Reproduce “Variational Temporal Abstraction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Taesup Kim, Sungjin Ahn, Yoshua Bengio

Researchers from Mila, Rutgers, and Kakao Brain propose the Hierarchical Recurrent State Space Model (HRSSM), a generative model that learns stochastic hierarchical temporal abstractions from sequential data. This approach enables "jumpy imagination" for model-based reinforcement learning, allowing agents to plan more efficiently over long horizons by reasoning at a higher level of temporal granularity.

Key insights

  • Learning hierarchical temporal structure as discrete latent variables (binary indicators for subsequence boundaries) allows adaptive segmentation of sequences.
  • Combining stochastic state-space models with hierarchical transitions enables modeling both fine-grained dynamics and high-level temporal abstractions.
  • A carefully designed prior on temporal structure is crucial for the model to learn meaningful and interpretable segmentations, preventing degeneracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.