Representation Trajectories Matters: Complementary Evidence for OOD Detection and Image Classification

2026-07-29

AI Summary

This work addresses the issue that existing visual models may discard intermediate information useful for out-of-distribution (OOD) detection and classification during their layer-wise representation construction. It pioneers modeling deep representation evolution as a continuous trajectory rather than isolated snapshots, proposing to distinguish semantically consistent transformations from input-specific innovations—and coordinate shifts from relational reorganizations—by leveraging trajectory continuity, transition surprisal scoring, and frozen-update probing. This approach extracts discriminative dynamic signals overlooked by final-layer representations. The method generalizes across diverse encoder architectures, including supervised, self-supervised, vision-language, and convolutional models, achieving significant FPR95 reductions in 131 out of 152 non-saturated comparisons on the OpenOOD benchmark, with particularly strong gains under severe perturbations or large semantic shifts, while also improving classification accuracy on 71 out of 72 clean datasets.

Abstract

Vision models do not form a representation at once; each block revises it. We ask whether the resulting computation path contains evidence that the final representation discards, and whether that evidence improves OOD detection and image classification on clean and shifted data. Unlike approaches that treat intermediate layers as separate snapshots, we retain sample identity across depth and study the transformations connecting successive states. We separate class-coherent transport from input-specific innovation, and coordinate movement from relational reorganization. Across supervised, self-supervised, vision--language, hierarchical, and convolutional encoders, these paths show strong sample-specific continuity and architecture-specific depth profiles that recur across datasets. They are also practically useful. An ID-only transition-surprise score complements strong final-state detectors, reducing FPR95 in 131/152 non-saturated comparisons on a balanced OpenOOD grid; gains are largest for visually disruptive and semantically far shifts, and remain positive on near-OOD for most detectors. Frozen update probes improve 71/72 clean model--dataset cases, while shifted-data gains vary with architecture and corruption type. Computation paths therefore provide a broadly useful reliability signal whose value is determined jointly by model organization and the shift encountered.

Problem

• representation trajectories
• OOD detection
• image classification
• distribution shift
• intermediate representations

Innovation

• representation trajectories
• OOD detection
• transition-surprise
• frozen update probes
• depth-wise dynamics