Research map / Label-efficient and robust learning / Recognition and long-tail

Understanding Generalization through Decision Pattern Shift

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Decision Pattern Shift (DPS), a GradCAM-based channel-contribution vector and a cosine deviation metric from class-mean training patterns, showing it correlates linearly with the generalization gap and organizes degradation scenarios into a spectrum.

Problem

Explaining why DNNs fail to generalize to unseen samples, given that existing analyses focus on external factors (data, representations, outputs) rather than the internal decision mechanism evolving from training to test.

Approach

Represent each sample's decision pattern as a GradCAM-derived channel-contribution vector and define generalization failure as the cosine deviation of a test sample's pattern from the training class-mean pattern; the DPS magnitude is argued (theoretically via a local expansion of cross-entropy and empirically) to be linearly related to the generalization gap.

Key results

• Decision patterns achieve intra-class consistency of 0.973 on CIFAR-10 and 0.925 on CIFAR-100 (ResNet), versus 0.894 and 0.810 for activation patterns.
• Sample-level DPS correlates with the generalization gap with Pearson r >= 0.78 (p<0.001) across datasets and architectures.
• On CIFAR-100 with VGG16, sample-level correlation drops from 0.930 (0% label noise) to 0.855 (20% label noise).

Surprising findings

• OOD data (STL-10) does not necessarily induce decision-pattern shift; its spectrum stays close to the in-distribution case.
• Activation pattern shift shows benign drift for about 40% of correct predictions, which DPS suppresses.

Limitations

• DPS reference relies on class-average patterns from clean training labels; label noise weakens the correlation.
• Evaluated only on CNN architectures (VGG, ResNet, GoogLeNet); Vision Transformers not studied.
• Decision pattern is a first-order approximation; the logit-sum equivalence is not exact for nonlinear heads.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 TinyImageNet ImageNet (50-class subset) CIFAR-10-C STL-10 Colored MNIST

Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet ImageNet CIFAR-10-C STL-10 Colored MNIST

Metrics: Pearson correlation intra-class consistency (cosine similarity) inter-class confusability (cosine similarity) generalization gap

Baselines: Activation pattern (APS)

Builds on: GradCAM

Closest papers by meaning

• Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings (CVPR 2026 · similarity 0.88)
• Towards Understanding Generalization via Decomposing Excess Risk Dynamics (ICLR 2022 · similarity 0.87)
• Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking (NeurIPS 2025 · similarity 0.86)
• Deep Networks Always Grok and Here is Why (ICML 2024 · similarity 0.86)
• Topological Generalization Bounds for Discrete-Time Stochastic Optimization Algorithms (NeurIPS 2024 · similarity 0.86)
• To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.86)