Neurally-Guided Structure Inference

Authors: Sidi Lu, Jiayuan Mao, Joshua Tenenbaum, Jiajun Wu

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 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. ... To evaluate the accuracy of our approach, we replicate the experiments in Grosse et al. (2012), including one synthetically generated dataset and two real-world datasets: motion capture and image patches.
Researcher Affiliation Academia 1Shanghai Jiao Tong University, 2MIT CSAIL, 3IIIS, Tsinghua University, 4Department of Brain and Cognitive Sciences, MIT, 5Center for Brains, Minds and Machines (CBMM), MIT.
Pseudocode Yes Algorithm 1 Neurally-Guided Structure Inference Function Infer(D, Type): rule Select Rule(D, Type) for each non-terminal symbol s in rule do Cs Decompose Data(D, rule, s) Replace s in rule with Infer(Cs, s) return rule
Open Source Code Yes Project Page: http://ngsi.csail.mit.edu.
Open Datasets Yes Looking at hand-written digits from the MNIST dataset (Le Cun et al., 1998)... The dataset of human motion capture (Hsu et al., 2005; Taylor et al., 2007)... The natural image patches dataset contains samples from the Sparsenet dataset proposed in Olshausen & Field (1996).
Dataset Splits No The paper mentions generating synthetic data for training and testing generalizability on programs of different lengths/depths, but it does not specify explicit training/validation/test split percentages or sample counts.
Hardware Specification Yes We ran all experiments on a machine with an Intel Xeon E5645 CPU and a GTX 1080 Ti GPU.
Software Dependencies No The paper mentions using the Adam optimizer and CNN-GRU models but does not specify version numbers for any software libraries or dependencies (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes We train the model with the Adam optimizer (Kingma & Ba, 2015). The hyperparameters for the optimizer are set to be β1 = 0.9, β2 = 0.9, α = 10 4. The model is trained for 100,000 iterations, with a batch size of 100. ... We adopt a unidirectional GRU with a hidden dimension of 256 as the code string encoder for production rule selection. We train the model using the Adam optimizer, with hyperparameters β1 = 0.9, β2 = 0.9, α = 10 4. The batch size is set to 64.