Factorizing Declarative and Procedural Knowledge in Structured, Dynamical Environments

Authors: Anirudh Goyal, Alex Lamb, Phanideep Gampa, Philippe Beaudoin, Charles Blundell, Sergey Levine, Yoshua Bengio, Michael Curtis Mozer

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments address the following questions. (1) Does SCOFF successfully factorize knowledge into OFs and schemata? (2) Do the learned schemata have semantically meaningful interpretations? (3) Is the factorization of knowledge into object files and schemata helpful in downstream tasks? (4) Does SCOFF outperform state-of-the-art approaches, both modular and non-modular, which lack SCOFF s strong inductive bias toward systematicity and knowledge factorization?
Researcher Affiliation Collaboration 1 Mila, University of Montreal, 2 IIT BHU, Varanasi, 3 Waverly, 4 UC Berkeley, 5 Deepmind, 6 Google Research, Brain Team
Pseudocode Yes Algorithm 1 provides a precise specification of SCOFF broken into four steps. and G PSEUDOCODE FOR SCOFF ALGORITHM
Open Source Code No We include more experimental results in the Appendix, and we will release the code.
Open Datasets Yes We used the Intuitive Physics Benchmark (Riochet et al., 2019)..., We consider a bouncing-balls environment in which multiple balls move with billiard-ball dynamics (Van Steenkiste et al., 2018)., We use the partially-observable Goto Obj Maze environment from Chevalier-Boisvert et al. (2018)
Dataset Splits No The paper specifies training and test set sizes (e.g., 'The dataset consists of 50,000 training examples and 10,000 test examples'), but does not explicitly provide details about a validation dataset split.
Hardware Specification Yes It takes about 2 days to train the proposed model on bouncing ball task for 100 epochs on V100 (32G).
Software Dependencies No The paper mentions software components such as 'Adam optimizer' and implies the use of a deep learning framework (e.g., PyTorch through pseudocode constructs like 'torch.zeros'), but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Unless otherwise indicated, we always train in an end-to-end fashing using the Adam (Kingma and Ba, 2014) optimizer with a learning rate of 0.0001 and momentum of 0.9. As a default, we use nf = 6 and ns = 4, except where we are specifically exploring the effects of manipulating these hyperparameters. and Table 2: Hyperparameters for the adding generalization task