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..
CRRL: Learning Channel-invariant Neural Representations for High-performance Cross-day Decoding
Authors: Xianhan Tan, Binli Luo, Yu Qi, Yueming Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed method on both simulation datasets with diverse channel-wise variabilities and multiple neural signal datasets with motor, handwriting and speech decoding tasks. With the neural decoding tasks, we evaluate the neural decoding performance of our approach with long-term neural signals with time spans over two months. Experimental results demonstrate that our approach achieves stable decoding performance compared with existing studies, especially for a long period. |
| Researcher Affiliation | Academia | 1MOE Frontier Science Center for Brain Science and Brain-machine Integration, Zhejiang University 2Affiliated Mental Health Center & Hangzhou Seventh People s Hospital, Zhejiang University 3College of Computer Science and Technology, Zhejiang University 4College of Computer Science and Technology, Central South University 5State Key Lab of Brain-Machine Intelligence, Zhejiang University Corresponding author (EMAIL) |
| Pseudocode | No | Figure 2 illustrates the conceptual CRRL training and evaluation processes. Algorithms 1 and 2 specify the two stages of CRRL training in detail. In the rearrangement training, the first step is to predict the permutation matrices. Then, CRRL rearranges the spike channels and obtains the rearranged spike. In the reconstruction training, we predict the masked signal and its frequency feature. During optimization, our approach learns to robustly restore the information of the day 0 signal. For evaluation, we can obtain the reconstructed data by well-trained CRRL. The obtained data can decode downstream tasks by a decoder trained on day 0. |
| Open Source Code | Yes | Yes, we upload the code in the supplementary file and list the link of open data source in the appendix. |
| Open Datasets | Yes | Yes, we upload the code in the supplementary file and list the link of open data source in the appendix. [...] 2https://datadryad.org/stash/dataset/doi:10.5061/dryad.cvdncjt7n 3https://datadryad.org/stash/dataset/doi:10.5061/dryad.wh70rxwmv 4https://datadryad.org/stash/dataset/doi:10.5061/dryad.x69p8czpq |
| Dataset Splits | Yes | In the cross-day setting, we use data from the first 5 consecutive days for training (held-in data). All data from these 5 days are merged, then randomly split into 90% training and 10% validation. Data from Day 6 and onward is treated as held-out test data. The model is evaluated on these later days without any additional training or adaptation, to assess generalization under temporal distribution shift. And the data split ratio used for training the day 0 decoder of all datasets is 8:1:1. |
| Hardware Specification | Yes | All experiments were conducted on NVIDIA Tesla A100 GPUs, and the maximum memory usage across all experiments was approximately 22 GB. |
| Software Dependencies | No | Our model is implemented using Py Torch, enabling flexible model design and training, and training and inference are performed on GPUs to accelerate computation. |
| Experiment Setup | Yes | We adopt the Adam optimizer with a learning rate of 0.001 for model training and utilize a batch size of 128/256 samples. The hyperparameters, such as the weights of different loss λfreq and λvq are chosen based on validation performance. During masked channel modeling, a masking ratio of 0.05 determines the proportion of channels to mask in each input, and a random channel shuffling rate of 0.1 is applied during the training of the permutation network to enhance robustness. |