Deep classifiers overfit by continuing to push every training example toward zero loss long after the useful signal in it has been learned. For ambiguous, atypical, or mislabeled examples, that final push is mostly memorization. Decision science has long studied a related problem: how an agent should decide when to stop gathering evidence. Its standard answer is the drift-diffusion model, in which noisy evidence accumulates until it crosses a threshold. When deliberation is costly, the best threshold is not fixed. It collapses over time, so the agent eventually commits with modest confidence instead of deliberating forever on a stimulus that will never become clear.
We propose to treat each training example as a decision trial. At every visit, the network sees several cheap augmented views of the example. It forms a reliability-normalized evidence score, which is the mean margin for the labeled class divided by how much that margin varies across views. The loss pushes this score up to a bound and no further. The bound for each example collapses with the number of epochs that example has spent without reaching it. Examples that quickly give consistent evidence are released early. Persistently ambiguous examples are allowed to settle at a low margin rather than being memorized. Because the score is normalized by variability across views, an example that the network fits confidently on its clean view but inconsistently under augmentation does not count as learned. That is the signature of memorization, and this is the property a plain confidence threshold lacks.
Core Claim
A per-example collapsing bound on view-consistent evidence closes more of the generalization gap than global regularization schedules or single-view confidence thresholds at matched compute. The gap here means training accuracy minus test accuracy. We predict the method will also give better-calibrated predictions. Two ingredients carry the claim, and both are testable by ablation. The first is variance normalization, taken from the reliability-weighting result described below. The second is that the bound collapses on each example's own elapsed "deliberation time" rather than on the global epoch counter. If either ablation matches the full method, the claim fails in that respect.
Mechanism
For training example $i$ with label $y_i$, draw $K$ augmented views. For view $k$, the margin $e_{ik}$ is the logit of $y_i$ minus the largest other logit. Let $\bar e_i$ and $s_i$ be the mean and standard deviation of these margins across views. The reliability-normalized evidence is
$$x_i = \frac{\bar e_i}{s_i/\sqrt{K} + \epsilon},$$
which is a t-statistic: accumulated evidence weighted by its inverse noise, with a small constant $\epsilon$ for stability. Let $n_i$ be the number of epochs example $i$ has spent with $x_i$ below its bound, which serves as its deliberation time. The training loss is
$$\mathcal{L}_i = \max\big(0,\; a(n_i) - x_i\big), \qquad a(n) = a_{\min} + (a_0 - a_{\min})\, e^{-n/\tau},$$
where $a_0$ is the initial bound, $a_{\min}$ the floor, and $\tau$ the collapse time constant. Gradients flow through both $\bar e_i$ and $s_i$. The network can therefore satisfy the bound either by raising the margin or by making the margin consistent across views, but it gains nothing by exceeding the bound. We will also test a standard cross-entropy term with small weight to keep early optimization stable. Its effect is reported separately so it cannot carry the result.
The design borrows two specific results from decision science. The Brain Uses Reliability of Stimulus Information when Making Perceptual Decisions shows that the classical diffusion model is equivalent to sequential Bayesian updating, with evidence weighted by an online estimate of its reliability. We supply this method directly: $s_i$ is that online reliability estimate, computed from views instead of time samples. Accumulation of evidence during decision making in OCD patients fits leaky diffusion with collapsing bounds and finds that reduced drift rate, not a different bound, produces long deliberation on hard trials. This motivates our reading of large $n_i$ as low drift, meaning slowly accumulating evidence, which is exactly where a fixed bound would force memorization.
Normalizing by variability also sidesteps a confound flagged in Tuning the speed-accuracy trade-off to maximize reward rate in multisensory decision-making. In that work, the fitted "scaled bound" mixes threshold height with noise level. Our bound is defined on a noise-normalized quantity, so its height and the per-example noise are separately observable. That separation makes the analysis of why an example was released interpretable.
Relation to Prior Work
The closest machine-learning work makes either its schedules or its boundaries dynamic, but never per example and never as a stopping rule. Asymptotic Midpoint Mixup for Margin Balancing and Moderate Broadening moves feature-interpolation ratios toward class midpoints on a schedule driven by global training accuracy. It treats every example in a batch identically, and the classifier boundary only passively receives gradients. We instead give each example its own bound driven by its own evidence history. Adaptive Cascading Network for Continual Test-Time Adaptation updates the classification head together with the features at test time, so that decision boundaries track a drifting domain. It addresses distribution shift after training, whereas we address memorization during training. The two are complementary, and we do not compete on its continual-adaptation protocol.
On the decision-science side, Strategically managing learning during perceptual decision making shows that learners in rats and in a small recurrent network keep thresholds high early, deliberately slowing decisions to gain the stimulus exposure that drives learning. This motivates our high initial bound $a_0$: every example receives sustained gradient before it can be released. The paper never implements this in deep networks or relates it to overfitting. We test that missing link, and we add the late-collapse half of the policy that this paper does not need.
Several regularizers constrain the design by showing what multi-view and perturbation methods already achieve. MaxUp: Lightweight Adversarial Training With Data Augmentation Improves Neural Network Training also draws several augmented copies per example, but it trains on the worst one for every example forever. That pushes in the opposite direction from releasing settled examples. Saliency-guided mixing, namely Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity and Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification, reduces label noise from mixing at the cost of extra saliency passes. Our method changes the objective rather than the inputs, so it can be stacked on top of these methods. Co-Mixup's report that Input Mixup reaches 17.7% expected calibration error tells us calibration must be measured, not assumed.
Research Program
The first stage builds the objective and verifies its basic behavior on CIFAR-10 and CIFAR-100 with PreActResNet18. This is the setting where Co-Mixup reports numbers. We use $K\in\{2,4\}$ views from standard crop-and-flip augmentation. We log per-example deliberation times $n_i$ and their distribution over training, along with the train-test gap.
The second stage is diagnostic. If the mechanism works as claimed, long deliberation times should flag the examples a network would otherwise memorize. To test this, we inject symmetric label noise at 20% and 40% into CIFAR-100. This is our protocol, and the file reports no Lacuna results for it. We then measure whether $n_i$ separates flipped labels from clean ones, and whether the method fits fewer flipped labels by the end of training than cross-entropy does.
The third stage tests robustness without adaptation. Models trained on clean CIFAR-10 and CIFAR-100 are evaluated on CIFAR-10-C and CIFAR-100-C across the 15 corruptions. The fourth stage checks scale on ImageNet with ResNet-50, where MaxUp and PatchUp give reference points, and on CIFAR100-LT. Long-tailed data is an important stress case. Tail classes naturally have slow evidence accumulation, so a collapsing bound might under-train them, and AM-mixup is strong exactly there.
Baselines
The naive reference is plain cross-entropy training with standard augmentation, run for the same epochs. It defines the generalization gap we aim to shrink.
The strongest existing method on clean CIFAR-100 in the file is Co-Mixup. On PreActResNet18 it reports 19.87% top-1 error and 1.9% expected calibration error, against 21.29% error for CutMix and 17.7% calibration error for Input Mixup. We re-run Co-Mixup and CutMix with their published settings on our pipeline, with identical epochs and the same tuning budget for each method (an equal number of hyperparameter trials). Beating Co-Mixup means lower test error at equal or better calibration, or matching its error at a clearly smaller train-test gap. Because our method is orthogonal, we also report Co-Mixup combined with our loss. A gain on top of Co-Mixup is the more meaningful win.
MaxUp is the compute-matched multi-view competitor, since it also uses several views per example. We run it with $m=K$ views so forward-pass budgets are identical. On ImageNet, its reported ResNet-50 top-1 accuracy with CutMix is 78.9%. On CIFAR-10 it reports certified robustness rather than clean accuracy, so we compare on our own runs.
AM-mixup is the strongest dynamic-schedule competitor and the reference on CIFAR100-LT. It reports 41.19% accuracy on tail classes at imbalance factor 100, versus 8.89% for cross-entropy. We report no clean CIFAR comparison number for it because the file does not give one.
The key ablations isolate each ingredient. The first removes variance normalization and bounds the mean margin $\bar e_i$ only. The second uses a fixed bound with no collapse. The third collapses on the global epoch counter, so every example shares one schedule. The fourth uses a single view with a stop-gradient once softmax confidence exceeds a threshold. Each ablation receives the same tuning budget as the full method.
Why not just…?
Why not just stop the gradient once an example's confidence passes a threshold? This is the most serious alternative, because it captures "push no further" with no machinery. Our argument against it is that single-view confidence cannot distinguish an example the network has learned from one it has memorized, since memorized examples are often fitted confidently on their training view. The variability term $s_i$ makes that distinction. The single-view threshold ablation, especially under label noise, settles this directly. If it matches the full method on the generalization gap and on the fraction of flipped labels fitted, then the reliability-weighting idea adds nothing, and we will report that.
Why not just average the loss over several augmented views, or train on the worst view as MaxUp does? Both use the same extra forward passes we do. Neither ever releases an example, and MaxUp concentrates gradient on the hardest view of every example indefinitely, which plausibly encourages memorization of ambiguous ones. The compute-matched MaxUp baseline answers this. If MaxUp closes the gap as much as our method at equal $K$, the stopping rule is not what matters.
Why not just use a global schedule, such as annealed label smoothing or an AM-mixup-style accuracy-driven schedule? A global schedule already relaxes targets late in training. However, it relaxes them equally for clean examples that deserve high margins and for noisy ones that do not. The global-collapse ablation is the direct test: it keeps everything except per-example deliberation time. The long-tailed comparison with AM-mixup shows whether per-example collapse hurts or helps rare classes. We consider this a genuine risk, and if it materializes we will report that a class-aware floor $a_{\min}$ is needed.
Why not just use early stopping? Early stopping halts all examples at once, at a point chosen on validation data. It therefore cannot let clean examples keep improving while noisy ones settle. We report cross-entropy with validation-selected early stopping alongside the naive baseline.
Decisive Checks
The method is worth pursuing if four outcomes hold together. On CIFAR-100 with PreActResNet18, it must reduce the train-test gap relative to cross-entropy and to compute-matched MaxUp, without raising test error. Calibration error must stay within the range Co-Mixup reports. Under label noise, deliberation time must separate flipped from clean labels, and the full method must fit fewer flipped labels than the single-view threshold ablation. Finally, both the per-example-collapse ablation and the variance-normalization ablation must lose to the full method. The single most informative experiment is the label-noise comparison between the full method and the single-view confidence stop. Of the alternatives, it is the one that most directly asks whether decision-science reliability weighting does real work in deep network training.
References
- The Brain Uses Reliability of Stimulus Information when Making Perceptual Decisions
- Accumulation of evidence during decision making in OCD patients
- Tuning the speed-accuracy trade-off to maximize reward rate in multisensory decision-making
- Asymptotic Midpoint Mixup for Margin Balancing and Moderate Broadening
- Adaptive Cascading Network for Continual Test-Time Adaptation
- Strategically managing learning during perceptual decision making
- MaxUp: Lightweight Adversarial Training With Data Augmentation Improves Neural Network Training
- Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity
- Attentive Cutmix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification
- PatchUp: A Feature-Space Block-Level Regularization Technique for Convolutional Neural Networks