Midway Network is a new self-supervised learning architecture that is the first to learn visual representations for both object recognition and motion prediction solely from natural videos by leveraging latent dynamics.

Our method Canonical pretrains visual representations by modeling camera viewpoint change as an SO(3) rotation in latent space, allowing visuomotor policies to generalize to unseen camera angles without calibration.

LaMo is a structured latent dynamics model that enables long-horizon prediction by autoregressively predicting latent motion tokens used to advance a visual scene's latent state over time.

We introduce Discrete-JEPA, which learns discrete semantic tokens for improved symbolic reasoning and long-horizon planning.

We propose a self-supervised learning framework that combines pooled and dense objectives to learn representations with spatial and semantic understanding from naturalistic videos.

We leverage successor features to formulate a graph-based planning framework and goal-conditioned policy, enabling long-horizon goal-reaching in visual environments.

We use an agent-based model and empirical game-theoretic analysis to study price manipulation in financial markets and propose mitigation mechanisms.

We design learning-based trading strategies that have improved robustness to market manipulation and evaluate them with agent-based simulation.