The Predictability Bias of Joint-Embedding Predictive Learning in Medical Imaging
Joint-embedding predictive architectures (JEPAs) learn by predicting teacher representations of masked regions from visible context.
ProofPaper ↗
Key points
- 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.
- Replacing visible matched-control representations increases surrounding-target prediction loss more than fully replacing pathology representations across four pathology categories.
Sources (1)
- [1]The Predictability Bias of Joint-Embedding Predictive Learning in Medical ImagingarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 08:53 AM
Joint-embedding predictive architectures (JEPAs) learn by predicting teacher representations of masked regions from visible context.
We call this constraint predictability bias.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 9, 2026RSI-Anything: Unlocking Recursive Self-Improvement with World Models
- Oct 9, 2026Learning Representations to Guide Pixel Diffusion Transformers
- Oct 9, 2026Rethinking Relative Policy Differences Transfer of On-Policy Distillation
- Oct 9, 2026InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators
- Oct 9, 2026MixTok: Adaptive Visual Tokenization with Global 1D Tokens and Sparse 2D Refinement
- Oct 8, 2026MCL: Meta Convolution Layer