Inducing Causal Structure for Interpretable Neural Networks
Authors: Atticus Geiger, Zhengxuan Wu, Hanson Lu, Josh Rozner, Elisa Kreiss, Thomas Icard, Noah Goodman, Christopher Potts
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate IIT on a structural vision task (MNIST-PVR), a navigational language task (Rea SCAN), and a natural language inference task (MQNLI). We compare IIT against multi-task training objectives and data augmentation. In all our experiments, IIT achieves the best results and produces neural models that are more interpretable in the sense that they more successfully realize the target causal model. |
| Researcher Affiliation | Academia | 1Stanford University, Stanford, California. Correspondence to: Atticus Geiger <atticusg@stanford.edu>, Zhengxuan Wu <wuzhengx@stanford.edu>. |
| Pseudocode | Yes | Pseudocode for interchange intervention training. |
| Open Source Code | Yes | We release our code at https://github.com/frankaging/Interchange-Intervention-Training. |
| Open Datasets | Yes | Our first benchmark is MNIST Pointer-Value Retrieval (MNIST-PVR; Zhang et al. 2021), a visual reasoning task constructed using the MNIST dataset (Le Cun et al., 2010). and Our second benchmark is Rea SCAN (Wu et al., 2021), a synthetic command-based navigation task that builds off the SCAN (Lake & Baroni, 2018) and g SCAN (Ruis et al., 2020) benchmarks. and Our final benchmark is MQNLI (Geiger et al., 2019), a synthetic natural language inference dataset |
| Dataset Splits | Yes | The train/test split designed by Zhang et al. (2021) creates a distributional shift between the training and testing data and The best model is picked by performance on a smaller development set of 2,000 examples, which is consistent with the training pipeline proposed in Ruis et al. (2020) for g SCAN. and For our experiments, we used a train set with 500K examples, a dev set with 60k examples, and a test set with 10K examples |
| Hardware Specification | Yes | The training time is about 1 day on a Standard Ge Force RTX 2080 Ti GPU with 11GB memory. |
| Software Dependencies | No | The paper mentions software components like PyTorch vision, Adam optimizer, and BERT, and the 'antra package' but does not provide specific version numbers for these software dependencies (e.g., 'PyTorch 1.9' or 'TensorFlow 2.x'). |
| Experiment Setup | Yes | The learning rate starts at 1e 4 and decays by 0.9 every 20,000 steps. We train the model for a fixed number of epochs (100,000) before stopping. and We use a batch size of 32. We use 5.0 10 5 as our learning rate, and use adamw optimization. We train for a maximum of 5 epochs. |