Research map / Architectures and efficiency / RNNs and normalization
Reproducibility of predictive networks for mouse visual cortex
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.
Studies how reproducible learned neuronal readout embeddings of CNN predictive models of mouse V1 are across seeds, and proposes adaptive per-neuron L1 regularization and iterative channel pruning to improve clustering consistency.
Problem
Neuronal readout embeddings from deep predictive models of mouse visual cortex are likely non-identifiable due to overparameterization, so it is unclear whether they can reliably define functional cell types via clustering.
Approach
Introduces an adaptive L1 regularization for the Gaussian readout with per-neuron learned coefficients under a log-normal hyperprior, plus an iterative channel-pruning procedure, and measures embedding consistency via ARI across seeds along with tuning-curve consistency.
Key results
- Adaptive L1 regularization (gamma=10, sigma=0.1) matches Gaussian readout predictive performance while giving more consistent embedding clusters than factorized and uniformly regularized Gaussian readouts (higher ARI in Fig. 2I).
- Iterative pruning removes 1/3 to 1/2 of core channels without performance loss and improves ARI for uniformly regularized Gaussian models, but not for the adaptive model.
- Regularized Gaussian readouts (gamma=50, 100) produce strongly biased tuning curves, while the factorized readout reproduces tuning properties more consistently across seeds (Table 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive per-neuron L1 regularization instead of uniform regularization | Gaussian readout · Sensorium 2022 | ARI of embedding clusters | Helped |
| Prune core feature channels after training (ablation) | Gaussian readout models · Sensorium 2022 | ARI of embedding clusters | Helped |
| Prune core feature channels after training (ablation) | Gaussian readout models · Sensorium 2022 | Predictive performance | Mixed |
| Add uniform L1 regularization to the Gaussian readout (ablation) | Gaussian readout · Sensorium 2022 | ARI of embedding clusters | No clear effect |
| Add uniform L1 regularization to the Gaussian readout (ablation) | Gaussian readout · Sensorium 2022 | Predictive performance | Hurt |
| Use Gaussian readout instead of factorized readout | Predictive network · Sensorium 2022 | ARI of embedding clusters | Hurt |
| Use older factorized readout instead of Gaussian readout | Predictive network · Sensorium 2022 | Tuning-index consistency | Helped |
| Increase uniform regularization to γ=50 or γ=100 (ablation) | Gaussian readout · Sensorium 2022 | Tuning-index consistency | Hurt |
Surprising findings
- Uniform L1 regularization of Gaussian readouts increases sparsity but does not improve cluster consistency across seeds.
- Older factorized readout reproduces neuronal tuning properties more consistently across model fits than the newer Gaussian readout, even though the newer readouts have better predictive performance.
Limitations
- Only one core architecture (rotation-equivariant CNN) is studied in the main experiments; the Vision Transformer core remains untested.
- Absolute cluster consistency remains modest (ARI low), and pruning trades off predictive performance.
- Tuning-curve validity is evaluated only qualitatively and by cross-seed consistency, not against ground truth.
Experimental setup
Datasets: Sensorium 2022 dataset
Benchmarks: Sensorium 2022 dataset validation/test correlation
Metrics: Pearson correlation adjusted rand index normalized mean absolute error of tuning indices trustworthiness Fowlkes-Mallows v-measure
Baselines: factorized readout Gaussian readout with uniform L1 regularization non-pruned Gaussian readout non-rotation-equivariant CNN control
Builds on: Sensorium 2022 baseline factorized readout (Klindt et al.) Gaussian readout (Lurz et al.) rotation-equivariant CNN (Ecker et al.) lottery ticket pruning
Closest papers by meaning
- Learning to cluster neuronal function (NeurIPS 2025 · similarity 0.93)
- Most discriminative stimuli for functional cell type clustering (ICLR 2024 · similarity 0.88)
- Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning (ICLR 2025 · similarity 0.87)
- Benchmarking Out-of-Distribution Generalization Capabilities of DNN-based Encoding Models for the Ventral Visual Cortex. (NeurIPS 2024 · similarity 0.85)
- NOBLE - Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models (NeurIPS 2025 · similarity 0.85)
- Long-Range Feedback Spiking Network Captures Dynamic and Static Representations of the Visual Cortex under Movie Stimuli (NeurIPS 2024 · 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.