Inference via Interpolation: Contrastive Representations Provably Enable Planning and Inference
Authors: Benjamin Eysenbach, Vivek Myers, Ruslan Salakhutdinov, Sergey Levine
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our theory using numerical simulations on tasks up to 46-dimensions.1 |
| Researcher Affiliation | Academia | Benjamin Eysenbach Princeton University eysenbach@princeton.edu Vivek Myers UC Berkeley vmyers@berkeley.edu Ruslan Salakhutdinov Carnegie Mellon University rsalakhu@cs.cmu.edu Sergey Levine UC Berkeley svlevine@eecs.berkeley.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code: https://github.com/vivekmyers/contrastive_planning |
| Open Datasets | Yes | We used two datasets from prior work [102]: door-human-v0 (39-dimensional observations) and hammer-human-v0 (46-dimensional observations). [102] Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. D4RL: Datasets for Deep Data-Driven Reinforcement Learning. ar Xiv:2004.07219, 2020. |
| Dataset Splits | No | The paper mentions using a "validation set" and performing operations on it (e.g., sampling 500 trajectories from it), but it does not specify the exact size, percentage split, or detailed methodology for creating this validation set from the overall dataset. |
| Hardware Specification | Yes | The expected compute time is a few hours on a A6000 GPU. |
| Software Dependencies | No | The paper mentions software components like 'Python', 'PyTorch', and 'CUDA' in a general context (e.g., the NeurIPS checklist), but it does not provide specific version numbers for these or other key libraries/solvers within the main text or appendices to ensure reproducibility. |
| Experiment Setup | Yes | In practice, we will impose this constraint by adding a regularization term λ Ep(x) ψ(x) 2 2 to the info NCE objective (Eq. 2) and dynamically tuning the weight λ via dual gradient descent. |