AI Science Tracker
EN

Masked patch training changes visual correspondence differently at different model sizes

National University of Singapore; Nanyang Technological University · 23 Oct 2026 · computer science

Discovery

Masked patch training changes visual correspondence differently at different model sizes

Source check: Source checked. Not independent scientific confirmation.

Finding: Under matched ImageNet-1K training, adding iBOT with its associated masking weakened patch correspondence in small and base Vision Transformers. Final-block patch matching improved by roughly 3.3 and 4.0 percentage points in the large and huge variants, while some local appearance details became harder to recover from deep features. These effects were measured with one training recipe.

Technology: DINO-style ViT-S/B/L/H; iBOT; fitted linear appearance probes

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Vision Transformers trained with and without iBOT supplied patch embeddings for crop matching and linear appearance decoding.
Test and result
Within each scale, the comparison toggled the iBOT branch and student masking. The position diagnostic used paired local and global crops. Separate linear probes assessed appearance.
Limitation / still unknown
Data volume, loss weight, training duration and specific diagnostics limit transfer to full-scale commercial checkpoints.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.33223v1 · https://arxiv.org/html/2610.33223v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: v1 training §3, patch and appearance diagnostics §4, Limitations.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.