[Paper Review] Generalized Category Discovery with Decoupled Prototypical Network

Wenbin An, Feng Tian+4 more|arXiv (Cornell University)|Nov 28, 2022

Text and Document Classification Technologies4 citations

TL;DR

This paper proposes Decoupled Prototypical Network (DPN), a novel framework for Generalized Category Discovery that decouples known and novel category learning through bipartite matching of prototypes, enabling explicit category-specific knowledge transfer and semantic-aware soft assignment to reduce pseudo-label noise. DPN achieves state-of-the-art performance across multiple benchmarks by improving feature discriminability and generalization on both known and novel categories.

ABSTRACT

Generalized Category Discovery (GCD) aims to recognize both known and novel categories from a set of unlabeled data, based on another dataset labeled with only known categories. Without considering differences between known and novel categories, current methods learn about them in a coupled manner, which can hurt model's generalization and discriminative ability. Furthermore, the coupled training approach prevents these models transferring category-specific knowledge explicitly from labeled data to unlabeled data, which can lose high-level semantic information and impair model performance. To mitigate above limitations, we present a novel model called Decoupled Prototypical Network (DPN). By formulating a bipartite matching problem for category prototypes, DPN can not only decouple known and novel categories to achieve different training targets effectively, but also align known categories in labeled and unlabeled data to transfer category-specific knowledge explicitly and capture high-level semantics. Furthermore, DPN can learn more discriminative features for both known and novel categories through our proposed Semantic-aware Prototypical Learning (SPL). Besides capturing meaningful semantic information, SPL can also alleviate the noise of hard pseudo labels through semantic-weighted soft assignment. Extensive experiments show that DPN outperforms state-of-the-art models by a large margin on all evaluation metrics across multiple benchmark datasets. Code and data are available at https://github.com/Lackel/DPN.

Motivation & Objective

• To address the limitation of current methods that couple known and novel category learning, leading to poor generalization and loss of category-specific knowledge.
• To explicitly transfer category-specific knowledge from labeled known categories to unlabeled data by aligning prototypes across datasets.
• To decouple known and novel categories in unlabeled data through prototype matching, enabling distinct training objectives for each.
• To improve feature discriminability and reduce noise from pseudo-labels via semantic-aware soft assignment.
• To enhance model robustness and performance across varying known category ratios and category estimation tasks.

Proposed method

• Formulate a bipartite matching problem between labeled and unlabeled category prototypes using the Hungarian algorithm to align known categories and identify novel ones.
• Use prototype alignment to enable explicit, category-specific knowledge transfer from labeled to unlabeled data without additional parameters.
• Introduce Semantic-aware Prototypical Learning (SPL) that assigns instances to prototypes using semantic similarity weights, reducing noise from hard pseudo-labels.
• Apply Exponential Moving Average (EMA) to iteratively update labeled prototypes, stabilizing representation learning and improving generalization.
• Leverage prototype-based soft assignment to capture high-level semantic information and enhance feature discriminability for both known and novel categories.
• Estimate the number of categories using the DAC algorithm, with DPN showing improved representation quality for better estimation.

Experimental results

Research questions

• RQ1 Can decoupling known and novel category learning improve model generalization and discriminative ability in generalized category discovery?
• RQ2 How can category-specific knowledge be transferred explicitly from labeled to unlabeled data without relying on implicit feature transfer?
• RQ3 To what extent does semantic-aware soft assignment reduce the impact of noisy pseudo-labels in unsupervised category discovery?
• RQ4 How effective is prototype-based bipartite matching in aligning known categories across labeled and unlabeled data?
• RQ5 How does the model perform under varying known category ratios and in estimating the true number of categories?

Key findings

• DPN outperforms state-of-the-art models across all evaluation metrics on CLINC, BANKING, and StackOverflow datasets, demonstrating superior generalization and robustness.
• The ablation study confirms that removing decoupling or soft assignment significantly degrades performance, proving their critical role in noise reduction and knowledge transfer.
• The model achieves lower category estimation error (8.7% on CLINC, 13.0% on BANKING, 10.0% on StackOverflow) than DAC, indicating better representation quality.
• t-SNE visualizations show that DPN learns more separable and discriminative feature embeddings compared to baseline methods.
• The alignment strategy effectively matches known category prototypes between labeled and unlabeled data, as shown in the heat map of prototype distances.
• The model maintains strong performance across different known category ratios (0.25, 0.50, 0.75), confirming its robustness to data distribution shifts.