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..
Learning to cluster neuronal function
Authors: Nina Nellen, Polina Turishcheva, Michaela Vystrčilová, Shashwat Sridhar, Tim Gollisch, Andreas S. Tolias, Alexander Ecker
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments Clustering loss hyperparameters. For the clustering loss, we fixed the degrees of freedom to ν = 2.1, just above the threshold where the variance ν ν 2Σ becomes defined (only for ν > 2). To balance model flexibility and robustness, we allowed each cluster to have its own diagonal scale matrix Σj, which alloed for different variances per embedding dimension while preventing overfitting. For each dataset, we adjusted the clustering strength β R such that it is in the same order of magnitude as the model loss at initialization. Evaluation of model performance. Building on previous research [8, 14, 16, 34, 43, 44], we evaluated the model s predictive performance by computing the Pearson correlation (across images in the test set) between the measured and predicted neural responses, averaged across neurons. Evaluation of embedding consistency. We wanted to assess the relative structure of the embedding space: Do the same groups of neurons consistently cluster together across models fit with different initial conditions? To quantify this notion, we took DECEMber s cluster assignments and measured how often neuron pairs are assigned to the same group using the Adjusted Rand Index (ARI) [22], which quantifies the similarity between two clustering assignments, X and Y . The ARI remains unchanged under permutations of cluster labels. ARI equals one if and only if the two partitions are identical and it equals zero when the partitions agreement is no better than random. |
| Researcher Affiliation | Academia | 1 Institute of Computer Science and Campus Institute Data Science, University Göttingen, Germany 2 University Medical Center Göttingen, Department of Ophthalmology, Göttingen, Germany 3 Bernstein Center for Computational Neuroscience Göttingen, Göttingen, Germany 4 Cluster of Excellence Multiscale Bioimaging: from Molecular Machines to Networks of Excitable Cells (MBEx C), University of Göttingen, Germany 5 Else Kröner Fresenius Center for Optogenetic Therapies, University Medical Center Göttingen, Germany 6 Department of Ophthalmology, Byers Eye Institute, Stanford University School of Medicine, CA, US 7 Stanford Bio-X, Stanford University, Stanford, CA, US 8 Wu Tsai Neurosciences Institute, Stanford University, CA, US 9 Department of Electrical Engineering, Stanford University, CA, US 10 Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany Shared contribution EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Model Training with clustering loss Inputs: Degrees of freedom ν, clustering weight β, core parameters θ, neuronal embeddings (readout) Z Output: Parameters µj, Σj, θ and Z Pretraining: Train the predictive model by optimizing Lmodel w.r.t. θ and Z for m epochs Initialize: Cluster centers µj with k-means and diagonal scale matrix Σj as within-cluster variance for epoch t = 1 to T do for minibatch b in dataset do (1) E-step (Expectation): Compute 1.1 Soft assignments qij = ft(zi;µj,Σj,ν) PJ j =1 ft(zi;µj ,Σj ,ν) (3.2) 1.2 Latent scales uij = ν+K ν+(zi µj) Σ 1 j (zi µj) (3.3) (2) M-step (Maximization): Update parameters 2.1 Update µj = PN i=1 qijuij zi PN i=1 qijuij (3.4) 2.2 Update Σj = PN i=1 qijuij(zi µj)(zi µj) PN i=1 qij (3.5) (3) Gradient step: Optimize predictive model parameters 3.1 Minimize L = Lmodel + βKL(Q||P) w.r.t θ, Z with pij = q2 ij/fj P k qik/fk and fj = P return µ, Σ, θ, Z |
| Open Source Code | Yes | The code is available at https://github.com/Nisone2000/DECEMber, https://github.com/ecker-lab/cnn-training. |
| Open Datasets | Yes | To check whether DECEMber works on real data, we applied it on marmoset retinal ganglion cells (RGCs) where the existence of discrete cell types is well established [47]. We used data from two male marmoset retinas published by Sridhar et al. [48], where the neural activity was recorded using a micro-electrode array while presenting grayscale natural movies. We used SENSORIUM 2022 dataset and baseline architecture to train a model to predict responses of mouse primary visual cortex to grayscale images. for seven mice (more detail on data in App. A.5). For the mouse retina we used both the data and the models from Höfling et al. [18]. |
| Dataset Splits | Yes | The validation set contains responses to roughly 500 and test set to 5000 images per mouse. |
| Hardware Specification | Yes | We use a local infrastructure cluster with 8 NVIDIA RTX A5000 GPUs with 24Gb of memory each for mouse experiments. For mouse retina, marmoset retina, and monkey V4 we used 40Gb NVIDIA A100. |
| Software Dependencies | No | The learning rate of both pretraining and training with the clustering loss was initially 0.005 and reduced during training using the Reduce LROn Plateau learning rate scheduler, patience 10 and minimal learning rate 1e 8. |
| Experiment Setup | Yes | For marmoset RGC dataset, we used the three layer CNN described in [15]. We trained it for a maximum of 1000 epochs, stopping early if validation correlation did not improve for 20 epochs. The learning rate of both pretraining and training with the clustering loss was initially 0.005 and reduced during training using the Reduce LROn Plateau learning rate scheduler, patience 10 and minimal learning rate 1e 8. For SENSORIUM 2022, we used their model and training hyperparameters for the baselines training. Pretraining duration, learning rates and clustering strength β is reported in every experiment. |