Research map / Label-efficient and robust learning / Recognition and long-tail
Learning to cluster neuronal function
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adds a Student's t-mixture EM-based DEC-style clustering loss to neuronal predictive models so per-neuron readout embeddings form more consistent clusters, finding mouse V1 shows only a weak, non-discrete structure.
Problem
Per-neuron functional embeddings from predictive models of visual cortex do not form clear clusters, so it is unclear whether discrete functional cell types exist in mouse V1 or whether current models fail to reveal them; cluster consistency across seeds is low.
Approach
Augments the predictive-model loss with a DEC-style KL clustering loss where clusters are multivariate Student's t-mixtures with learned diagonal scale matrices, updated by EM each batch, avoiding the centroid collapse of vanilla DEC when the embedding scale is set by regression.
Key results
- On marmoset retina, DECEMber reaches ARI 0.96 ± 0.01 with test correlation 0.81 ± 0.07, outperforming GMM and k-means on consistency.
- On mouse V1, ARI roughly doubles relative to the GMM baseline for clustering-loss strengths up to 1e4, but predictive performance drops for larger beta and no clear ARI peak at any cluster count appears.
- On macaque V4 with an attention readout, embedding consistency roughly doubles while predictive performance is preserved.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add clustering loss with EM-updated t-mixture parameters | CNN neuronal response model · Marmoset retina 1 | ARI | Helped |
| Add clustering loss with EM-updated t-mixture parameters | CNN neuronal response model · Mouse visual cortex | ARI: ARI values double compared to the baseline model | Helped |
| Increase clustering strength above 10^4 (ablation) | DECEMber on mouse visual cortex · Mouse visual cortex | predictive performance | Hurt |
| Increase clustering strength above 10^4 (ablation) | DECEMber on mouse visual cortex · Mouse visual cortex | ARI | No clear effect |
| Pretrain for 10 epochs (ablation) | DECEMber · Mouse visual cortex | ARI | Helped |
| Vary the number of pretraining epochs while tuning the learning rate (ablation) | DECEMber · Mouse visual cortex | predictive performance | No clear effect |
| Use a learned diagonal scale matrix for each cluster instead of fixed unit scale (ablation) | DECEMber clustering loss · Synthetic linear-neuron toy example | cluster separation | Helped |
| Use vanilla DEC with fixed unit scale instead of learned cluster scales (ablation) | Clustering of pretrained linear-neuron weights · Synthetic linear-neuron toy example | cluster separation | Hurt |
| Use DECEMber with different pretraining lengths (ablation) | Höfling et al. mouse retina model · Mouse retina test set | Pearson correlation: 0.4727 $\pm$ 0.0008; 0.4695 $\pm$ 0.0009; 0.4732 $\pm$ 0.0008; 0.4727 $\pm$ 0.0009 | No clear effect |
| Use DECEMber with different pretraining lengths (ablation) | Pierzchlewicz et al. macaque V4 model · Macaque V4 test set | Pearson correlation: 0.308 $\pm$ 0.004; 0.308 $\pm$ 0.006; 0.304 $\pm$ 0.005; 0.305 $\pm$ 0.003 | No clear effect |
Surprising findings
- Vanilla DEC loss collapses cluster centers to a single point on a toy linear regression example, because the fixed unit-scale t-distribution is ill-suited to regression-scaled embeddings.
- The authors conclude mouse V1 excitatory neurons likely form a functional continuum rather than discrete clusters, since ARI showed no peak at a particular cluster number.
Limitations
- Requires a predefined number of clusters and multiple runs to choose it, increasing compute cost.
- Clustering strength beta must be tuned, trading off ARI against predictive performance; the pretraining length also needs tuning.
- High-dimensional covariance matrices can be ill-conditioned; the clamping fix is heuristic.
Experimental setup
Datasets: SENSORIUM 2022 marmoset retinal ganglion cell recordings (Sridhar et al.) mouse retina dataset (Höfling et al.) macaque V4 recordings (Willeke et al.)
Benchmarks: SENSORIUM 2022 marmoset retinal ganglion cell classification mouse retina macaque V4
Metrics: Adjusted Rand Index Pearson correlation Fowlkes-Mallows index V-measure
Baselines: GMM k-means vanilla DEC loss Turishcheva et al. rotation-equivariant model baseline predictive model without clustering loss
Builds on: Deep Embedding Clustering Sensorium 2022 core-readout model Turishcheva et al. adaptive log-norm regularization Gaussian readout
Closest papers by meaning
- Reproducibility of predictive networks for mouse visual cortex (NeurIPS 2024 · similarity 0.93)
- Rotation-invariant clustering of neuronal responses in primary visual cortex (ICLR 2020 · similarity 0.90)
- Most discriminative stimuli for functional cell type clustering (ICLR 2024 · similarity 0.90)
- Neural Regression, Representational Similarity, Model Zoology & Neural Taskonomy at Scale in Rodent Visual Cortex (NeurIPS 2021 · similarity 0.89)
- Multi-Modal Latent Variables for Cross-Individual Primary Visual Cortex Modeling and Analysis (AAAI 2025 · similarity 0.85)
- Modeling Dynamic Neural Activity by combining Naturalistic Video Stimuli and Stimulus-independent Latent Factors (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.