Improving Open Set Recognition via Visual Prompts Distilled from Common-Sense Knowledge

Seongyeop Kim, Hyung-Il Kim, Yong Man RoAAAI 2024 · DOI

TL;DR

This paper tackles the "overconfidence problem" in Open Set Recognition (OSR), where models confidently misclassify unknown objects. They introduce "visual prompts" – learned image-like patterns infused with common-sense knowledge distilled from large language models – to help visual models better distinguish known from unknown classes. This approach is model-agnostic and computationally efficient, showing significant improvements in OSR metrics.

What Problem Does This Paper Address?

Open Set Recognition (OSR) is crucial for real-world AI systems, as they inevitably encounter objects not seen during training. The major hurdle in OSR is the "overconfidence problem": models often assign high confidence scores to incorrect predictions for unknown objects. This happens because traditional visual recognition models lack the common-sense reasoning abilities that humans naturally possess, such as understanding that a fox and a dog are distinct, even if they share some visual features. Language models excel at this kind of reasoning, but integrating it into visual systems has been challenging. This paper aims to bridge this gap by injecting common-sense knowledge directly into the visual recognition process.

Methods

The core innovation is the creation of "visual prompts." These are small, learned image-like patterns that are added to input images. The process involves:

  1. Common-Sense Knowledge Distillation: Large Language Models (LLMs) are used to extract common-sense knowledge about known classes from sources like ConceptNet. This knowledge is then formulated into textual prompts.
  2. Visual Prompt Learning: For each known class, a unique visual prompt is trained. This prompt is designed to have the same spatial dimensions as the input image but only uses a border region to avoid obscuring image content. The training involves two key loss functions:
  3. Alignment Loss ($L_{Align}$): This contrastive loss encourages the visual features of an image and its corresponding common-sense textual prompt to be similar.
  4. Suppression Loss ($L_{Supp}$): This loss is applied when a mismatched visual prompt is used with an image. It aims to force the model's prediction distribution to be uniform, indicating uncertainty rather than overconfidence.
  5. Strategic Optimization: Label smoothing ($L_{LS}$) is also employed during training for matching scenarios to promote more generalized predictions. The final loss is a weighted sum of these components.
  6. Model-Agnostic Approach: The visual prompts are trained independently, and the visual recognition model's parameters are frozen during this process. This means the learned prompts can be applied to various pre-trained visual recognition architectures.

Datasets

The paper evaluates their method on several standard OSR benchmarks:

  • CIFAR10 (C10): A subset of 10 image classes, with 6 designated as known and 4 as unknown.
  • CIFAR+N (C+N): Uses 4 known classes from CIFAR10 and varying numbers of unknown classes from CIFAR100, creating a more complex scenario.
  • TinyImagenet (TI): A smaller version of ImageNet, with 20 known classes and 180 unknown classes.

Key Insights and Ideas

  • Common Sense as a Bridge: The paper highlights that common-sense knowledge, readily available in language models, is the missing piece for robust OSR in visual models.
  • Visual Prompts as Knowledge Carriers: By framing common-sense knowledge as learnable "visual prompts," the authors create a novel way to inject this information directly into the visual pipeline.
  • Suppressing Overconfidence is Key: The explicit "suppression loss" is crucial for teaching models to express uncertainty when faced with unfamiliar inputs, directly combating the overconfidence problem.
  • Model Agnosticism is a Strength: The ability to train prompts separately and apply them to existing models without retraining the entire network makes this approach highly practical and adaptable.

Key Results

The proposed method, "Visual Prompt," consistently outperforms baseline methods and other state-of-the-art OSR techniques across various metrics:

  • AUC: Significant improvements were observed, with the baseline model showing a +24.5% boost in AUC on CIFAR10 after applying the visual prompts.
  • F1 Score: The method effectively handles data imbalance, with an increase of +8.0% for the ARPL+CS model on CIFAR10.
  • OSCR: The approach enhances both open-set and closed-set recognition, demonstrating a +1.7% OSCR improvement for ARPL+CS on CIFAR10.
  • Computational Efficiency: The method is computationally light, focusing on training the visual prompts rather than the entire model, and achieves optimal performance with a prompt size of P=30.

Limitations and Future Work

  • Limitations: The paper notes that the effectiveness of the common-sense text prompts can depend on the chosen weight threshold for extraction from ConceptNet, requiring careful tuning. The performance also shows a slight decline when the prompt size exceeds a certain point (P=30).
  • Future Work: The authors suggest that their model-agnostic approach can be further explored with different visual recognition architectures. The framework's ability to integrate knowledge from various sources could also be expanded. The paper implies that further research into optimizing the distillation and application of common-sense knowledge could lead to even more robust OSR systems.

Summary updated

BibTeX
@inproceedings{kim2024improving,
  author = {Seongyeop Kim and Hyung-Il Kim and Yong Man Ro},
  title = {{Improving Open Set Recognition via Visual Prompts Distilled from Common-Sense Knowledge}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i3.28058},
}
Figure