Locally Sparse Neural Networks for Tabular Biomedical Data

Authors: Junchen Yang, Ofir Lindenbaum, Yuval Kluger

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that our method outperforms state-of-the-art models when applied to synthetic or real-world biomedical datasets using extensive experiments. Furthermore, the proposed framework dramatically outperforms existing schemes when evaluating its interpretability capabilities.
Researcher Affiliation Academia 1Interdepartmental Program in Computational Biology and Bioinformatics, Yale University, New Haven, CT, USA 2Faculty of Engineering, Bar Ilan University, Ramat Gan, Israel 3Applied Math Program, Yale University, New Haven, CT, USA 4Department of Pathology, School of Medicine, Yale University, New Haven, CT, USA.
Pseudocode Yes Algorithm 1 Locally SParse Interpretable Networks (LSPIN) Pseudo-code
Open Source Code No The software dependencies are specified in the associated codes.
Open Datasets Yes We use MNIST handwritten dataset as a table... We apply the integrated models (COX-LLSPIN/COX-LSPIN) on a Surveillance, Epidemiology, and End Results (SEER) breast cancer dataset www.seer.cancer.gov to perform survival analysis... The purified Peripheral Blood Mononuclear Cells (PBMC) dataset is collected from (Zheng et al., 2017).
Dataset Splits Yes For the BASEHOCK, RELATHE, PCMAC, and PBMC datasets, 5% of each dataset is set aside as a validation set. We split the remaining 95% of the data into 5 non-overlapping folds, with 1 fold for training and the remaining folds for testing each time (see details in Appendix section B.5).
Hardware Specification Yes The CPU model used for the experiments is Intel(R) Xeon(R) Gold 6150 CPU @ 2.70GHz (72 cores total). GPU model is NVIDIA Ge Force RTX 2080 Ti. The operating system is Ubuntu 20.04.2 LTS. The memory storage is 1 TB in total.
Software Dependencies No The software dependencies are specified in the associated codes.
Experiment Setup Yes For lasso, we optimize the l1 regularization parameter with 20 trials and the grid range is [1e-2, 1e3]... For MLP/STG(Linear&Nonlinear)/LLSPIN/LSPIN/INVASE/L2X/REAL-x, the parameter settings and grids are listed in Table B.2. The number of hidden layers and nodes are identical for these models... For LLSPIN/LSPIN/STG/Neural Network model, the prediction network architecture is set to 3 hidden layers with 100, 50, 30 neurons, respectively, for all the datasets.