Authors
Sidi Lu, Jiayuan Mao, Joshua Tenenbaum, and Jiajun Wu
Overview
Neurally-Guided Structure Inference, or NG-SI, combines the robustness of explicit search with the efficiency of learned inference. A neural network guides a hierarchical, layer-wise search over a compositional structure space instead of predicting the complete structure in one step.
This hybrid design is intended to generalize beyond the structural complexity seen during training without paying the full cost of exhaustive search.
Results
The method is evaluated on probabilistic matrix decomposition and symbolic program parsing. On both tasks, it outperforms purely data-driven and purely search-based alternatives.
Paper
Status
ICML 2019 Short Oral.
Citation
Sidi Lu, Jiayuan Mao, Joshua Tenenbaum, and Jiajun Wu. “Neurally-Guided Structure Inference.” Proceedings of the 36th International Conference on Machine Learning, PMLR 97:4144–4153, 2019.