Causally Aligned Curriculum Learning

Authors: Mingxuan Li, Junzhe Zhang, Elias Bareinboim

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we validate our proposed methodology through experiments in discrete and continuous confounded tasks with pixel observations. and Finally, we validate the proposed framework through extensive experiments in various decision-making tasks.
Researcher Affiliation Academia Causal Artificial Intelligence Lab, Columbia University, USA
Pseudocode Yes Algorithm 1: FINDMAXEDIT, Algorithm 2: CURRICULUM LEARNING, Algorithm 3: FINDCAUSALCURRICULUM, Algorithm 4: ISEDIT, Algorithm 5: LISTEDIT, Algorithm 6: CAUSAL CURRICULUM LEARNING
Open Source Code No No explicit statement for open-source code release for their own work, nor a direct link to their repository, was found. The paper mentions using "Klink et al. (2022) s official implementations (https://github.com/psclklnk/currot)" for curriculum generators, which is third-party code.
Open Datasets Yes Colored Sokoban and Button Maze are implemented based on Sokoban (Schrader, 2018) and Grid World (Chevalier-Boisvert et al., 2018), respectively.
Dataset Splits No No explicit details about train/validation/test dataset splits (percentages, counts, or predefined splits) were provided.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions using 'Py Torch' for implementation but does not specify a version number. Other software dependencies are not listed with version numbers.
Experiment Setup Yes Both networks have three convolutional layers and two linear layers with rectified linear unit (Re LU) activation functions. The number of output channels of the convolutional layers is [32, 64, 64] with each layer. For convolutional kernels in those three convolutional layers, we use 8 × 8 with a stride of 4, 4 × 4 with a stride of 2, and 3 × 3 with a stride of 1, respectively. We flatten the output of the convolutional layers and feed it into the two linear layers with the intermediate feature dimension set to 512. The input for the network is an image observation of size 84 × 84 × 3 for both environments.