Just Say the Name: Online Continual Learning with Category Names Only via Data Generation (2403.10853v3)

Abstract: Requiring extensive human supervision is often impractical for continual learning due to its cost, leading to the emergence of 'name-only continual learning' that only provides the name of new concepts (e.g., classes) without providing supervised samples. To address the task, recent approach uses web-scraped data but results in issues such as data imbalance, copyright, and privacy concerns. To overcome the limitations of both human supervision and webly supervision, we propose Generative name only Continual Learning (GenCL) using generative models for the name only continual learning. But na\"ive application of generative models results in limited diversity of generated data. So, we specifically propose a diverse prompt generation method, HIerarchical Recurrent Prompt Generation (HIRPG) as well as COmplexity-NAvigating eNsembler (CONAN) that selects samples with minimal overlap from multiple generative models. We empirically validate that the proposed GenCL outperforms prior arts, even a model trained with fully supervised data, in various tasks including image recognition and multi-modal visual reasoning. Data generated by GenCL is available at https://anonymous.4open.science/r/name-only-continual-E079.

• The paper introduces G-NoCL, a framework that leverages text-to-image generators and the DISCOBER ensembling method to create training data solely from category names.
• The methodology combines a prompt refiner, multiple generative models, and a complexity-aware ensembling technique to boost both data diversity and training challenge.
• Empirical results show that G-NoCL enhances online continual learning performance by reducing reliance on manual annotations and improving out-of-distribution generalization.

Integrating Text-to-Image Generative Models for Online Continual Learning

Introduction to Generative Name only Continual Learning (G-NoCL)

The landscape of online continual learning (CL) is evolving, prompted by the challenges in data acquisition and maintenance, such as data privacy concerns, cost, and the dynamic nature of real-world scenarios. The Generative Name only Continual Learning (G-NoCL) framework emerges as a novel approach addressing these challenges by integrating text-to-image (T2I) generative models with the continual learning process. G-NoCL leverages automatically generated images from textual descriptions (or category names) for training, eliminating the need for manual annotation or web-scraping efforts. This paper unpacks the workings of G-NoCL, showcasing its effectiveness through extensive evaluations and highlighting a pivotal innovation: the DIverSity and COmplexity enhancing ensemBlER (DISCOBER) for optimal data sampling.

The G-NoCL Framework Components

The G-NoCL framework comprises four critical components:

1. Prompt Refiner Module (ψ): Generates fine-grained, diverse text prompts from basic category names to enhance the variability and quality of generated images.
2. Generators (G): A set of T2I generative models that produce images based on the refined prompts, contributing to both intra- and inter-diversity of the training data.
3. Ensembler (∆): A novel technique, DISCOBER, ensembles the generated images based on data complexity to ensure a balanced and challenging training set.
4. Continual Learner (fθ): The main learning model trained on the dynamically generated and ensembled data, capable of online adaptation to new concepts without forgetting previous knowledge.

DISCOBER: Enhancing Diversity and Complexity

The DISCOBER method stands out as a groundbreaking approach within the G-NoCL framework. It assesses the complexity of images generated for each concept and optimizes the ensemble of training data to include a variety of easy and challenging examples. This ensures that the learner is not only exposed to a diverse set of images but also faces varying degrees of challenge, enhancing its learning capability and generalizability. The mechanism behind DISCOBER relies on the novel concept of Relative Mahalanobis Distance (RMD) to measure sample complexity, guiding the selective inclusion of samples in the training set.

Empirical Evaluations and Insights

Extensive experiments validate the G-NoCL framework’s effectiveness, demonstrating notable improvements in online continual learning benchmarks, spanning both in-distribution (ID) and out-of-distribution (OOD) evaluations. Comparative analyses highlight the superior performance of G-NoCL against approaches utilizing manually annotated or web-scraped data, especially in OOD generalization tasks. Such outcomes underline the potential of generative models in synthesizing diversified and complex training data, fostering model robustness and adaptability.

Theoretical and Practical Implications

From a theoretical standpoint, the G-NoCL framework and the DISCOBER method introduce an innovative perspective on leveraging generative AI for continual learning, expanding the horizons of online learning research. Practically, the ability to generate training data on-the-fly can significantly reduce the costs and logistical hurdles associated with manual data curation, promoting scalable and privacy-conscious AI solutions.

Future Directions

While G-NoCL marks a substantive advancement in CL, it also opens avenues for future research, particularly in multi-concept learning scenarios and other machine learning tasks beyond classification, like semantic segmentation. Furthermore, exploring the integration of G-NoCL with more advanced generative and LLMs could unlock even greater efficiencies and capabilities in continual learning systems.

Conclusion

The G-NoCL framework, empowered by the DISCOBER data ensembling technique, represents a significant step forward in online continual learning. By intelligently leveraging generative models for data generation and optimization, G-NoCL addresses several critical challenges in the domain, setting a new benchmark for future developments in continual learning research and applications.