Directed Graph Grammars for Sequence-based Learning

Authors: Michael Sun, Orion Foo, Gang Liu, Wojciech Matusik, Jie Chen

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4. Experiments Directed Graph Grammars for Sequence-based Learning Table 2. Prior validity, uniqueness and novelty (%). We follow the same settings as Zhang et al. (2019). Methods Neural architectures Bayesian networks Accuracy Validity Uniqueness Novelty Accuracy Validity Uniqueness Novelty D-VAE 99.96 100.00 37.26 100.00 99.94 98.84 38.98 98.01 S-VAE 99.98 100.00 37.03 99.99 99.99 100.00 35.51 99.70 Graph RNN 99.85 99.84 29.77 100.00 96.71 100.00 27.30 98.57 GCN 98.70 99.53 34.00 100.00 99.81 99.02 32.84 99.40 Deep GMG 94.98 98.66 46.37 99.93 47.74 98.86 57.27 98.49 DIGGED (GNN) 100 100 98.7 99.9 100 100 97.6 100 DIGGED (TOKEN) 100 100 25.4 37.8 100 100 98.67 26.67
Researcher Affiliation Collaboration 1MIT CSAIL 2MIT 3University of Notre Dame 4MIT-IBM Watson AI Lab, IBM Research. Correspondence to: Michael Sun <EMAIL>.
Pseudocode Yes Further details and pseudocode are in App. B. In Algo. 4, we give the pseudocode of the disambiguation algorithm. Algorithm 1: function grammar induction(dataset)
Open Source Code Yes Code is available at https://github.com/ shiningsunnyday/induction.
Open Datasets Yes 1. Neural Architectures (ENAS). The ENAS dataset contains 19,020 neural architectures from the ENAS software and their weight-sharing accuracy (WS-Acc) on CIFAR10 (Pham et al., 2018). ... 2. Bayesian Networks (BN). The BN dataset contains 200,000 random, 8-node Bayesian networks from the R package bnlearn (Scutari, 2009) and their Bayesian Information Criterion (BIC) score for fitting the Asia dataset (Lauritzen & Spiegelhalter, 1988). ... 3. Analog Circuits (CKT). The CKT dataset contains 10000 operational amplifiers (op-amps) released by Dong et al. (2023)
Dataset Splits Yes 2. Predictive Performance. For property prediction, we train a Sparse Gaussian Process (SGP) regressor, following the same setup and hyperparameters as Zhang et al. (2019); Thost & Chen (2021); Dong et al. (2023). 3. Bayesian Optimization. We run batched Bayesian Optimization based on the SGP model for 10 rounds with 50 acquisition samples per round. We follow the same setup as Zhang et al. (2019) for ENAS and BN and Dong et al. (2023) for CKT
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or processor types used for running its experiments.
Software Dependencies No The paper mentions software components like the 'Subdue library' and 'networkx's cliques library' (implicitly referring to the Python library 'networkx') in Appendix B.1, but it does not provide specific version numbers for any of the software dependencies used in their methodology.
Experiment Setup Yes The optimal parameters for our model were determined using a hyperparameter scan sweeping over various properties of the VAE, using validation loss as the guide. During the scan, we explore varying architecture properties such as: number of encoder layers, number of decoder layers, latent dimension, embedding dimension, batch size, and KL divergence loss coefficient.