Neurally-Guided Structure Inference

June 17, 2019 · Entered Twilight · 🏛 International Conference on Machine Learning

"No code URL or promise found in abstract"
"Code repo scraped from project page (backfill)"

Evidence collected by the PWNC Scanner

Repo contents: .gitignore, CNNGuider.py, README.md, algorithms, calc_time.py, concat_experiment.py, config.py, data_loader.py, example.py, example_data, experiments.py, experiments_roger.py, grammar.py, guided_synthetic_data.py, image_patch_experiment.py, initialization.py, models.py, motion_experiment.py, observations.py, parallel.py, parser.out, parsetab.py, parsing.py, predictive_distributions.py, presentation.py, recursive.py, scoring.py, single_process.py, train_guider.py, utils

Authors Sidi Lu, Jiayuan Mao, Joshua B. Tenenbaum, Jiajun Wu arXiv ID 1906.07304 Category cs.LG: Machine Learning Cross-listed cs.AI, cs.SC, stat.ML Citations 7 Venue International Conference on Machine Learning Repository https://github.com/desire2020/NGSI ⭐ 9 Last Checked 10 days ago

Abstract

Most structure inference methods either rely on exhaustive search or are purely data-driven. Exhaustive search robustly infers the structure of arbitrarily complex data, but it is slow. Data-driven methods allow efficient inference, but do not generalize when test data have more complex structures than training data. In this paper, we propose a hybrid inference algorithm, the Neurally-Guided Structure Inference (NG-SI), keeping the advantages of both search-based and data-driven methods. The key idea of NG-SI is to use a neural network to guide the hierarchical, layer-wise search over the compositional space of structures. We evaluate our algorithm on two representative structure inference tasks: probabilistic matrix decomposition and symbolic program parsing. It outperforms data-driven and search-based alternatives on both tasks.