New paper proves predictive SSL provably identifies stochastic signals under nuisance
hisspikeness · x · 2026-10-08
Fabian Mikulasch and Friedemann Zenke released an arXiv paper showing that predictive self-supervised learning (predicting in latent space without reconstructing inputs) can provably separate prediction-relevant stochastic signals from nuisance variation.
- Key puzzle: both stochastic variation of a relevant latent signal and true nuisance make observations partially unpredictable — how can SSL tell them apart?
- Theory: common SSL methods implicitly instantiate a latent-variable model with stochastic dynamics and observation-private nuisance; predictive mutual information maximization retains prediction-relevant information, while latent distribution matching constrains encoding, making the retained signal identifiable (recoverable up to an affine transformation for Gaussian predictors).
- Experiments: on a MuJoCo hopper with random colors, lighting, camera and noisy backgrounds, latents allow linear decoding of pose and velocity; adding a pixel-reconstruction loss makes latents encode nuisance and breaks decodability.
- Ongoing work: learned latent dynamics rollout from a few observations recovers the hopper's signals with uncertainty estimates. Code is open-sourced.
More from Research
- Talk: how to RL-train an agent running inside a harness you didn't write — SergioPaniego · 2026-10-08
- AI labs accused of press-release deception: none of 520 claimed proofs Lean-formalized — gerardsans · 2026-10-08
- Commercial detector re-runs NeurIPS AI-text check: 7.3% of 2025 papers flagged vs Pangram's 1% — AltruisticCouple3491 · 2026-10-08
- Burned 300B Tokens with Nothing; Internal Model Broke Through in 3 Hours — burny_tech · 2026-10-08
- NVIDIA's UNREAL paper lets one LLM both retrieve and answer, lifting recall from 49% to 73% — mark_k · 2026-10-08
- NAMVIS: next-scale autoregression beats diffusion for multi-view synthesis, 3x faster — Ramil Khafizov · 2026-10-08