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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning with Statistical Equality Constraints
Authors: Aneesh Barthakur, Luiz Chamon
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we demonstrate the empirical performance of Algorithm 1 on instances of (P) presented in Section 5. Detailed descriptions of the experiments can be found in Appendix E. We showcase its effectiveness and the new formulations enabled by equality constraints in fair learning, interpolating classifiers, and boundary value problems. |
| Researcher Affiliation | Academia | Aneesh Barthakur University of Stuttgart EMAIL Luiz F. O. Chamon École polytechnique, Institut Polytechnique de Paris EMAIL |
| Pseudocode | Yes | Algorithm 1 Primal-dual constrained learning algorithm |
| Open Source Code | Yes | The python code used to reproduce all experiments in the camera-ready will be uploaded to Git Hub. |
| Open Datasets | Yes | Figure 1 compares models trained on the COMPAS dataset [66] using the fairness formulation in (P-DP)... We used the CIFAR-10 and CIFAR-100 datasets from the torchvision library. |
| Dataset Splits | Yes | Split the data into train (70%) and test (30%) sets. This process was repeated 10 times to produce 10 different train-test splits. |
| Hardware Specification | Yes | The fairness experiments were run on CPU-only nodes with 32 threads, with about 10-20 runs in parallel. ... The BVP experiments were run (one run at a time) on an accelerated node with 16 CPU threads and one GPU. ... The interpolation experiments were run (one run at a time) on a workstation with 12 CPU threads and one GPU. |
| Software Dependencies | No | Our implementation was made with pytorch and other standard ML libraries, and our codebase can be found at https://github.com/abarthakur/equality-constrained-learning. All experiments were run on an internal computing resources. ... The Adam optimizer was used for both primal (step 5 of Alg. 1) and dual updates (step 6 7 of Alg. 1)... |
| Experiment Setup | Yes | The Adam optimizer was used for both primal (step 5 of Alg. 1) and dual updates (step 6 7 of Alg. 1) with learning rates of 0.2 (primal) and 0.001 (dual). The other hyperparameters for Adam were set to their default values (ϵ = 10 8, β1 = 0.9, β2 = 0.999). A logistic regression classifier was used as the model. The cross-entropy loss was used as the objective. ... All models were trained for 300k iterations or epochs. Both methods (PINN/ (P-BVP)) used a 4 layered MLP with 50 hidden neuron layers and Tanh activation to represent the primal model fθ. Adam was used to optimize the primal model, with an initial learning rate of 1e-3, and a learning rate scheduler was used that multiplies the learning rate by a factor 0.9 every 5000 steps/epochs (available as the class Step LR from pytorch). ... SGD was used to optimize the primal model for both datasets, with a batch size of 128, and an initial learning rate of 1e-3, and a momentum hyperparameter of 0.9. |