Select Page

The paper “Attentive Explanations: Justifying Decisions and Pointing to the Evidence” by Dong Huk Park et al. introduces the Pointing and Justification (PJ-X) model, which aims to provide explanations for visual decision-making processes in a manner that mimics human justifications. This model integrates an attention mechanism to generate natural language explanations and highlight relevant visual evidence, addressing the opacity of deep learning models.

Introduction and Motivation

Humans naturally justify their decisions with language and visual cues. The authors aim to imbue deep learning models with similar capabilities. By providing both textual justifications and visual attention maps, PJ-X helps elucidate the decision-making process of AI systems, making them more transparent and interpretable.

Related Work

Previous works on explanations focus on textual outputs, often using template-based methods for specific domains such as medical applications and simulated training. Visual explanations have been less explored, typically aiming to identify discriminative visual patches or understanding neuron representations. The PJ-X model extends this by combining textual and visual justifications, offering a more comprehensive explanation mechanism.

Datasets

The authors introduce two new datasets for training and evaluating the PJ-X model:

VQA-X (Visual Question Answering Explanation): This dataset extends the VQA dataset by including natural language explanations for answers to questions about images. It contains 20,000 question-answer pairs from the VQA training set and 2,000 pairs from the validation set, each annotated with explanations.

ACT-X (Activity Explanation): Derived from the MPI Human Pose dataset, ACT-X focuses on human activities in images. It includes 17,019 images across 367 activities, with explanations for each activity, aiming to provide context for activity recognition.

Model Architecture

PJ-X is designed to jointly train the tasks of decision-making, justification, and evidence pointing. The model uses a double attention mechanism: one for answering questions and another for generating explanations. The attention mechanism allows the model to focus on relevant parts of the image, enabling it to provide visual justifications alongside textual ones.

Visual Question Answering

For VQA, the model predicts answers based on a combination of image features and question embeddings. The attention mechanism helps pinpoint areas of the image that provide evidence for the answer. The textual justification is generated by conditioning on the question, the predicted answer, and the image features.

Activity Recognition

The approach for activity recognition is similar, but without the question component. The model focuses on identifying relevant visual cues that justify the predicted activity.

Experiments and Results

The authors evaluate PJ-X on both VQA-X and ACT-X datasets, using metrics for textual justification (BLEU, METEOR, ROUGE, CIDEr, SPICE) and visual pointing (Weighted Overlap, Earth Mover’s Distance). The model demonstrates strong performance in generating high-quality explanations and attention maps that align with human annotations.

Textual Justification

PJ-X’s explanations are evaluated against human-provided ground truth using several metrics. The results show that integrating attention mechanisms significantly improves the quality of explanations compared to models without attention.

Visual Pointing

The visual pointing task assesses how well the model’s attention maps correspond to human-annotated evidence. PJ-X outperforms baseline methods, indicating its effectiveness in highlighting relevant image regions for both answers and explanations.

Summary

PJ-X represents a significant step towards explainable AI, offering a model that can justify its decisions with natural language and point to visual evidence. The creation of VQA-X and ACT-X datasets provides a valuable resource for further research in this area. Future work may focus on refining the attention mechanisms and extending the approach to other visual tasks.

For detailed information, please refer to the original paper: Attentive Explanations: Justifying Decisions and Pointing to the Evidence​.