Abstract
Zero-shot classification is a promising paradigm to solve an applicable problem when the training classes and test classes are disjoint. Achieving this usually needs experts to externalize their domain knowledge by manually specifying a class-attribute matrix to define which classes have which attributes. Designing a suitable class-attribute matrix is the key to the subsequent procedure, but this design process is tedious and trial-and-error with no guidance. This paper proposes a visual explainable active learning approach with its design and implementation called semantic navigator to solve the above problems. This approach promotes human-AI teaming with four actions (ask, explain, recommend, respond) in each interaction loop. The machine asks contrastive questions to guide humans in the thinking process of attributes. A novel visualization called semantic map explains the current status of the machine. Therefore analysts can better understand why the machine misclassifies objects. Moreover, the machine recommends the labels of classes for each attribute to ease the labeling burden. Finally, humans can steer the model by modifying the labels interactively, and the machine adjusts its recommendations. The visual explainable active learning approach improves humans' efficiency of building zero-shot classification models interactively, compared with the method without guidance. We justify our results with user studies using the standard benchmarks for zero-shot classification.
Introduction
Zero-Shot Classification (ZSC) is a powerful machine learning paradigm where a model learns to recognize objects it has never seen before. For example, a model might identify a "zebra" after being trained only on "horses" and the concept of "stripes." To achieve this, ZSC models rely on a class-attribute matrix—a structured map defining which semantic traits (attributes) belong to which classes.
Traditionally, creating this matrix is a significant bottleneck. Experts must manually specify dozens of attributes for hundreds of classes through a tedious process of trial and error. Often, there is no feedback on whether a chosen attribute (like "has ears") actually helps the model distinguish between classes, leading to a "semantic gap" where the machine's internal logic and human domain knowledge are misaligned.
The Semantic Navigator System
To address these challenges, we propose a Visual Explainable Active Learning approach implemented in a system called the Semantic Navigator. This system transforms the model-building process into a collaborative "human-AI teaming" exercise.
The interface provides a comprehensive view of the model's state. The Semantic Map (Figure 1b) is the centerpiece, projecting the machine's high-dimensional feature space and the human's semantic space into a shared Mutual Mental Space. Using t-SNE, it visualizes class clusters, "prototypes" (the attribute-based definitions), and "exemplars" (actual image data). This allows users to see exactly why a model might be failing—for instance, if the prototype for "Wolf" is incorrectly positioned within the "Dog" cluster.
The Interaction Loop
The system guides users through a four-step iterative loop: Ask, Explain, Recommend, and Respond.
- Ask: Instead of broad queries, the machine asks contrastive questions via the Hint View (e.g., "What is the most visible difference between a deer and a bobcat?"). This leverages human analogical reasoning, which is more efficient than absolute description.
- Explain: The Semantic Map explains the model's current logic. Users can observe Semantic Trajectories, tracking how class prototypes move toward their data clusters as new attributes are added, providing immediate visual confirmation of progress.
- Recommend: To reduce the labeling burden, the system uses a Semi-Supervised Support Vector Machine ($S^3VM$) to predict attribute labels for classes the user hasn't yet addressed.
- Respond: Users provide feedback by lasso-selecting classes on the map to confirm or correct the machine's recommendations, directly steering the underlying EXEM model.
Evaluation and Results
The system was evaluated using two standard ZSC benchmarks: AWA2 (Animals with Attributes 2) and SUB-CUB200 (fine-grained bird species). User studies demonstrated that the Semantic Navigator significantly improves both efficiency and accuracy:
- Efficiency Gains: Participants saved an average of 14.42 seconds per attribute compared to traditional manual matrix interfaces.
- Accuracy Boost: Models built with the guided system saw a significant performance increase, with test accuracy improving by 4.96% for AWA2 and 4.13% for SUB-CUB200 compared to unguided baselines.
- Qualitative Success: Users reported that the system made a "boring" task feel intuitive and prevented the "brain stuck" feeling common in manual data labeling.
Conclusion and Future Work
The Semantic Navigator bridges the semantic gap in zero-shot classification by providing a visual, explainable framework for active learning. By guiding users with contrastive questions and providing real-time visual feedback, the system makes model building more efficient and accurate. Future work will explore supporting relative attributes (e.g., "Animal A is furrier than Animal B") and improving scalability for datasets with thousands of classes.
Summary updated