Dynamic content-addressable memory based on global centroid features for online task-free continual learning

Abstract

In online continual learning, a neural network learns from a continuous data stream in which each data point is processed only once and never revisited in the next learning stages. This non-revisiting mechanism often leads to Catastrophic Interference (CI), i.e., the model can dramatically lose the learned features of previous tasks when new information is learned. This forgetting issue becomes even more severe in online task-free continual learning, where the model has to adapt to the latest information autonomously without knowing task boundaries in advance. Furthermore, the inner sparse distributed memory used to store data points can become overburdened due to the rapid growth of new data points that need to be stored for upcoming tasks. To deal with these challenges, in this article, we propose an efficient model for online task-free continual learning built on the following two novel concepts. Firstly, we introduce a new powerful operator based on Global Centroid Features (GCF) to take advantage of the condensed knowledge of data points stored in dynamic content-addressable memory. Thanks to a suitable clustering mechanism, the GCF accumulation is able to control the storage overload issue caused by the rapid increase of newly stored data points in the upcoming tasks. Secondly, we design a continual learning model that effectively leverages the extraction of GCF-based information, enabling it to alleviate catastrophic forgetting through cognitive condensation of GCF. Thereby, the proposed GCF-based model simultaneously exploits the global centroid information of the stored data points and efficiently manages the rapid growth of stored data points in the storage space. The experimental results on benchmark image classification datasets emphasize the superior performance of our approach compared to state-of-the-art methods.

Keywords: Online task-free continual learning, Class-incremental learning, Clustered features, Image classification

Dataset

• Split CIFAR-10
• Split CIFAR-100
• Split CUB-200
• CORe-50

Feature extractor

• Reduced Resnet-18
• Resnet-18
• Resnet-50

Sample commands to run GCF

Dataset: Split CIFAR-10, Feature extractor: Reduced Resnet-18, Memory size: 1000

  
    python General_main.py --dataset cifar10 --backbone reduced --memory 1000 
  

Dataset: CORe-50, Feature extractor: Resnet-18, Memory size: 2000

  
    python General_main.py --dataset core50 --backbone resnet18 --memory 2000
  

Dataset: Split CIFAR-100, Feature extractor: Resnet-50, Step: 2

  
    python General_main.py --dataset cifar100 --backbone resnet50 --step 2
  

Dataset: Split CUB-200, Feature extractor: Resnet-50, Step: 5

  
    python General_main.py --dataset cub200 --backbone resnet50 --step 5
  

Case study: Split CIFAR-10, Feature extractor: Resnet-18, Step: 2

  
    python General_main.py --case_study True --dataset cifar10 --backbone resnet18 --step 2
  

Runtime with respect to the number of seen classes: Split CIFAR-100, Feature extractor: Resnet-50, Step: 2, Memory size: 3000

  
    python General_main.py --runtime True
  

Citation

If you use this code in your research, please cite the following relevant work:

  
    @article{tran2026dynamic,
      title={Dynamic content-addressable memory based on global centroid features for online task-free continual learning},
      author={Tran, Cong Tu and Nguyen, Thanh Tuan and Nguyen, Thanh Phuong and Thirion-Moreau, Nad{\`e}ge},
      journal={Machine Vision and Applications},
      volume={37},
      number={5},
      pages={117},
      year={2026},
      publisher={Springer}
    }