Research map / Learning theory and algorithms / Prediction and decision losses
Locally Sparse Neural Networks for Tabular Biomedical Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use sample-specific locally sparse gates with a jointly trained gating network | LSPIN · Synthetic datasets E1–E5 | feature-selection F1 and prediction performance: E1 F1 0.9761; E2 F1 0.8600; E3 F1 0.9296; E4 F1 0.9615; E5 F1 1; E5 MSE 0.0019 | Helped |
| Use linear locally sparse prediction (LLSPIN) instead of nonlinear LSPIN | LSPIN · Biomedical tabular classification | accuracy: LLSPIN median rank 1; LSPIN median rank 4.5 | Mixed |
| Add the second, affinity-based gate regularization term (λ2) (ablation) | LLSPIN · MNIST | stability of selected features | Helped |
| Add the second regularization term (λ2) (ablation) | LLSPIN · MNIST | stability of selected features | Helped |
| Use LLSPIN rather than a fully connected MLP | MNIST · Low sample size MNIST | accuracy: 94.2% -> 94.8% | Helped |
| Use LLSPIN rather than a fully connected network | MNIST · Rotated MNIST, angles 10–70 | NLL: 10: 1.987 -> 1.403; 20: 6.455 -> 3.777; 30: 19.859 -> 10.341; 40: 43.771 -> 20.532; 50: 71.346 -> 31.006; 60: 96.177 -> | Helped |
| Use LLSPIN rather than a fully connected network | LLSPIN · Rotated MNIST, angles 80–90 | accuracy: 80: 14.24 vs 14.6; 90: 11.7 vs 12.3 | Hurt |
| Use LLSPIN rather than fully connected network or INVASE | LLSPIN · MNIST and real-world datasets | NLL: MNIST 0.826 vs 1.043 and 2.008; low sample size MNIST 1.985 vs 3.311 and 4.675; TOX-171 0.760±0.411 vs 3.968±0.500 and 2 | Helped |
| Modify the second regularizer to encourage dissimilar gates for dissimilar cells | LLSPIN · Single-nucleus RNA-sequencing marker gene identification | marker-gene F1: 0.9950 vs 0.4900 (INVASE) | Helped |
| Use a nonlinear prediction network for unequal group-specific regression coefficients | LSPIN · Unequal-coefficient synthetic regression | test MSE: 0.000199 | Helped |
| Use sample-specific feature selection rather than one global feature subset | LSPIN · Synthetic data E1–E5 | feature-selection F1: E1: 0.9761 vs 0.8306; E2: 0.8600 vs 0.7089; E3: 0.9296 vs 0.7823; E4: 0.9615 vs 0.8511 | Helped |
Closest papers by meaning
- ProtoGate: Prototype-based Neural Networks with Global-to-local Feature Selection for Tabular Biomedical Data (ICML 2024 · similarity 0.85)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.84)
- BiSHop: Bi-Directional Cellular Learning for Tabular Data with Generalized Sparse Modern Hopfield Model (ICML 2024 · similarity 0.84)
- Causal Representation Learning from Multimodal Biomedical Observations (ICLR 2025 · similarity 0.84)
- BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation (NeurIPS 2026 · similarity 0.83)
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.