Reproduce “Tree-CNN: A Hierarchical Deep Convolutional Neural Network for Incremental Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Deboleena Roy, Priyadarshini Panda, Kaushik Roy

Researchers from Purdue University developed Tree-CNN, a hierarchical deep convolutional neural network that dynamically expands its tree structure to incorporate new classes over time. This approach significantly reduces catastrophic forgetting and training effort, achieving competitive accuracy, including over 10% higher accuracy than iCaRL on the CIFAR-100 dataset.

Key insights

  • A dynamically growing hierarchical DCNN structure effectively mitigates catastrophic forgetting in incremental learning tasks by localizing network modifications.
  • Reducing training effort by selectively retraining only affected DCNN nodes within a tree hierarchy can maintain high classification performance while being computationally efficient.
  • Feature similarity, as captured by DCNN outputs, can guide the self-organization of new classes into meaningful super-classes within a growing network, forming a feature-driven hierarchy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.