Source-linked AI summary
Predicting trends in the quality of state-of-the-art neural networks without access to training or testing data
Charles H. Martin, Tongsu, Peng, Michael W. Mahoney
TL;DR
The paper asks how to assess pretrained neural-network quality when training/testing data and development details are unavailable. It conducts a large meta-analysis of publicly available models using norm-based and power-law-based weight metrics, finding that power-law metrics generally distinguish model quality better, while norm metrics work mainly for well-trained models. These metrics can help identify problematic pretrained networks beyond training or testing accuracies alone.
Problem
Pretrained-model quality is difficult to assess when users lack training/testing data, development details, or reliable metrics aligned with downstream goals.
Method
The paper performs a large empirical meta-analysis of hundreds of publicly available pretrained models using norm-based and power-law-based weight-matrix metrics.
Results
Power-law-based metrics better predict quality trends and discriminate well-trained from poorly trained models, whereas norm-based metrics work reasonably well mainly for well-trained models.
Takeaways & Limitations
Weight-based metrics can reveal problems in pretrained neural networks that training and testing accuracies alone may not detect.
Takeaways & Limitations
Power-law and norm-based metrics should be used cautiously when comparing models with very different architectures.
Abstract
from arXiv · showhide
In many applications, one works with neural network models trained by someone else. For such pretrained models, one may not have access to training data or test data. Moreover, one may not know details about the model, e.g., the specifics of the training data, the loss function, the hyperparameter values, etc. Given one or many pretrained models, it is a challenge to say anything about the expected performance or quality of the models. Here, we address this challenge by providing a detailed meta-analysis of hundreds of publicly-available pretrained models. We examine norm based capacity control metrics as well as power law based metrics from the recently-developed Theory of Heavy-Tailed Self Regularization. We find that norm based metrics correlate well with reported test accuracies for well-trained models, but that they often cannot distinguish well-trained versus poorly-trained models. We also find that power law based metrics can do much better -- quantitatively better at discriminating among series of well-trained models with a given architecture; and qualitatively better at discriminating well-trained versus poorly-trained models. These methods can be used to identify when a pretrained neural network has problems that cannot be detected simply by examining training/test accuracies.
1 Introduction
Evaluating pretrained models without their training or testing data is practically important because standard accuracy-based evaluation can be inaccessible or poorly aligned with downstream goals. The paper addresses this gap through a large empirical analysis of weight-based quality metrics, finding different strengths for norm-based and power-law-based measures.
- Motivation: Model users may lack training and testing data, model-development details, or even a reliable metric aligned with downstream goals.This is especially relevant when models are developed and used by separate groups, or when objectives such as perplexity are poor proxies for human evaluation.
- Motivation: Industry needs metrics that evaluate trained-model quality without detailed knowledge of the training or testing process.The paper frames this as a need for a practical theory that can predict when pretrained models may perform well or poorly.
- Approach: The study performs a large-scale empirical meta-analysis of hundreds of publicly available pretrained models, mainly from computer vision and natural language processing.The models lack accessible training and testing data and known training/testing protocols.
- Main findings: Norm-based metrics predict quality trends reasonably well in well-trained models but can produce spurious results for poorly trained models.The paper identifies Scale Collapse as one example of misleading norm-based behavior.
- Main findings: Power-law-based metrics better predict quality trends and discriminate both models within architecture series and well-trained versus poorly trained models.The unweighted average power-law exponent is reported as qualitatively better for distinguishing training quality, while weighted power-law measures perform well for model-quality trends.
2 Results
The study evaluates norm-based and PL-based metrics on pretrained model series, finding that both track quality among well-trained models while PL metrics better distinguish complex-model trends and poorly trained models. Layer-wise PL behavior also reveals architecture-dependent patterns and diagnostic anomalies such as Scale Collapse.
- RMSE 0.66, R2 0.9, and Kendall-τ −1.0 describe Log α-Norm correlation for ResNet, compared with RMSE 1.9, R2 0.88, and Kendall-τ −0.88 for ResNet-1K.The ResNet-1K relationship is good but weaker than the ResNet relationship.
- Norm-based and PL-based metrics both distinguish well-trained from very-well-trained models within established architecture series.PL-based metrics perform quantitatively better on the larger, more complex ResNet series.
- PL-based Weighted Alpha and Log α-Norm metrics perform best on larger, more complex ResNet models, while Log Spectral norm performs very well on smaller VGG and DenseNet models.Weighted Alpha and Log α-Norm also tend to perform better when hyperparameters vary more broadly.
- Model diagnostics: In GPT versus GPT2-small, the average PL exponent distinguishes the deficient GPT from the improved GPT2-small, whereas scale-dependent norms can produce spurious quality signals.The deficient GPT has many unusually large α values and anomalously small Spectral norms.
- Layer analysis: Layer-wise α is architecture dependent: it rises systematically through VGG, remains near α ∼2.0 across much of ResNet-1K, and is more variable in DenseNet.VGG reaches α ∼5.0 in later Conv2D layers, whereas DenseNet can range to α ∼8.0.
- Model diagnostics: Scale Collapse can spuriously increase norms in distilled layers, making norm metrics misleading even when PL exponents diagnose the anomaly.In the Group Regularization example, two Conv2D pre-activation layers show spuriously increased norms, a problem not readily detected from training/test curves.
3 Discussion
The paper connects layer-wise spectral patterns to architectural differences and argues that weight-matrix metrics can assess pretrained-model quality without training or testing data. It also positions the work as an extensive empirical analysis using WeightWatcher and HT-SR Theory.
- Architecture comparisons: ResNet models show smaller, more stable α ∼2.0 across layers, while DenseNet models show larger α and weaker layer-wise regularity.The authors associate these patterns with stronger inner-layer correlations in ResNet and diluted interactions in DenseNet.
- Practical implication: The authors report that examining weight matrices can provide substantial insight into model quality without access to training or testing data.They describe their analysis as broader than subsequent studies that considered smaller models and narrower application ranges.
- Approach: The study analyzes weight matrices with WeightWatcher and applies HT-SR Theory to publicly available pretrained neural networks.The paper describes this as a detailed analysis of hundreds of models across computer vision and language applications.
- Contribution: The paper frames its contribution as a first detailed empirical demonstration that norm-based and PL-based metrics can predict quality trends across pretrained models.It reports that PL-based metrics perform better than norm-based metrics for this practical problem.
4 Methods
The methods analyze pretrained networks through their layer weight matrices, using norm-based and spectral power-law measures. Practical choices address convolutional representations, normalization, and unusually scaled NLP embedding layers.
- Weight-matrix analysis: The analysis focuses on pretrained layer weight matrices while ignoring architectural details such as regularization and skip connections.The models are understood as having been trained by minimizing a loss over labeled data, but the analysis does not use that data.
- Convolutional layers: For convolutional layers, the study considers three SVD strategies: per-map SVD, stacked-map SVD, and Fourier-domain SVD.These alternatives produce different singular-value representations and involve different computational tradeoffs.
- Convolutional layers: The analysis selects per-map SVD because it is fastest and easiest to reproduce, despite tradeoffs among the three convolutional-layer methods.The Fourier-based method is described as mathematically sound but computationally expensive, while the stacked-map method is ambiguous.
- Normalization: PL fits are scale-invariant, whereas norm-based metrics depend strongly on weight-matrix scale.The empirical matrices are therefore generally used as-is, with selected rescaling for convolutional pre-activation maps.
- NLP models: NLP embedding layers require special care because their eigenvalue scales can differ greatly from those of other layers.For GPT and GPT2, the study treats layers as-is while noting that normalizing the first layers or treating them as outliers may be considered.
A.1 Supplementary Details
The supplementary material emphasizes reproducibility through notebooks, datasets, and WeightWatcher analyses. It also documents the filtering and outlier-removal decisions used in the large-scale regression study.
- Reproducibility: The authors provide notebooks and a repository to reproduce results from Sections 2.2 and 2.3 using WeightWatcher.The materials include additional analyses and pretrained-model resources from PyTorch, Hugging Face, and related repositories.
- Regression design: The regression analysis covers over 450 PyTorch models and excludes datasets with fewer than 15 models and regressions with fewer than 5 datapoints.The omitted datasets are CUB-200-2011, Pascal-VOC2012, ADE20K, and COCO.
- Data filtering: The study removes two outlier models and the entire CIFAR-100 ResNeXT series because it showed no trends with norm metrics.The final architecture series and their model counts are listed in Table 5.
- Supplementary analysis: The supplementary section presents increasingly fine-grained tables and figures summarizing the large-scale analysis.These materials extend the summary results rather than introducing a separate experimental setting.
A.2 Supplementary Tables and Figures
The supplementary figures evaluate how weight-based metrics relate to reported Top1 accuracy across pretrained models. They report regression and rank-correlation statistics while illustrating that fit quality varies substantially across analyses.
- Analysis scope: The supplementary analysis spans numerous pretrained models from OSMR/imgclsmob and applies OLS regressions across datasets and architecture series.The figures provide a more detailed view of results summarized in Table 3.
- Figure design: Each figure row compares average norm-based and power-law metrics with reported Top1 test accuracy for a pretrained model and dataset.Shaded regions represent 95% confidence bounds, and the figures report RMSE, R2, and Kendall-τ statistics.
- Illustrative result: For ImageNet-1K PreResNet, Log Spectral norm has RMSE = 3.93, R2 = 0.36, and Kendall-τ = 0.54, with 6 of 13 points outside the 95% confidence bands.The passage contrasts this weaker fit with the smaller RMSE reported for Log Frobenius norm.
- Fit variability: The figures and tables contain fits ranging from remarkably good to completely uncorrelated.The authors direct readers to the reproducibility repository for additional conclusions and regenerated results.
A.3 Supplementary Discussion: Additional Details on HT-SR Theory
Across nearly 10,000 layer weight matrices from hundreds of pretrained architectures, truncated power laws provide a widespread description of eigenvalue spectra. Most fitted exponents fall between 2 and 4, although exceptions extend from approximately 1.5 to 10 or higher.
- Empirical prevalence: Nearly every examined layer weight matrix can be fit to a truncated power law.The analysis reports this pattern across hundreds of different modern pretrained DNN architectures.
- Scope and exceptions: Fitted α values may range from approximately 1.5 to 10 or higher, and large α values may indicate that the power law is not a good model.Thus, the observed heavy-tailed behavior has exceptions despite its broad prevalence.