Skip to content

ssl-lab documentation

A research lab for self-supervised learning (SSL) — tracking the SSL frontier and going deep on the methods most worth mastering. The current focus is JEPA (joint-embedding predictive architectures) and two ways to extend it.

This site is built from the docs/ folder of pleiadian53/ssl-lab and rendered with full LaTeX math support.

Research directions

1. Generative JEPA — make the representation sampleable. Add a prior over the latent and a decoder back to data space, turning a representation learner into a generative model. Built as a walking skeleton on MNIST (modality-agnostic core). → Full tutorial: Generative JEPA.

flowchart LR
    NOISE(["noise ε"]) --> PRIOR["flow-matching<br/>prior p(z)"]
    PRIOR -- "sample z ~ p(z)" --> DEC["decoder<br/>z → x"]
    DEC --> GEN(["generated sample"])
    ENC["JEPA encoder<br/>(frozen)"] -. "defines the latent z<br/>the prior is fit to" .-> PRIOR

    classDef accent fill:#eef2ff,stroke:#6366f1,color:#1e1b4b;
    classDef io fill:#f8fafc,stroke:#94a3b8,color:#0f172a;
    class ENC,PRIOR,DEC accent;
    class NOISE,GEN io;

2. Action operators on JEPA — make prediction active. Promote JEPA's fixed "predict region \(q\)" mask to a learned operator the model chooses — sensing (where to look) and perturbing (what an action does) — so the system can form and test hypotheses rather than only in-fill what's masked. The gentlest way in is the Time-Series JEPA series (JEPA pointed at time series — the natural first arena for action operators); then the Action Operators foundation and the Operator World Models synthesis. Builds on the action-operator formalism from the sibling project GRL.

  • Generative JEPA — a four-part tutorial on extending a JEPA encoder into a sampleable generative model: the encoder, the flow-matching prior, the decoder, and sampling + evaluation.
  • Time-Series JEPA — the accessible entry point for the operator track: JEPA pointed at time series, multimodal channels, and the one blind spot that motivates action operators.
  • Action Operators — the foundation: From actions to operators → Augmenting JEPA with Action Operators → A gallery of operators.
  • Operator World Models — the JEPA + operator world-model series, building on that foundation (state and latent operators, temporal prediction, conditioning on interventions).

About

ssl-lab is a spin-off of genai-lab (generative AI for computational biology). Methods are kept deliberately modality-agnostic so they can mature here and feed back into genomic generative models. See the repository README for setup and the code layout.