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..

CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices

Authors: XUCHEN FENG, Siyu Liao

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results demonstrate that CDFlow excels in density estimation for natural image datasets and effectively models data with inherent periodicity. In terms of computational efficiency, our method speeds up the matrix inverse and log-determinant computations by 1.17 and 4.31 , respectively, compared to the general dense matrix, when the number of channels is set to 96.
Researcher Affiliation Academia Xuchen Feng School of Integrated Circuits Sun Yat-sen University EMAIL Siyu Liao School of Integrated Circuits Sun Yat-sen University EMAIL
Pseudocode No The paper describes the methods and model architecture using text, mathematical equations, and diagrams (e.g., Figure 1). It does not contain any explicit pseudocode or algorithm blocks.
Open Source Code No We plan to release the anonymized version of the code and data upon publication.
Open Datasets Yes To compare our method with previous approaches, we evaluate its performance on the CIFAR-10 [Krizhevsky, 2009] and Image Net [Deng et al., 2009] datasets, using bits per dimension (BPD) as the evaluation metric. ... The Galaxy datasets [Ackermann et al., 2018] is a small classification dataset that includes 3,000 merged galaxies and 5,000 non-merged galaxies in the training and test sets.
Dataset Splits Yes To compare our method with previous approaches, we evaluate its performance on the CIFAR-10 [Krizhevsky, 2009] and Image Net [Deng et al., 2009] datasets, using bits per dimension (BPD) as the evaluation metric. ... The Galaxy datasets [Ackermann et al., 2018] is a small classification dataset that includes 3,000 merged galaxies and 5,000 non-merged galaxies in the training and test sets. We train CDFlow on non-merged galaxy images and compare its performance to other models under the same experimental setup.
Hardware Specification Yes For a fair comparison, all methods were implemented in Py Torch, and all experiments were conducted on an NVIDIA A800 GPU.
Software Dependencies No For a fair comparison, all methods were implemented in Py Torch, and all experiments were conducted on an NVIDIA A800 GPU.
Experiment Setup Yes On standard images datasets, all of these models are configured with three blocks and 32 flow steps. On the structured dataset, we instead adopt a lighter configuration with two blocks and eight flow steps. For Woodbury, we adhere to the original configuration with ds = dc = 16, while for Butterfly Flow we use the default setting with butterfly level = 1. For our proposed CDFlow, we set m = 2, meaning we use two diagonal vectors and a circulant vector. During training, we apply spectral normalization [Miyato et al., 2018] to the convolutional layers to ensure stability. In addition, we employ a channel-aware learning rate scaling for the CD-Convolution parameters, which prevents abrupt updates to the structured matrices.