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..
Generalized and Invariant Single-Neuron In-Vivo Activity Representation Learning
Authors: Wei Wu, Yuxing Lu, Zhengrui Guo, Chi Zhang, Can Liao, Yifan Bu, Fangxu Zhou, Jinzhuo Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the following sections, we will present a series of experiments, such as cell-type prediction across different stimuli or across different animals. For each experiment, we will first detail the specific setup and protocol, and then immediately present the results, providing a clear and selfcontained evaluation of model generalization. 4 Benchmark Experiment Setup 4.1 Cell Type Prediction Across Visual Stimulus 4.2 Cell Type Prediction Across Animals 4.3 Anatomical Brain Region Prediction Across Animals Table 1: Cell Type Prediction Across Visual Stimulus (V1-Cell Type). Each cell shows the top-1 accuracy. Table 2: Cell Type Prediction Across Animals (V1-Cell Type). Each cell shows the top-1 accuracy. Table 3: Anatomical Brain Region Prediction Across Animals (IBL Brain-wide Map). Each cell shows the top-1 accuracy. |
| Researcher Affiliation | Academia | Wei Wu1, Yuxing Lu1, Zhengrui Guo2, Chi Zhang1, Can Liao3, Yifan Bu1, Fangxu Zhou1 , Jinzhuo Wang1 1 Peking University, Beijing, China 2 Hong Kong University of Science and Technology, Hong Kong, China 3 University of Georgia, Athens, USA Corresponding author: EMAIL, EMAIL |
| Pseudocode | No | The paper describes algorithms (LOLCAT, Neu PRINT, VAE) using mathematical formulations and textual explanations in sections 'B LOLCAT Framework', 'C Neu PRINT Framework', and 'D VAE Framework', but it does not provide explicit pseudocode blocks with structured, step-by-step instructions labeled as 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | The original code can be accessed via the following links: Neu PRINT at https:// github.com/lumimim/Neu PRINT/, LOLCAT at https://github.com/nerdslab/lolcat, and NEMO at https://github.com/Haansololfp/NEMO_ICLR. |
| Open Datasets | Yes | The data can be found at the links: IBL at https://www.internationalbrainlab.com/brainwide-map and V1-Cell Type at https://figshare.com/articles/dataset/A_transcriptomic_axis_ predicts_state_modulation_of_cortical_interneurons/19448531. |
| Dataset Splits | Yes | Specifically, we train and validate both the representation model and the cell type classifier using neurons recorded under three of the four stimulus conditions, and then evaluate the model on neurons recorded under the remaining, held-out stimulus condition. We train and validate both the representation model and the cell type classifier using neurons from three mice, and then evaluate the model on neurons from the remaining mouse, which is held out entirely from model optimization. Specifically, we train and validate both the representation model and the anatomical region classifier using neurons recorded from a subset (0.8) of animals, and then evaluate the model on neurons from a held-out animal that is excluded from all stages of model optimization. |
| Hardware Specification | Yes | All work can be completed on a single A100. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and ReLU activations, but does not provide specific version numbers for software libraries or frameworks (e.g., PyTorch, TensorFlow) used for implementation. |
| Experiment Setup | Yes | For all experiments, models were trained using the Adam optimizer with an initial learning rate of 1e-4. The learning rate was reduced by a factor of 0.1 if the validation loss plateaued for 5 consecutive epochs. We used a batch size of 512 and trained with an early stopping criterion based on validation accuracy to prevent overfitting. The discriminator was implemented as a two-hidden-layer MLP with 256 units per layer and Re LU activations. The adversarial weight, which balances the loss of primary tasks with the loss of adversarial, was selected among 8 values between 0.00001 and 100 on a logarithmic scale based on reverse cross-validation. |