Learning Generalized Gumbel-max Causal Mechanisms
Authors: Guy Lorberbom, Daniel D. Johnson, Chris J. Maddison, Daniel Tarlow, Tamir Hazan
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically we show that the mechanisms can be learned using a variant of Gumbel-softmax relaxation [Maddison et al., 2017, Jang et al., 2017], and that the resulting mechanisms improve over Gumbel-max and other fixed mechanisms. Further, we show that the learned mechanisms generalize, in the sense that we can learn a causal mechanism from a training set of observed outcomes and counterfactual queries and have it generalize to a test set of observed outcomes and counterfactual queries that were not seen at training time. |
| Researcher Affiliation | Collaboration | Guy Lorberbom Technion Haifa, Israel guy_lorber@campus.technion.ac.il Daniel D. Johnson Google Research Toronto, ON, Canada ddjohnson@google.com Chris J. Maddison University of Toronto & Vector Institute Toronto, ON, Canada cmaddis@cs.toronto.edu Daniel Tarlow Google Research Montreal, QC, Canada dtarlow@google.com Tamir Hazan Technion Haifa, Israel tamir.hazan@technion.ac.il |
| Pseudocode | No | The paper describes the proposed methods (Gadget 1 and Gadget 2) using mathematical equations and textual descriptions, but it does not include any formally labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | An implementation of our approach and instructions for reproducing our experiments is available at https: //github.com/google-research/google-research/tree/master/gumbel_max_causal_gadgets. |
| Open Datasets | No | The paper mentions using "datasets of fixed and random logits" and drawing "20,000 patient trajectories from the simulator" but does not provide specific access information (link, DOI, formal citation) to these datasets or explicitly state they are publicly available. |
| Dataset Splits | No | The paper refers to "training set" and "test set" but does not specify exact percentages, absolute counts, or detailed methodologies for train/validation/test splits. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory, or cloud instances) used for running the experiments. |
| Software Dependencies | No | The paper states: "Our code is implemented in JAX [Bradbury et al., 2018] and PyTorch [Paszke et al., 2019]." While it names the software, it does not specify exact version numbers for JAX or PyTorch. |
| Experiment Setup | No | The paper provides some general experimental conditions, such as using "20,000 patient trajectories" and averaging results across "10 different random seeds", but it lacks specific hyperparameter values (e.g., learning rate, batch size, number of epochs) or detailed training configurations required for reproducibility. |