Generative Multi-modal Models are Good Class-Incremental Learners

TL;DR

This work proposes a novel generative multi-modal model (GMM) framework for class-incremental learning and improves by at least 14% accuracy over all the current state-of-the-art methods with significantly less forgetting in the Few-shot CIL setting.

Abstract

In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic forgetting caused by the classifier's bias towards the current task has long posed a significant challenge. It is mainly caused by the characteristic of discriminative models. With the growing popularity of the generative multi-modal models, we would explore replacing discriminative models with generative ones for CIL. However, transitioning from discriminative to generative models requires addressing two key challenges. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, it requires formulating the task of CIL within a generative framework. To this end, we propose a novel generative multi-modal model (GMM) framework for class-incremental learning. Our approach directly generates labels for images using an adapted generative model. After obtaining the detailed text, we use a text encoder to extract text features and employ feature matching to determine the most similar label as the classification prediction. In the conventional CIL settings, we achieve significantly better results in long-sequence task scenarios. Under the Few-shot CIL setting, we have improved by at least 14\% accuracy over all the current state-of-the-art methods with significantly less forgetting. Our code is available at \url{https://github.com/DoubleClass/GMM}.

Figures (4)

• Figure 1: Illustration of conventional discriminative models for class-incremental learning (CIL) and our generative multi-modal models (GMM) for CIL. Discriminative models pose potential risk of classifier bias toward current task with network expasion. Our GMM framework consists generation and classification phases. And It is adapted to CIL based on the similarity of generated text and the true category names.
• Figure 2: The overview structure of our proposed method. The conceptual illustration of the generative multi-modal model (GMM) is shown on the left. In order to adapt this model for CIL, We have to turn the GMM model for classification and further adapt to our objective benchmark for learning (see Sec. \ref{['sec:core']}). On the right side, we demonstrate how the final evaluation is taken for CIL with all seen classes. Text encoder is used for obtaining the embeddings for similarity prediction.
• Figure 3: Comparison of our method with other SOTA baselines on CIFAR100 and Tiny-ImageNet under the conventional CIL setting.
• Figure 4: Visual comparison examples of our method against frozen MiniGPT-4 zhu2023minigpt (Zero-shot). Text with a gray background is generated by MiniGPT-4 based on the image, while text with a orange background represents our method's output. The ground-truth labels are displayed on the left side of each row of images. All images displayed here are random sampled from ImageNet-R.