Self-supervised learning (SSL) extracts representations from massive unlabeled data that transfer to a variety of downstream tasks. Its empirical success has produced a dizzying array of objectives and procedures, yet a clear understanding of what representation is desired, and why these seemingly different methods work, is still missing. In this paper, we develop a unified spectral framework for representation learning. We first show that a representation is sufficient for arbitrary downstream prediction with a linear head if it spans the spectral decomposition of the pointwise mutual information, which can be extracted purely from unlabeled data. This view organizes representation learning along two axes, the parametrization of the spectral representation (linear, energy-based, denoising, nonlinear, latent-variable, and multi-modal) and the learning objective for fitting it, under which a broad range of existing methods, including SimCLR, MoCo, BYOL, Barlow Twins, VICReg, DINO, MAE, JEPA, BERT, GPT, and CLIP, as well as classic PCA, CCA, Laplacian eigenmaps, LLE, and SNE, are recovered as particular combinations. The framework also clarifies optimization behaviors: the batch-size requirement is determined by whether the objective is unbiased, a consistent surrogate, or biased, rather than by contrastive vs. non-contrastive losses. Finally, the framework inspires new scalable algorithms and extends spectral representations to causal inference and reinforcement learning.

翻译：暂无翻译