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 [1].
Sparsity regularization via tree-structured environments for disentangled representations
Authors: Elliot Layne, Jason Hartford, Sebastien Lachapelle, Mathieu Blanchette, Dhanya Sridhar
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the theory empirically with both simulations and ground-truth gene expression data. We find that TBR recovers the latent causal variables better than related methods across these settings, even under settings that violate some assumptions of the theory. |
| Researcher Affiliation | Collaboration | Elliot Layne EMAIL Mc Gill University, Mila Jason Hartford EMAIL Valence Labs Sébastien Lachapelle EMAIL Université de Montréal, Mila Samsung SAIT AI Lab, Montreal Mathieu Blanchette EMAIL Mc Gill University, Mila Dhanya Sridhar EMAIL Université de Montréal, Mila |
| Pseudocode | No | The paper describes methods and proofs using mathematical formulations (e.g., Eq. 1, Eq. 2) and prose, but does not include any clearly labeled pseudocode or algorithm blocks. Section 3.2 is a mathematical proof. |
| Open Source Code | Yes | Software and Data Relevant code for replicating experimentation is released on Github. |
| Open Datasets | Yes | Then, we apply TBR to a dataset of gene expression measurements across different cell types, where we hold out some genes as latent variables and simulate phenotypes derived from these latents. We performed all experimentation on the publicly available GTEx V8 sn RNA-Seq dataset, described by Eraslan et al. (2022). |
| Dataset Splits | Yes | We trained models on 50% of the data. A validation set of 25% of the data was used to tune hyper-parameters (learning rate, λ for TBR models), and the remaining 25% was held out as a test set for evaluation. |
| Hardware Specification | Yes | Experimentation on the fully simulated dataset was performed on a Mac Book Pro with 18GB memory and an M3 Pro CPU. Experimentation on gene-expression data described in 4.1 was performed on a shared compute server with a total of 755GB of ram and 96 CPU cores. Training leveraged shared use of a Quadro RTX 6000 GPU, with an average GPU memory usage of approximately 1GB. |
| Software Dependencies | No | All models were implemented in Py Torch Paszke et al. (2019). We instantiated ˆΨ as a neural network with two hidden layers (256 units, 64 units) and Leaky Relu activations. All preprocessing of the GTEx data was performed using standard functions within the Scanpy library (Wolf et al., 2018). We jointly optimized ˆΨθ, ˆw0 and ˆ to produce optimal predictions ˆYe across all environments. Additionally, we regularize our estimate ˆ with a sparsity-inducing norm || ˆ ||0. In each experiment, ˆΨ and the relevant estimation of ˆΦ were jointly optimized with the Adam optimizer Kingma & Ba (2014). The paper mentions several software components like PyTorch, Scanpy, and Adam optimizer, but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | Implementation We instantiated ˆΨ as a neural network with two hidden layers (256 units, 64 units) and Leaky Relu activations. Performance was compared across two methods of estimating linear map ˆΦ. In each experiment, ˆΨ and the relevant estimation of ˆΦ were jointly optimized with the Adam optimizer Kingma & Ba (2014). Hyper-parameters are summarized in Table 1. Table 1: Summary of hyper-parameter settings. Experiment / Dataset Parameter name Setting or range Simulated λ 0.001 Simulated Learning rate 0.001 Gene Expression λ {0.0, 0.1, 0.01, 0.001, 0.0001} Gene Expression Learning rate {0.001, 0.0001} |