Research map / Label-efficient and robust learning / Self-supervised and contrastive
Position: An Empirically Grounded Identifiability Theory Will Accelerate Self Supervised Learning Research
Patrik Reizinger, Randall Balestriero, David Klindt, Wieland Brendel
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A position paper arguing that an empirically grounded extension of identifiability theory (Singular Identifiability Theory) is needed to explain when and why self-supervised representations converge, as posited by the Platonic Representation Hypothesis.
Problem
Current identifiability theory cannot explain key empirical SSL phenomena (representation convergence, dimensional collapse, the projector phenomenon, loss saturation, OOD generalization), leaving a gap between SSL theory and practice.
Approach
Proposes expanding identifiability theory into Singular Identifiability Theory (SITh), a framework built on empirically grounded data generating processes covering the whole SSL pipeline, and lays out research questions on augmentations, finite data/batch size, learning dynamics, architecture, projectors, OOD generalization, and evaluation.
Limitations
- Current identifiability theory relies on idealized assumptions such as infinite data and converged training
- Proposed SITh is a future framework rather than a developed theory
- Existing identifiability results (e.g., hypersphere vMF DGP) do not match realistic augmentations
Experimental setup
Builds on: Platonic Representation Hypothesis nonlinear ICA Singular Learning Theory SimCLR DIET
Closest papers by meaning
- Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.86)
- Position: The Platonic Representation Hypothesis (ICML 2024 · similarity 0.86)
- Nonparametric Identification of Latent Concepts (ICML 2025 · similarity 0.84)
- Self-supervised contrastive learning performs non-linear system identification (ICLR 2025 · similarity 0.84)
- Contrastive Identification and Generation in the Limit (NeurIPS 2026 · similarity 0.84)
- Objective drives the consistency of representational similarity across datasets (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
