A Solver-free Framework for Scalable Learning in Neural ILP Architectures

Authors: Yatin Nandwani, Rishabh Ranjan, - Mausam, Parag Singla

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present several experiments on problems which require learning of ILP constraints and cost, with both symbolic as well as perceptual input. These include solving a symbolic sudoku as well as visual sudoku... Experiments on several problems, both perceptual as well as symbolic, which require learning the constraints of an ILP, show that our approach has superior performance and scales much better compared to purely neural baselines and other state-of-the-art models that require solver-based training.
Researcher Affiliation Academia Yatin Nandwani , Rishabh Ranjan , Mausam & Parag Singla Department of Computer Science, Indian Institute of Technology Delhi, INDIA {yatin.nandwani, rishabh.ranjan.cs118, mausam, parags}@cse.iitd.ac.in
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code available at: https://github.com/dair-iitd/ilploss
Open Datasets Yes For 9x9 sudokus, we use a standard dataset from Kaggle [Park, 2017], for k = 4 we use publically available data from Arcot and Kalluraya [2019], and for k = 6 we use the data generation process described in Nandwani et al. [2022]. We randomly select 10,000 samples for training, and 1000 samples for testing for each k. To generate the input images for visual-sudoku, we use the official train and test split of MNIST[Deng, 2012].
Dataset Splits No We randomly select 10,000 samples for training, and 1000 samples for testing for each k. (No explicit mention of validation set size or splitting method, though it refers to 'best val set accuracy')
Hardware Specification No We thank IIT Delhi HPC facility for computational resources. ... See appendix for the details of the ILP solver used in our experiments, the hardware specifications, the hyper-parameters, and various other design choices. (The main paper text does not provide specific hardware models.)
Software Dependencies No See appendix for the details of the ILP solver used in our experiments... Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2022. URL https://www.gurobi.com. (No specific software version numbers are provided in the main text).
Experiment Setup No µ+ and µ are the hyperparameters representing the margins for the positive and the negative points respectively. and The temperature parameter τ needs to be annealed as the training progresses. and We pick an ϵ small enough.... However, it also states: See the appendix for the details... the hyper-parameters, and various other design choices. The main text describes the type of hyperparameters but does not provide concrete values for them.