Counterfactual Visual Explanations
Authors: Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, Stefan Lee
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply our approach to multiple image classification datasets generating qualitative results showcasing the interpretability and discriminativeness of our counterfactual explanations. To explore the effectiveness of our explanations in teaching humans, we present machine teaching experiments for the task of fine-grained bird classification. |
| Researcher Affiliation | Collaboration | 1Georgia Institute of Technology 2Siemens Corporation. |
| Pseudocode | Yes | Algorithm 1 Greedy Sequential Search Data: query image I with class c, distractor I with class c Result: list of edits S that change the model decision S [ ] F f(I) F f(I ) /* Until decision is changed to c */ while c = argmax g(F ) do /* Find single best edit excluding previously edited cells in S */ i, j Best Edit(F , F , S) /* Apply the edit and record it */ F i, = F j , S.append({i, j }) end |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code or links to a code repository for the described methodology. |
| Open Datasets | Yes | We apply our approach on four different datasets SHAPES (Andreas et al., 2016) (in supplement due to space constraints), MNIST (Le Cun et al., 1998), Omniglot (Lake et al., 2015) and Caltech-UCSD Birds (CUB) 2011 Dataset (Wah et al., 2011), and present results showcasing the intepretability and discriminativeness of our counterfactual explanations. |
| Dataset Splits | No | The paper mentions a 'random train/test split of 80/20%' for Omniglot, but does not explicitly state a validation split for any of the datasets used, nor does it specify exact splits for MNIST or CUB beyond implying a test set. |
| Hardware Specification | Yes | Our approach takes 15 µs per image on a Titan XP GPU. Runtime is 9 µs per image on a Titan XP GPU. Runtimes are 1.85 and 1.34 sec/image for random and NN distractor classes respectively on a Titan XP GPU. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific library versions). |
| Experiment Setup | Yes | In all our experiments, we operate in output space of the last convolutional layer in the CNN but our approach is equally applicable to the output of any convolutional layer. We train a CNN model consisting of 2 convolutional layers and 2 fully-connected layers on this dataset. We use gradient descent with a learning rate of 0.3. |