Source-linked AI summary
Generating Visual Explanations
Lisa Anne Hendricks, Zeynep Akata, Marcus Rohrbach, Jeff Donahue, Bernt Schiele, Trevor Darrell
TL;DR
Deep visual recognition systems often lack justification text, while captioning models may omit class-discriminative evidence. The paper jointly predicts classes and generates image-grounded explanations using a sampled-sentence discriminative loss, achieving more discriminative and higher-quality explanations on fine-grained bird classification.
Problem
Deep recognition methods generally provide no justification text, while image descriptions may fail to mention class-discriminative features needed to justify predictions.
Method
The model jointly predicts a class and generates language conditioned on image features and the predicted label, using a reinforcement-learning loss over sampled sentences to enforce class specificity.
Results
The model generates explanations that are image relevant and class relevant, outperforming description-only baselines in discriminativeness and traditional sentence-generation metrics.
Takeaways & Limitations
Natural-language explanations can incorporate classification information to produce more convincing justifications than captioning systems.
Takeaways & Limitations
The evaluation uses CUB, whose images and sentences are associated with a single bird-class label, while its sentences were not collected specifically for visual explanation.
Abstract
from arXiv · showhide
Clearly explaining a rationale for a classification decision to an end-user can be as important as the decision itself. Existing approaches for deep visual recognition are generally opaque and do not output any justification text; contemporary vision-language models can describe image content but fail to take into account class-discriminative image aspects which justify visual predictions. We propose a new model that focuses on the discriminating properties of the visible object, jointly predicts a class label, and explains why the predicted label is appropriate for the image. We propose a novel loss function based on sampling and reinforcement learning that learns to generate sentences that realize a global sentence property, such as class specificity. Our results on a fine-grained bird species classification dataset show that our model is able to generate explanations which are not only consistent with an image but also more discriminative than descriptions produced by existing captioning methods.
1 Introduction
The paper defines visual explanations as image-grounded justifications that identify class-discriminative features. Its model jointly classifies images and generates class-specific language optimized for discriminative content.
- Visual explanations must be both class discriminative and accurate for a specific image instance.
- Unlike image descriptions or class definitions, explanations connect image evidence to why a predicted category is appropriate.
- The model conditions language generation on both image features and a predicted class label to produce class-specific sentences.
- A sampled-sentence reinforcement-learning loss specifically encourages global class-specificity rather than only alignment with ground-truth sentences.
- The proposed formulation produces more discriminative explanations than description-only baselines while improving traditional sentence-generation metrics.
2 Related Work
Prior work includes explanation systems, visual description models, discriminative-feature methods, and fine-grained classification. The paper distinguishes its approach by linking class-discriminative visual evidence to natural-language explanations through a global sentence constraint.
- Earlier explanation systems span applications such as medical diagnosis, simulation, and robotics, often using rules or predetermined templates.
- Computer-vision methods can discover features supporting classification decisions but generally do not connect those features to natural-language expressions.
- Modern image-description systems generate fluent, accurate image captions but are not designed to ensure class-discriminative explanations.
- Standard LSTM captioning commonly uses cross-entropy over predicted and ground-truth words, whereas this work specifies a global constraint on generated sentences.
- Fine-grained explanation requires focusing on aspects that are both class-specific and depicted in the image.
- Attribute-based approaches provide interpretable decision statements but require costly fine-grained expert annotation.
- The paper incorporates reinforcement-learning techniques for backpropagation through sentence or attention sampling, unlike related captioning work using standard cross-entropy.
3 Visual Explanation Model
The model generates image-grounded explanations that also emphasize class-discriminative properties. It combines relevance training with a sampled-sentence reinforcement-learning objective to favor discriminative descriptions.
- Visual Explanation Model: The explanation model requires sentences to describe the specific image and explain why it belongs to the predicted category.It uses relevance and discriminative losses to enforce these two requirements.
- Image-Relevant Generation: Category predictions are provided alongside image features to the sentence-generation LSTM, helping generate class-specific words and attributes.This extends the LRCN framework with category information as an additional input.
- Relevance Loss: The relevance loss trains the model to generate sentences corresponding to image content from ground-truth word sequences.Each training instance contains an image, category label, and ground-truth sentence.
- Discriminative Loss: The overall objective combines relevance loss with the expected negative discriminator reward over sentences sampled from the model.Because the expectation is intractable, training estimates it with Monte Carlo sampling from timestep softmax distributions.
- Discriminative Loss: The discriminative loss samples a sentence, scores its probability of indicating the ground-truth category, and increases the likelihood of highly rewarded descriptions.The reward is RD(˜w) = p(C|˜w), and REINFORCE provides the expected-reward gradient.
4 Experimental Setup
The experiments evaluate explanation quality on the CUB bird dataset using automatic image- and class-relevance metrics, ablations, and expert human judgments. The explanation model is compared with description, definition, and component-removal baselines.
- Dataset: The CUB dataset contains 200 North American bird classes, 11,788 images, and five sentences collected for each image.The sentences include detailed visual attributes such as beak shape, feather color, and face patches.
- Dataset: CUB is suited to visual explanation because every image and its sentences are associated with a single class label.The collected sentences were not originally written to explain why an image belongs to its class.
- Baselines and Ablations: The baselines include an image-conditioned description model, a label-only definition model, and ablations removing discriminative loss or class conditioning.These comparisons isolate the contributions of category conditioning and discriminative training.
- Metrics: METEOR and CIDEr measure image relevance, while class similarity and class rank measure class relevance.Higher is better for METEOR, CIDEr, and class similarity; lower is better for class rank.
- Metrics: CIDEr emphasizes distinctive class attributes through TF-IDF-weighted n-grams, so phrases such as “red eyes” receive more weight than common attributes.Class rank additionally compares generated sentences against references from the 200 bird classes.
- Human Evaluation: Experienced bird watchers rank which generated sentence best explains each classification, and the explanation model performs best on all reported metrics.Expert judgment is used because relevant bird-classification features may not be obvious to average readers.
5 Results
The model generates explanations that are both image-relevant and class-relevant, outperforming baselines and ablations across quantitative and qualitative evaluations. Visual comparisons show that its discriminative loss and class conditioning produce more specific attributes while retaining image grounding.
- The explanation model produces sentences that are both image relevant and class relevant, satisfying the paper’s definition of visual explanations.
- 5.1 Quantitative Results: The model achieves higher METEOR and CIDEr scores than the baselines and explanation-label and explanation-discriminative ablations.The reported comparison attributes better sentence quality to both label conditioning and the discriminative loss.
- 5.1 Quantitative Results: The explanation model obtains the highest sentence-classifier accuracy at 59.13%, compared with 22.32% for the description model.
- 5.1 Quantitative Results: Bird experts give the explanation model the best mean rank, while label-conditioned models generally receive lower ranks.Experts evaluated 91 randomly selected images, and lower rank indicates a preferred explanation.
- 5.2 Qualitative Results: Qualitative examples show the model mentioning fine-grained class attributes such as “black cheek patch,” “red eyes,” “black head,” “long neck,” and “orange beak.”The examples are presented as visually correct and more class-specific than outputs from comparison models.
- 5.2 Qualitative Results: Unlike definitions, explanations adjust to image evidence, avoiding attributes that are absent from a particular image instance.For example, the explanation model adapts across Downy Woodpecker images while the definition remains constant for the class.
- 5.2 Qualitative Results: The discriminative loss increases class-specific properties in generated sentences even when the model receives no class label at test time.Examples include “black head” for black-capped vireo and “long neck” and “orange beak” for white pelican.
6 Conclusion
The framework advances visual explanation by combining image descriptions with class-discriminative information. Its evaluations indicate that the model can produce convincing explanations beyond description-only captioning systems.
- The approach addresses a need for systems that can explain deep visual decisions, which may support adoption of sophisticated AI systems.
- The explanation-discriminative model captures more class-specific attributes than the description model while accurately describing visual attributes.
- Explanations change when generation is conditioned on different classes, with categories such as painted bunting strongly influencing the text.
- The framework provides explanations of a visual classifier using natural language.
- The conclusion reports quantitative and qualitative evidence for the proposed model and the effectiveness of its novel loss function.