Source-linked AI summary
Generating Counterfactual Explanations with Natural Language
Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, Zeynep Akata
TL;DR
Textual explanations often describe features present in an image, while explanations of absent evidence that might change a decision are less readily available. This paper generates counterfactual explanations for an image and a counterfactual class by identifying potentially relevant missing evidence. Both proposed models outperform the baseline on the reported metrics, and examples show how absent features can help distinguish similar bird classes.
Problem
Textual explanation datasets describe attributes present in images, but ground-truth explanations of absent attributes that might change a classification decision are not readily available.
Source: §1 · p. 1Method
Given an image and counterfactual class, the method identifies class-discriminative candidate evidence, checks whether it is absent using flipped attributes, and generates a negated explanation.
Source: §2 · p. 1, §2.2 · p. 2Results
16.26% phrase error for the baseline; both CF models outperform it on Phrase Error and Accuracy w/FC Test.
Source: §3 · p. 3Takeaways & Limitations
For similar birds, an absent feature such as a black face can help distinguish classes when present-feature explanations overlap.
Source: §3 · pp. 3–4Without ground-truth counterfactual evidence, the evidence checker relies on human image descriptions and the assumption that many visual attributes are exclusive.
Source: §2.2 · p. 2
Abstract
from arXiv · showhide
Natural language explanations of deep neural network decisions provide an intuitive way for a AI agent to articulate a reasoning process. Current textual explanations learn to discuss class discriminative features in an image. However, it is also helpful to understand which attributes might change a classification decision if present in an image (e.g., "This is not a Scarlet Tanager because it does not have black wings.") We call such textual explanations counterfactual explanations, and propose an intuitive method to generate counterfactual explanations by inspecting which evidence in an input is missing, but might contribute to a different classification decision if present in the image. To demonstrate our method we consider a fine-grained image classification task in which we take as input an image and a counterfactual class and output text which explains why the image does not belong to a counterfactual class. We then analyze our generated counterfactual explanations both qualitatively and quantitatively using proposed automatic metrics.
1. Introduction
Existing textual explanations discuss why a network’s decision is reasonable, while counterfactual explanations reason about information absent from an image that might affect the decision if present. Because ground-truth counterfactual explanations are unavailable, the paper treats class-discriminative attributes absent from another class’s image as counterfactual evidence.
1. Introduction
- Counterfactual explanations describe information missing from an image that might affect a decision if it were present.They can help people understand why an input belongs to one class and not another.Source: §1 · p. 1
- Datasets often describe attributes present in images, but counterfactual explanations are harder to collect and ground-truth examples are unavailable.Collecting explanations for why an image does not belong to every other possible class would be required.Source: §1 · p. 1
- The paper treats an attribute discriminative for class A but absent from a class B image as counterfactual evidence against class A.The authors use this reasoning because they lack ground-truth counterfactual explanations.Source: §1 · p. 1
- Rather than perturbing natural images, the method reasons about counterfactuals in semantic space and checks whether class-discriminative evidence is present.The generated explanations are post-hoc and are not intended to reflect the neural network’s internal reasoning.Source: §1 · p. 1
2. Generating Counter-Factual Explanations
Given an image and a counterfactual class, the system identifies evidence associated with that class, checks whether it is absent from the image, and generates a sentence explaining why the image does not belong to that class. It uses class explanations, image descriptions, and rule-based text generation to implement these steps.
2. Generating Counter-Factual Explanations
- The system takes an image and a counterfactual class, then predicts candidate evidence, checks its presence, and generates a sentence about evidence that is absent.The sentence explains why the image does not belong to the counter-class.Source: §2 · pp. 1–2, Figure 1 · p. 2
2.1. Determining Candidate Counterfactual Evidence
- Candidate counterfactual evidence consists of noun phrases extracted from generated explanations for images in the counter-class.The method uses the explanation model of Hendricks et al. (2016) and a rule-based noun-phrase chunker.Source: §2.1 · p. 2
2.2. Evidence Checker
- Because ground-truth counterfactual evidence is unavailable, the evidence checker mines negative attributes using the assumption that many visual attributes are mutually exclusive.The system explores a phrase-image classifier and a phrase-critic; each selects candidate evidence with the minimum score, while the phrase-critic uses positive original sentences and negative examples with flipped attributes.Source: §2.2 · pp. 2–3
2.3. Generating Counterfactual Text
- A rule-based system negates absent noun phrases and appends them to text identifying the counter-class being compared.For example, it can turn “this bird has brown wings” into “this bird does not have brown wings.”Source: §2.3 · p. 3
3. Experiments
The experiments evaluate counterfactual explanations on CUB birds using nearby different-class images and proposed text-based metrics. Qualitative examples and quantitative comparisons show how missing attributes can distinguish closely related bird classes.
Dataset
- The CUB-200-2011 dataset and visual descriptions provide the fine-grained bird images and text used in the experiments.Source: §3 · p. 3
Determining Counter-Classes
- Counter-classes are chosen as the nearest different-class images in fc7 space, making them non-trivial because they generally share at least one attribute with the input.Source: §3 · p. 3
Baseline
- The baseline randomly selects an attribute from candidate counterfactual evidence rather than using a trained evidence checker.Source: §3 · p. 3
Metrics
- Phrase error measures whether counterfactual text appears in ground-truth image descriptions, while Accuracy w/Counterfactual Text examines sentence-based classification after adding that text.Source: §3 · p. 3
Results
- 16.26% phrase error for the random baseline; both CF models outperform it on phrase error and Accuracy w/FC Test.All three models reduce Accuracy w/FC Test from 54.38% without counterfactual text.Source: §3 · p. 3, Table 1 · p. 3
- The CUB-trained CF: Classifier performs worse than CF: Phrase-Critic; the authors hypothesize that external Visual Genome training and phrase localization help explain the difference.Source: §3 · p. 3
Qualitative Results
- Examples show counterfactual explanations identifying absent class cues, including a yellow nape for Bobolink and a black face for Common Yellowthroat.The latter cue helps distinguish a Common Yellowthroat from a Blue-Winged Warbler when a present-feature explanation could apply to both.Source: §3 · pp. 3–4, Figure 2 · p. 4