026; arXiv 2609.01683) targets test-time adaptation (TTA) for CNNs deployed on microcontrollers, where the mandatory pipeline — BN folding into convolutions plus int8 quantization — removes both the backprop machinery gradient methods need (TENT, CoTTA) and the BN statistics forward-only recalibration needs (AdaBN, BN-adapt, LeanTTA). The claims are: (1) an adaptation gap exists and is caused by folding (+20.1 to +0.0 recovery on identical corruption error, Fig. 2); (2) a forward-only per-channel recalibration restores most of it (+20.9 of TENT's +24.9 on CIFAR-10-C, Table 2); (3) 3 of 21 layers retain 93% of the benefit; (4) batch-size-1 streaming survives via window-matched momentum m=bs/640; (5) measured on-device cost is 8.3 mJ / 21.9 ms on an ESP32-S3. The authors state upfront that novelty lies in the problem and the evidence, not the adaptation rule — an honest framing this review takes at face value and holds them to.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FORGE and what problem does it solve?
FORGE is a test-time adaptation method for vision models deployed on microcontrollers. Such models are quantized to int8 and their batch normalization layers are folded into convolutions, which removes both the machinery backpropagation needs and the statistics that existing adaptation methods update. FORGE restores adaptation by re-normalizing each folded convolution's per-channel output to clean training statistics (two constants per channel recorded at fold time), using only forward passes — no gradients, no learnable parameters, no extra forward passes.
How much energy and memory does FORGE's adaptation cost on a microcontroller?
Measured on an ESP32-S3 with a Nordic PPK2 power profiler: adaptation added 8.3 mJ (6.8% of the 121.7 mJ inference energy) and 21.9 ms on top of 286 ms inference. Memory added is 6.1 KB of SRAM for running statistics plus 6.1 KB of Flash for the clean targets across all 21 adaptation sites (784 channels) of a ResNet-20; adapting only the top-3 layers needs 1.8 KB total.
How does FORGE compare to TENT in accuracy?
On CIFAR-10-C at severity 5, FORGE recovers +20.9 accuracy points on the deployed int8 model, versus +24.9 for TENT — but TENT requires backpropagation and cannot run on a deployed integer-only model at all, so the practical on-device alternative is the unadapted model at +0.0. FORGE essentially matches forward-only BN adaptation (+20.6), which also cannot run once BN is folded.
Does FORGE work with transformers or LayerNorm?
No. The paper explicitly targets folded Conv-to-BN sites found in convolutional backbones, including depthwise-separable ones like MobileNetV2. Transformer LayerNorm is not fused into a preceding convolution in the same way, so the method does not cover transformer backbones; extending the post-fold restoration idea to other normalizers is stated as future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-5 are extracted only as scattered axis labels and caption text; plotted curve values (e.g., the exact recovery at batch size 1 in Fig. 5) cannot be read from the extraction, so those specific values are not assessed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.