vicreg

solo-learn: a library of self-supervised methods for visual representation learning powered by Pytorch Lightning

Toolkit for training and evaluating Self-Supervised Learning (SSL) frameworks for Speaker Verification (SV).

JEPA agent playing Minecraft from pixels: latent world model + MPC planning, 664K params on one 8GB GPU, trained on raw gameplay with no labels. A complete lab notebook - including a 20-attempt research dead end, documented with its root cause.

Latent JEPA world model (SIGReg + VICReg) with a CEM planner in latent space, trained from scratch on a parking task.

Train a JEPA world model on a set of pre-collected trajectories from an environment involving an agent in two rooms.

This a tensorflow implementation of VICReg - a self-supervised learning architecture that prevents collapse in an intuitive manner using a loss function that 1. maintains the variance of each embedding over a batch above a threshold and 2. decorrelating pairs of embeddings over a batch and attracting them to 0. Training was done using TPU on colab

A VIcReg Implementation in pytorch

Modifications to the Histomorphological Phenotype Learning pipeline.

A hands-on lab for reproducing representation collapse and preventing it with Barlow Twins and VICReg.

Self-supervised world model for 2D navigation. Dual-channel CNN encoder plus an action-conditioned autoregressive predictor, trained on 2.5M frames with VICReg to prevent representational collapse. No reconstruction objective. Linear probe recovers agent (x, y) at 1.89 MSE.

Causal, action-conditioned audio JEPA. A latent world model that forecasts future audio and steers it with control signals, plus the evaluation work needed to show it works.