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..
PLEIADES: Building Temporal Kernels with Orthogonal Polynomials
Authors: Yan Ru Pei, Olivier Coenen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimented with three event-based benchmarks and obtained state-of-the-art results on all three by large margins with significantly smaller memory and compute costs. We conduct experiments on standard computer vision tasks with event-based datasets. For all baseline experiments, we preprocess the event data into 4d tensors of shape (2, H, W, T), with the 2 polarity channels retained. General details of data and training pipelines are given in Appendix B. With the exception of the Prophesee GEN4 experiments (Section 5.3), we run 25 trials for each experiment and report the mean and standard error (which assumes a normal distribution of noise). |
| Researcher Affiliation | Industry | Yan Ru Pei NVIDIA Corporation Santa Clara, CA 95051 EMAIL. Olivier Coenen Independent Researcher Encinitas, CA 92024 EMAIL. |
| Pseudocode | No | The paper describes methods and algorithms using mathematical formulations and textual explanations, but it does not include any clearly labeled pseudocode blocks or algorithm listings. |
| Open Source Code | Yes | The code for building the structured temporal kernels, along with a pre-trained PLEIADES network for evaluation on the DVS128 dataset is available here: https://github.com/PeaBrane/Pleiades. Our code will also be made available upon as supplementary material, and will be open-sourced upon potential acceptance of the paper. |
| Open Datasets | Yes | The DVS128 dataset (CC BY 4.0) contains recordings of 10 hand gesture classes performed by different subjects [Amir et al., 2017]. The 10-pixel, 5-pixel, and 3-pixel tolerances for the CVPR 2024 AIS eye tracking challenge [Wang et al., 2024]. The Prophesee GEN4 Dataset is a road-scene object detection dataset collected with a megapixel event camera [Perot et al., 2020]. All the datasets used in this study are open sourced. |
| Dataset Splits | Yes | Following the standard benchmarking procedure on this dataset, we only train and evaluate on the first 1.5 seconds of each trial. The splits we use are also the standard splits for the corresponding datasets. |
| Hardware Specification | Yes | All training jobs are done on a single NVIDIA A30 GPU. |
| Software Dependencies | No | The paper mentions using 'Adam W optimizer', 'PyTorch default keywords', 'cosine decay learning rate scheduler', and 'automatic mixed precision (float 16)', but it does not specify exact version numbers for PyTorch or any other software libraries used. |
| Experiment Setup | Yes | For all training runs, we use the Adam W optimizer with a learning rate of 0.001 and weight decay of 0.001 (with PyTorch default keywords), along with the cosine decay learning rate scheduler (updated every step) with a warmup period of around 0.01 of the total training steps. The runs are performed with automatic mixed precision (float 16) with the model torch.compile d. For the total training epochs and walltimes (on a single A30 GPU): DVS128: 100 epochs and around 32 minutes, using a batch size of 64. 3ET+: 100 epochs and around 9 minutes, using a batch size of 64. Prophesee GEN4: 25 epochs and around 8 hours, using a batch size of 4. |