JEPA · 2026-10-12
The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging
Fabian Drexel, Bernhard Kainz, Jakob N. Kather, Daniel Rueckert, Lisa C. Adams, Keno K. Bressem, Cosmin I. Bercea
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors identify a “predictability bias” in joint-embedding predictive learning: point prediction preserves only what hidden targets are predictable from visible context, not clinically relevant variation that the context cannot resolve. In TAP-JEPA (ViT-B/3D on 113,611 CT volumes), prediction reduces pathology discrimination much more than anatomy and shifts pathology representations toward matched controls.
Why it matters
If predictive objectives preferentially retain context-resolvable information, they may underrepresent localized, sparse findings that drive clinical interpretation. The authors’ audits (mask/teacher/predictor replay and representation interventions) show that contextual predictability is not a sufficient proxy for clinical relevance, motivating training objectives that keep diagnostic features regardless of predictability from surrounding anatomy.
Method
- Introduce predictability bias for JEPA point-prediction learning: optimal predictions summarize conditional target distributions given visible context, so unresolved variation contributes no conditional mean training signal via the visible path.
- Train TAP-JEPA (ViT-B/3D) on 113,611 thoraco-abdomino-pelvic CT volumes; then replay the exact masking, teacher, predictor, and loss computations at a fixed checkpoint to audit prediction effects on pathology vs anatomy.
- Quantify downstream and intervention effects by deleting pathology tokens and by replacing visible pathology/control representations with matched control/pathology surrogates to trace changes in prediction and pathology–control separation.
Abstract (from arXiv)
Joint-embedding predictive architectures (JEPAs) learn by predicting teacher representations of masked regions from visible context. Point prediction summarizes the conditional target distribution and cannot reproduce variation that the context does not resolve. We call this constraint predictability bias. In computed tomography (CT), recurrent anatomy is predictable from context, whereas the sporadic, localized pathology that determines clinical interpretation often is not. We introduce TAP-JEPA, a ViT-B/3D trained on 113,611 CT volumes that is competitive with state-of-the-art encoders (COLIPRI, Merlin, VoxelFM) in scan-level classification. We then audit the learned predictive pathway by replaying the exact masking, teacher, and predictor computations used during training. Prediction lowers the mean Skill (prevalence-normalized average precision) from 0.388 to 0.271 for pathology, but only from 0.921 to 0.873 for anatomy. Along the teacher direction, separating pathology from matched control regions in the same organ, prediction moves pathology 34.7% toward the controls and controls 13.6% toward pathology, with larger shifts farther from visible context. Replacing visible matched-control representations increases surrounding-target prediction loss more than fully replacing pathology representations across four pathology categories. Downstream, removing all final-layer pathology tokens reduces the decision margin by at most 0.45% in four of seven scan-level probes. Together, these results show that contextual predictability is not a sufficient proxy for clinical relevance in medical representation learning. Training objectives should therefore retain diagnostic features regardless of their predictability from surrounding anatomy.
Related papers
- CI-JEPA: A Counterfactual Analysis of Latent Representations in Joint-Embedding Predictive Architectures for Self-Supervised Learning
- Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures
- Are Joint-Embedding Predictive Objectives Effective for Longitudinal EHRs?
- JEPA Learns What the Mask Leaves Unrecoverable
- Intervention anchors and scientific verification in synthetic vascular predictive representations
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-09 · Indexed by Pub-AI: 2026-10-12 UTC
The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging
Preprint posted: 2026-10-09 · Original record ↗
- arxiv
- 2610.13237
Explicit source versions (1)
- v1 · 2026-10-09
- Preprint available
Source date: 2026-10-09 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Joint-embedding predictive architectures (JEPAs) learn by predicting teacher representations of masked regions from visible context. Point prediction summarizes the conditional target distribution and cannot reproduce variation that the context does not resolve. We call this constraint predictability bias. In computed tomography (CT), recurrent anatomy is predictable from context, whereas the sporadic, localized pathology that determines clinical interpretation often is not. We introduce TAP-JEPA, a ViT-B/3D trained on 113,611 CT volumes that is competitive with state-of-the-art encoders (COLIPRI, Merlin, VoxelFM) in scan-level classification. We then audit the learned predictive pathway by replaying the exact masking, teacher, and predictor computations used during training. Prediction lowers the mean Skill (prevalence-normalized average precision) from 0.388 to 0.271 for pathology, but only from 0.921 to 0.873 for anatomy. Along the teacher direction, separating pathology from matched control regions in the same organ, prediction moves pathology 34.7% toward the controls and controls 13.6% toward pathology, with larger shifts farther from visible context. Replacing visible matched-control representations increases surrounding-target prediction loss more than fully replacing pathology representations across four pathology categories. Downstream, removing all final-layer pathology tokens reduces the decision margin by at most 0.45% in four of seven scan-level probes. Together, these results show that contextual predictability is not a sufficient proxy for clinical relevance in medical representation learning. Training objectives should therefore retain diagnostic features regardless of their predictability from surrounding anatomy.