acceptodds
Under review as a conference paper at ICLR 2027

LayerShield-QAT: Accuracy-Preserving Integer Quantization for Physics-Grounded Edge Regression

Abstract

Integer accelerators can make scientific inversion practical at the edge, yet regression models are unusually vulnerable to output discretization: errors that leave a classifier's decision unchanged can invalidate a continuous physical estimate. We present \method, a physics-grounded multi-task regression and quantization workflow for gamma-ray well-logging inversion on 8-bit neural accelerators. The model maps 11 detector-window measurements to five coupled formation and mud properties, using shared representations and task-weighted objectives to prioritize formation density and photoelectric factor. Layer-wise diagnostics identify the input and output transformations as the dominant sources of quantization error; \method therefore quantizes the hidden representation to INT8 while retaining higher precision at the boundary layers. Training additionally injects systematic, random, and count-statistical perturbations to reduce the synthetic-to-field gap. On the reported synthetic benchmark, multi-task learning increases the fraction of predictions within application tolerances from 23.61% to 89.53% for formation density and from 19.82% to 88.15% for photoelectric factor, while reaching 98.18–99.79% on three auxiliary mud targets. Accuracy-preserved quantization and deployment on NPU SoC systems reports 3.5 CPU and 23 NPU speedups over a legacy C implementation, with stable field data trends. The results support a broader principle: for high-accuracy edge regression, precision should be allocated according to physical sensitivity and layer-wise error propagation rather than uniformly.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Regression Integer Quantization, Accuracy-preserved regression QAT, NPU, 8-bit quantization, Physics, Nuclear geophysics, subsurface exploration.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_15GMOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose an engineering workflow for deploying a NN in a geology application. They discuss quantization, ablate first & last layer choice, the targeted hardware platform, data augmentation and provide synthetic experiments.

Critical strengths

This paper shows a case-study of applied ML to a geology problem. As such, it is probably interesting and a worthwhile read to that specific community.

Critical weaknesses

This paper is more of a lab-report than a new scientific contribution to ICLR. I see novelty as the major weakness, along with a lack of isolated baselines: The central LayerShield method of retaining higher precision at the first and last layers while quantizing the rest with standard QAT losely resembles established quantization practice, including prior work cited, such as PACT. The main accuracy table compares MLP versus multi-task learning, without isolating quantization. Speed-up claims wrt. to a legacy C implementation are somewhat arbitrary - presumably this is a CPU vs. accelerator comparison. I can see many AI formulations, which isn't necessarily bad but also not entirely clear how we as a community want to deal with this.

Questions for authors
  • What is different between LayerShield-QAT from established QAT approaches?
  • Compare different aspects of your QAT approach: FP32 everywhere on one side to LayerShield at the other end, including first/last layer choices.
  • Your graphics are low-resolution and partly impossible to read.
Reviewer_GinFOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors target discretization errors for edge inference and present a noise-aware, multi-task learning based solution with selective boundary-layer QAT. The authors claim multi-task learning increases synthetic tolerance compliance and the selective-layer part reduces quantization error. The authors show NPU speedup over legacy C code.

Critical strengths

The idea did not translate into substantial experiments and results: the physics-grounded part is just a task-weighted MSE, which is just a heuristic. There are no quantitative results tables for the selective-layer precision claim which is part of the main claims and they show only a hard to read image on synthetic data.

Comparisons to baselines: the main contribution is their method outperforming INT8 PTQ, and full-model QAT, and there are no results quantitative or qualitative to support this.

Critical weaknesses

Many of the claimed results are not supported in a fact-base manner that is convincing

Questions for authors
  1. Please show the full quantitative benchmark comparisons as a table.
  2. Hardware costs for the NPU implementation are missing; no reproducibility material either. The reported NPU speedups are different in the abstract vs Sec.1 and its impossible for a reader to verify this themselves.
  3. The physics-grounding being just a simple heuristic seems like a missed opportunity. Did you consider other more explicit ways to ground your QAT workflow in useful physics? Have you considered looking at the layer-wise precision sensitivity directly through the local Jacobians?
Reviewer_i1R1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents LayerShield-QAT, a selective quantization method for high-precision physical regression on edge devices, using gamma-ray well-logging inversion as the application scenario. The model predicts five formation- and mud-related parameters from 11 detector-window measurements. It employs a shared multilayer perceptron (MLP) and a task-weighted loss for multi-task learning, with systematic perturbations, Gaussian noise, and Poisson noise incorporated during training. Based on layer-wise normalized distortion analysis, the authors retain higher numerical precision in the first and last layers while applying 8-bit integer quantization-aware training (QAT) to the intermediate hidden layers. The paper uses 100,000 synthetic samples for training and evaluation, reporting improvements in tolerance accuracy from multi-task learning over a conventional MLP, as well as reductions in quantization-induced errors for key outputs through selective quantization. In addition, the authors present field well-log curves and inference acceleration results on edge processors to illustrate the method’s application and deployment potential.

Critical strengths

The paper addresses high-precision physical regression in gamma-ray well logging by designing a multi-task learning model and evaluation metrics around practical error tolerances, giving the work clear application value. Beyond experiments on synthetic data, it provides preliminary demonstrations using field well-log curves and reports inference acceleration on edge central processing units (CPUs) and neural processing units (NPUs), offering initial evidence of its practical deployment potential.

Critical weaknesses
  1. The paper lists the floating-point multi-task model, post-training quantization, full-network QAT, and other quantization strategies as comparators, but does not present a complete comparison of their errors and tolerance accuracies across all five physical outputs. Table 1 primarily compares a conventional MLP with the multi-task model and therefore does not directly substantiate LayerShield-QAT’s ability to preserve accuracy under quantization. Section 5.2 claims that quantization-induced errors for key outputs are reduced by more than 50%, but does not clearly specify the corresponding error metric, comparison baseline, or numerical results, making this central claim difficult to verify. The hardware experiments mainly report speedups relative to a legacy C implementation rather than controlled comparisons of different quantization strategies using the same network on the same hardware. Furthermore, the specific precision retained in the first and last layers and how these layers are executed are insufficiently described. Consequently, the current results do not disentangle the respective contributions of model replacement, the inference framework, hardware acceleration, and selective quantization.
  2. The proposed strategy retains higher precision in the first and last layers primarily based on layer-wise normalized activation distortion, but the paper does not establish how this metric relates to the final physical-parameter prediction errors or tolerance accuracy. Larger local activation distortion does not necessarily imply a greater impact on the final task error. It therefore remains unclear whether this diagnostic provides a task-relevant basis for precision allocation or merely supports a fixed policy of protecting the first and last layers in the current network. As currently described, the method primarily integrates multi-task learning, task weighting, noise augmentation, and selective quantization, without presenting a new precision-allocation or quantization-optimization mechanism. Such integration can have practical value, but sufficiently controlled experiments are needed to establish its distinctive benefits. The paper currently lacks ablations that separately isolate the contributions of the multi-task architecture, weighted loss, noise augmentation, and boundary-layer precision preservation, leaving both the sources of the substantial performance gains and the conditions under which the method is effective insufficiently clear.
Questions for authors
  1. Please provide a comparison of the floating-point multi-task model, 8-bit post-training quantization, full-network 8-bit QAT, and LayerShield-QAT in terms of regression errors and tolerance accuracy across all five outputs. What specific metric, baseline, and numerical results support the claim in Section 5.2 that quantization-induced errors are reduced by more than 50%? What precision is actually used for the first and last layers during deployment, on which types of processing units are these layers executed, and were the reported speedups obtained using the same selectively quantized model?
  2. Have the authors analyzed the correlation between the normalized distortion metric dℓd_\ell and the final regression error or tolerance accuracy? Are there experiments that quantize layers individually or preserve higher precision in the first and last layers separately, demonstrating that this metric identifies the layers to which final predictions are most sensitive rather than merely reflecting local numerical discrepancies?
  3. Do the conventional MLP and the multi-task model in Table 1 have comparable parameter counts, and are they trained using the same data, training budgets, and hyperparameter-tuning criteria? Could existing results or targeted ablations distinguish the contributions of the network’s branching structure, task-weighted loss, and noise augmentation to the performance improvements?
  4. The paper specifies an approximately 1% systematic perturbation, but the full parameter settings and calibration procedures for the Gaussian and Poisson noise remain unclear. How do these settings correspond to noise in field well-logging measurements? Could the authors provide a control without noise augmentation and ablations that remove each noise component individually to clarify their effects on synthetic test performance and predictions on field data?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.