Core Concepts
A novel framework, I2CANSAY, that enhances the capability of learning from data streams while eliminating the dependence on memory buffers.
Abstract
The paper proposes a novel framework, I2CANSAY, for Online Task-Free Continual Learning (OTFCL). OTFCL is a more challenging variant of continual learning that emphasizes the gradual shift of task boundaries and learns in an online mode.
The key components of the I2CANSAY framework are:
Inter-Class Analogical Augmentation (ICAN) module:
Generates diverse pseudo-features for old classes based on the inter-class analogy of feature distributions for different new classes.
Serves as a substitute for the memory buffer, eliminating the need for storing old samples.
Intra-Class Significance Analysis (ISAY) module:
Analyzes the significance of attributes for each class via its distribution standard deviation.
Generates an importance vector as a correction bias for the linear classifier, enhancing the capability of learning from new samples.
The framework effectively prevents catastrophic forgetting and exhibits strong learning capabilities when dealing with online data streams, without relying on memory buffers.
Comprehensive experiments on CIFAR-10, CIFAR-100, CoRe50, and CUB-200 datasets demonstrate that the proposed I2CANSAY framework achieves state-of-the-art performance in three protocols: Online Task-Free Continual Learning, Online Continual Learning, and Offline Task-Free Continual Learning.
Stats
The paper does not provide specific numerical data or statistics to support the key logics. The results are presented in the form of comparative accuracy percentages across different datasets and experimental settings.