Robust Explanation Constraints for Neural Networks
Authors: Matthew Robert Wicker, Juyeon Heo, Luca Costabello, Adrian Weller
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, our method surpasses the robustness provided by previous heuristic approaches. We find that our training method is the only method able to learn neural networks with certificates of explanation robustness across all six datasets tested. |
| Researcher Affiliation | Collaboration | Matthew Wicker ,1, Juyeon Heo ,2, Luca Costabello3, Adrian Weller1,2 1 The Alan Turing Institute, 2 University of Cambridge, 3 Accenture Labs |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | No explicit statement or link to open-source code for the methodology described in this paper was found. |
| Open Datasets | Yes | Empirically, we test our framework on six datasets of varying complexity from tabular datasets in financial applications to medical imaging datasets. We study the Adult dataset, predicting whether or not an individual makes more or less than fifty thousand dollars a year, and the German Credit dataset, predicting whether an individual has a good or bad credit rating. We train two networks, a two-layer fully-connected neural network with 128 neurons per layer and a four layer convolutional neural network inspired by the architecture in (Gowal et al., 2018).We consider two datasets from the Med MNIST benchmark and a third in Appendix E. The first task, Derma MNIST, is to classify images into one of 11 different skin conditions. The second task, Pneumonia MNIST, is to classify chest x-rays into positive or negative diagnosis for pneumonia. |
| Dataset Splits | No | The paper does not explicitly provide specific percentages, sample counts, or citations to predefined splits for training, validation, or test sets in the provided text. It mentions 'test-set accuracy' and discusses training, but detailed split information, especially for validation, is not present here. |
| Hardware Specification | No | No specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments were provided in the paper. The paper mentions that "We report exact training hyper-parameters and details in the Appendix" but these are not in the main text. |
| Software Dependencies | No | The paper does not provide specific ancillary software details, such as library names with version numbers, needed to replicate the experiment. It mentions that "We report exact training hyper-parameters and details in the Appendix" but these are not in the main text. |
| Experiment Setup | Yes | At training time, we consider the input perturbation budget t as well as the model perturbation budget γt. The input interval is simply taken to be [x t, x + t] and the interval over a weight parameter W is [W |W|γt, W + |W|γt]. We train two networks, a two-layer fully-connected neural network with 128 neurons per layer and a four layer convolutional neural network inspired by the architecture in (Gowal et al., 2018). We train models with t = 0.01 and γt = 0.01 and test with = 0.025 and γ = 0.025. |