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..
MEIcoder: Decoding Visual Stimuli from Neural Activity by Leveraging Most Exciting Inputs
Authors: Jan Sobotka, Luca Baroni, Ján Antolík
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments We compare MEIcoder to state-of-the-art baselines on three datasets, two of which represent dataand neuron-constrained settings. |
| Researcher Affiliation | Academia | Jan Sobotka1, Luca Baroni2 Ján Antolík2 1EPFL, Switzerland 2Charles University, Czechia |
| Pseudocode | No | The paper describes the architecture and training objective in Section 3 and Figure 1 using natural language and mathematical equations, but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code for our experiments and instructions on obtaining the data are available in our public repository at https://github.com/Johnny1188/meicoder. |
| Open Datasets | Yes | Brainreader dataset. The BRAINREADER4 data comes from mouse V1 and was originally introduced by [11]. SENSORIUM 2022 dataset. We repurpose the mouse dataset published by the SENSORIUM 2022 competition [48] for our decoding task. Synthetic cat V1 dataset. Lastly, since most of the currently publicly available datasets suitable for decoding in V1 are greatly data-constrained and limited to mouse or monkey data, we leverage a highly biologically realistic spiking model of cat V1 from [3] to generate a large synthetic dataset. |
| Dataset Splits | Yes | We divide the dataset into training, validation, and test sets of 4,500, 500, and 100 samples, respectively. (BRAINREADER) ... We split this additional synthetically generated dataset into training (45,000), validation (5,000), and test (250) sets. (SYNTHETIC CAT V1) ... We split the dataset from each mouse into a training (4,500) and a validation (500) set. (SENSORIUM 2022) |
| Hardware Specification | Yes | Each of our experiments used one NVIDIA Tesla V100 GPU and required less than 32 GB of VRAM. |
| Software Dependencies | No | The paper mentions using the Adam W optimizer [26] and references an implementation from Mind Eye2 [36] for evaluation metrics, but it does not specify explicit version numbers for other key software components such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | We train MEIcoder for 300 epochs using the Adam W optimizer [26] with a learning rate and weight decay found using hyperparameter search and the validation dataset. Similar to early stopping [29], we pick the best model from training based on the Alex(5) score measured on the validation dataset. ... We provide hyperparameters of MEIcoder used for the final experiments in Table 5. Additional settings that we kept the same across all datasets include: Number of compressed neural map channels dc (readin): 480, Number of CNN channels (core): 480, 256, 256, 128, 64, 1, Kernel sizes (core): 7, 5, 5, 3, 3, 3, Padding (core): 3, 2, 2, 1, 1, 1, Stride (core): 1, 1, 1, 1, 1, 1, Dropout probability (core): 0.35. |