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..
FAN: Fourier Analysis Networks
Authors: Yihong Dong, Ge Li, Yongding Tao, Xue Jiang, Kechi Zhang, Jia Li, Jinliang Deng, Jing Su, Jun Zhang, Jingjing Xu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we demonstrate the superiority of FAN in periodicity modeling tasks and the effectiveness and generalizability of FAN across a range of real-world tasks. To verify the effectiveness of FAN, we conduct extensive experiments from three main aspects: 1) For periodicity modeling, FAN achieves significant improvements in fitting both basic and complex periodic functions... 2) FAN shows superior performance in various real-world tasks... 3) Compared to existing Fourier-based networks, FAN accommodates both periodicity modeling and general-purpose modeling well. |
| Researcher Affiliation | Collaboration | 1School of Computer Science, Peking University 2The Hong Kong University of Science and Technology 3Byte Dance |
| Pseudocode | No | The paper describes the model architecture and equations (e.g., Eq. 9, Eq. 10, Eq. 11, Eq. 12, Eq. 13, Eq. 14) but does not include a section explicitly labeled 'Pseudocode' or 'Algorithm' with structured steps for a method or procedure. |
| Open Source Code | Yes | The code is available at https://github.com/Yihong Dong/FAN. We provide source code with comprehensive instructions and data acquisition methods at (https://anonymous.4open.science/r/FAN-D43C). |
| Open Datasets | Yes | We employ four public datasets of this task to assess the model performance on time series forecasting, including Weather [Wu et al., 2021], Exchange [Lai et al., 2018], Traffic [Wu et al., 2021], and ETTh [Zhou et al., 2021] datasets. We conduct language modeling using the SST-2 [Socher et al., 2013] dataset and evaluate the model s performance on its test set, as well as on the related datasets such as IMDB [Maas et al., 2011], Sentiment140 [Sahni et al., 2017], and Amazon Reviews [Linden et al., 2003]. Our evaluation contains four public benchmarks of image recognition: MNIST [Le Cun et al., 2010], MNIST-M [Ganin et al., 2016], Fashion-MNIST [Xiao et al., 2017], and Fashion-MNIST-C [Weiss and Tonella, 2022]. |
| Dataset Splits | Yes | Specifically, we generate data from periodic functions over a large domain, using a portion of this domain as training data and the entire domain as test data, i.e., a part of test data would be out of the domain of training data. Each dataset contains 3000 training samples and 1000 test samples, with all input variables randomly sampled from the range [-1, 1]. For each dataset, we input 96 previous time steps and forecast the subsequent time steps of {96, 192, 336, 720}. |
| Hardware Specification | Yes | We conduct our experiments on a single GPU of Tesla A100-PCIe-40G. |
| Software Dependencies | No | The paper mentions using Adam W optimizer [Loshchilov and Hutter, 2019] and GELU activation function [Hendrycks and Gimpel, 2016], and implicitly refers to 'Py Torch s optimization of MLP'. However, it does not provide specific version numbers for any key software components or libraries (e.g., PyTorch version, Python version, CUDA version) required for full reproducibility. |
| Experiment Setup | Yes | Unless otherwise specified, we use the following hyperparameters in the experiments. The model architecture consists of 3 to 24 layers, the activation function σ is set to GELU [Hendrycks and Gimpel, 2016], and the dimension of the projection matrix Wp is set to dp = 1/4dh, where dh denotes the dimension of the hidden layers. We employ the Adam W optimizer [Loshchilov and Hutter, 2019] for the model s training process. We apply a learning rate of 1e-5 for training all models. We ensured that the data density of each period in tasks was consistent, meaning that each cycle contained a fixed quantity of 10,000 training data points. |