Tree-CNN: A hierarchical Deep Convolutional Neural Network for incremental learning.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep Convolutional Neural Networks (DCNNs) excel in computer vision but struggle with evolving datasets due to catastrophic forgetting and hyper-parameter sensitivity.
• Traditional DCNN models are static after training, posing challenges for continuous learning and adaptation to new information.
• The need for adaptive deep learning models is critical in a world with constantly changing data.

Purpose of the Study:

• To propose a novel adaptive hierarchical network structure composed of DCNNs that can grow and learn incrementally.
• To enable deep learning models to accommodate new data classes without losing previously learned information.
• To improve upon existing hierarchical Convolutional Neural Network (CNN) models by incorporating self-growth capabilities.

Main Methods:

• Developed an adaptive hierarchical network structure utilizing DCNNs.
• Implemented a tree-like growth mechanism to accommodate new data classes.
• Organized incrementally available data into feature-driven super-classes.
• Compared the proposed model against fine-tuning a deep network.

Main Results:

• The proposed hierarchical model demonstrated self-growth capabilities, adapting to new data.
• The network successfully preserved the ability to distinguish previously trained classes while learning new ones.
• Significant reduction in training effort was achieved compared to fine-tuning a deep network.
• Competitive accuracy was maintained on CIFAR-10 and CIFAR-100 datasets.

Conclusions:

• The adaptive hierarchical network offers an effective solution for incrementally learning DCNNs.
• This approach mitigates catastrophic forgetting and reduces computational overhead in evolving data scenarios.
• The self-growing hierarchical structure provides a scalable and efficient method for continuous learning in computer vision.