Thursday 13 March 2025
Deep learning models have made tremendous progress in recent years, but they still struggle with one major challenge: how to learn new tasks while retaining knowledge of previously learned ones. This problem is known as incremental learning (IL), and it’s a crucial issue for many applications, from self-driving cars to medical diagnosis.
One approach to IL is to use replay-based methods, which store samples from previous tasks in memory and periodically review them to prevent forgetting. Another approach is to use regularization-based methods, which modify the model’s optimization objective to encourage retention of old knowledge. However, these methods often require significant modifications to the model architecture or training procedure.
A new paper presents a novel weakly-supervised learning paradigm called Incremental Partial Label Learning (IPLL), which tackles IL in a more efficient and flexible way. IPLL assumes that sequentially arriving data relates to a set of candidate labels rather than the true label, making it easier to handle ambiguous annotations.
The proposed method, called Prototype-Guided Disambiguation and Replay Algorithm (PGDR), leverages class prototypes as a proxy to mitigate two intertwined challenges in IPLL: label ambiguity and catastrophic forgetting. PGDR uses a momentum-based pseudo-labeling algorithm along with prototype-guided initialization, resulting in a balanced perception of classes.
In the training phase, PGDR first generates a set of candidate labels for each sample, which are then used to calculate pseudo-labels. The model is trained to predict these pseudo-labels, and the loss function combines three terms: a classification error term, a knowledge distillation term, and a consistency regularization term.
To alleviate forgetting, PGDR incorporates a memory replay technique that collects well-disambiguated samples while maintaining representativeness and diversity. By jointly distilling knowledge from curated memory data, PGDR achieves superior disambiguation ability for new tasks and exhibits less forgetting of old knowledge.
Experiments on several datasets, including CIFAR100 and Tiny-ImageNet, demonstrate the effectiveness of PGDR in balancing category perception and preventing catastrophic forgetting. The method outperforms baseline models with a significant margin, especially when dealing with ambiguous annotations or limited training data.
PGDR’s advantages lie in its simplicity and flexibility: it can be easily integrated into existing deep learning architectures and trained using standard optimization algorithms. This makes it an attractive solution for many applications where IL is crucial, such as autonomous vehicles, medical diagnosis, and personalized recommendations.
Cite this article: “Prototype-Guided Disambiguation and Replay Algorithm for Incremental Learning”, The Science Archive, 2025.
Incremental Learning, Deep Learning, Prototype-Guided Disambiguation, Replay Algorithm, Weakly-Supervised Learning, Partial Label Learning, Catastrophic Forgetting, Momentum-Based Pseudo-Labeling, Knowledge Distillation, Consistency Regularization.
