1. What is catastrophic forgetting in class-incremental learning?
Catastrophic forgetting refers to the severe degradation of performance on previously learned classes when a deep neural network model is trained on new data or classes. In class-incremental learning, models are trained sequentially on new incoming data, and they must be incrementally updated using a limited number of new classes at a time. This paradigm is challenging because models tend to forget previously learned information when exposed to new data. To mitigate catastrophic forgetting, researchers have proposed using rehearsal techniques, such as storing samples from each previously encountered class in a memory. However, the size of this memory is often limited due to storage or privacy concerns. Some authors have suggested leveraging additional external training data from large curated datasets, like ImageNet, to work around this constraint. In this work, the authors propose using the generative model Stable Diffusion to generate synthetic samples belonging to the same classes as the ones previously encountered by the model. This allows for the use of these samples in both knowledge distillation and replay in the classification loss, resulting in improved general class-incremental learning methods on large-scale datasets.
read more
2. What methods mitigate bias in class-incremental learning?
Several methods have been proposed to mitigate bias in class-incremental learning (CIL). These include Nearest-Mean-of-Exemplars (NME) classifier, cosine classifier, bias correction layer, finetuning on a balanced subset, and specific losses. Additionally, model inversion and training a generative model in parallel to generate samples from past classes have been explored. These methods aim to overcome the limitations of memory size and catastrophic forgetting. Furthermore, approaches like Global distillation (GD) and Deep Model Consolidation (DMC) utilize external data to enhance the learning process. Synthetic data generation using GANs and pretrained generative models like Stable Diffusion have also been proposed to improve CIL performance. These methods help preserve previous knowledge and improve the model's ability to learn new classes effectively.
read more
3. What is the structure of the training procedure in the proposed method?
The training procedure consists of T + 1 steps, starting with the base step and followed by T incremental steps. Each step uses a training dataset D t with images from new and previously unseen classes Y t. The model is trained on D t M, where M is the replay memory containing samples from previously encountered classes. After each step, the model is evaluated on the test set of all learned classes without access to step or task descriptors. Additionally, the model has access to an external data source S, which can be an online or offline data stream, during every step of training.
read more
4. How does WordNet assist in generating prompts for class names and descriptions?
WordNet, a lexical database, provides lemmas of synsets as class names 'c' and definitions of synsets as descriptions 'd c'. For datasets like ImageNet, the WordNet association is included, while for others like CI-FAR100, it can be done semiautomatically. This helps in generating accurate prompts for class names and descriptions, reducing semantic errors and improving the quality of synthetic samples. WordNet's role is crucial in ensuring that the generated prompts accurately represent the intended class, especially in cases where homographs may cause confusion. By leveraging WordNet, researchers can automatically generate prompts without extensive manual engineering, enhancing the efficiency of the synthetic data generation process.
read more