Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Optimal Neural Compressors for the Rate-Distortion-Perception Tradeoff
Authors: Eric Lei, Hamed Hassani, Shirin Saeedi Bidokhti
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, we investigate the roles that these two components of our design, lattice coding and randomness, play in the performance of neural compressors on synthetic and real-world data. We observe that performance improves with more shared randomness and better lattice packing. ... 5 Experimental Results |
| Researcher Affiliation | Collaboration | JPMorgan Chase Global Technology Applied Research, University of Pennsylvania EMAIL, EMAIL |
| Pseudocode | No | The paper includes diagrams of the architectures (Figure 1) and mathematical definitions of the processes, but it does not contain explicit pseudocode blocks or algorithms labeled as such. |
| Open Source Code | Yes | Code can be found at https://github.com/leieric/LTC-RDP. |
| Open Datasets | Yes | We use MNIST (Lecun et al., 1998), Physics and Speech datasets (Yang and Mandt, 2022). ... The data used is all open source and publicly available. |
| Dataset Splits | No | All models are trained for 100 epochs on the training data split, and reported metrics are averaged over the test split. The paper does not specify exact percentages or sample counts for these splits for the synthetic or real-world datasets. |
| Hardware Specification | Yes | Training is performed on a NVIDIA RTX5000 GPU. |
| Software Dependencies | No | For NTC models, we use the factorized py of Ballé et al. (2018), and for LTC models, we use the Real NVP normalizing flow (Dinh et al., 2017). These are frameworks/models used, but no specific software versions (e.g., Python, PyTorch versions) are provided. |
| Experiment Setup | Yes | For the synthetic (i.i.d. Gaussian source), we set n L = 8, ga and gs to be linear functions... For the Speech and Physics datasets, we use MLPs for ga and gs of depth 3, hidden dimension 100, and softplus nonlinearities. For MNIST, we follow the same exact experimental setup of Blau and Michaeli (2019), including model architecture... All models are trained for 100 epochs on the training data split... we swept a variety of λ1, λ2 values. |