Position: Many generalization measures for deep learning are fragile

In this position paper, we argue that many post-mortem generalization measures – those computed on trained networks – are \textbf{fragile}: small training modifications that barely affect the performance of the underlying deep neural network can substantially change a measure’s value, trend, or scaling behavior. For example, minor hyperparameter changes, such as learning rate adjustments or switching between SGD variants, can reverse the slope of a learning curve in widely used generalization measures such as the path norm. We also identify subtler forms of fragility. For instance, the PAC-Bayes origin measure is regarded as one of the most reliable, and is indeed less sensitive to hyperparameter tweaks than many other measures. However, it completely fails to capture differences in data complexity across learning curves. This data fragility contrasts with the function-based marginal-likelihood PAC-Bayes bound, which does capture differences in data-complexity, including scaling behavior, in learning curves, but which is not a post-mortem measure. Beyond demonstrating that many post-mortem bounds are fragile, this position paper also argues that developers of new measures should explicitly audit them for fragility.

💡 Research Summary

In this position paper the authors argue that many post‑mortem generalization measures—metrics computed on a trained deep neural network—are surprisingly fragile. “Fragile” here means that tiny changes to the training pipeline (learning‑rate tweaks, optimizer swaps, modest label‑noise adjustments) that leave the test accuracy essentially unchanged can cause a measure’s absolute value, its trend across sample sizes, or its scaling behavior to shift dramatically, sometimes even reversing sign.

The paper first motivates the problem by recalling that classical capacity measures (VC‑dimension, Rademacher complexity) have long been used to explain generalization, and that modern deep‑learning research has produced a plethora of norm‑based, margin‑based, flatness‑based, and PAC‑Bayes‑based diagnostics. While these are often called “bounds”, in practice they serve more as diagnostics because tight guarantees are rarely achieved. The central question the authors pose is whether these diagnostics inherit the empirical robustness of deep nets themselves: if a model’s performance is stable under modest hyper‑parameter changes, should a good diagnostic be similarly stable?

To answer this, the authors design a systematic “fragility audit”. The audit holds data, architecture, and most training settings fixed while perturbing a single knob at a time. Three stressors are examined: (1) learning‑curve behavior as the training set size grows, (2) temporal behavior after the model has interpolated the training data, and (3) response to changes in data complexity (label noise, dataset swaps). For each stressor they compute a “fragility score” that captures how much a measure deviates from a baseline under the perturbation.

Hyper‑parameter fragility. The authors focus first on the path norm, a widely used norm‑based proxy for ReLU networks. Using ResNet‑50 on Fashion‑MNIST, they show that switching from SGD with momentum (lr = 0.01) to Adam (same lr) collapses the path norm from ≈10⁵ to ≈10⁻¹ and makes its curve non‑monotonic. A tiny learning‑rate reduction for Adam (lr = 0.001) pushes the norm back up to 10⁵–10⁶, but now the norm plateaus instead of decreasing with more data. Test error remains almost identical across all three runs. The authors relate this to recent linear‑regression theory: depending on which ℓₚ‑minimizer the optimizer implicitly selects, the ℓᵣ norm can be monotonic, plateauing, or even U‑shaped as a function of sample size. Thus, the same “norm” does not guarantee a stable learning‑curve scaling; the optimizer’s implicit bias matters.

Temporal fragility. After a network reaches 100 % training accuracy (the interpolation point T_int), a stable diagnostic should stop moving. The authors demonstrate that this is not the case for many norm‑based measures. With SGD+momentum, the path norm settles near 5 × 10² and slowly declines after T_int ≈ 27 epochs. With Adam (same lr = 0.01), the path norm instead rises from ≈10⁻² to ≈10⁰ after T_int ≈ 114, and the margin‑normalized version follows the same upward trend. Reducing Adam’s lr to 0.001 yields a different pattern: both curves drop sharply then increase modestly after interpolation. The authors attribute this to continued logit‑scale drift in cross‑entropy loss; Adam amplifies this drift while SGD+momentum damps it. Consequently, weight‑norm‑based measures are not monotonic indicators of optimization progress.

Data‑complexity fragility. The authors then examine how measures react to changes in data difficulty. Adding label noise or swapping datasets (e.g., CIFAR‑10 ↔ SVHN) should cause a reliable measure to reflect the increased complexity, typically by changing its scaling with sample size. They find that the PAC‑Bayes “origin” bound—often praised for robustness to hyper‑parameter changes—fails to capture these differences; its curve is essentially flat across noise levels. In contrast, a function‑space marginal‑likelihood PAC‑Bayes bound (ML‑PAC‑Bayes) does track data complexity and exhibits appropriate scaling, but it is not a post‑mortem measure; it requires a prior over functions and does not depend on the training dynamics.

To formalize these observations, the authors introduce a quantitative fragility score that aggregates deviations across the three stressors. Across a broad suite of measures (path norm, Frobenius norm, spectral norm, margin, VC‑style proxies, PAC‑Bayes variants), most receive high fragility scores, indicating pronounced sensitivity. The only exception is the function‑space ML‑PAC‑Bayes bound, which remains stable but is fundamentally a different kind of object.

Finally, the paper provides a theoretical contribution: for scale‑invariant networks, a fixed learning‑rate with fixed weight decay is non‑asymptotically equivalent to a schedule with exponentially increasing learning rate and time‑varying weight decay. This equivalence explains why magnitude‑sensitive measures can inflate by orders of magnitude while test error stays flat, further highlighting the need to account for such invariances when designing diagnostics.

Conclusions and call to action. The authors argue that developers of new generalization measures must explicitly audit for fragility using the proposed protocol. Without such testing, a measure may appear to correlate with test error under a narrow set of conditions but fail dramatically under benign, realistic variations. They also suggest that future work should explore function‑space diagnostics, which appear more robust to hyper‑parameter and data changes, and investigate the underlying causes of fragility (implicit bias, scale invariance, optimizer dynamics). In sum, the paper provides both empirical evidence and methodological tools to reassess the reliability of post‑mortem generalization measures in deep learning.