Transfer Neural Trees: Semi-Supervised Heterogeneous Domain Adaptation and Beyond.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Heterogeneous domain adaptation (HDA) is challenging due to dissimilar data domains and feature types.
• Existing methods often struggle to unify feature mapping, domain adaptation, and classification.
• Advancements in deep learning offer new possibilities for tackling HDA.

Purpose of the Study:

• To propose a novel deep learning model, Transfer Neural Trees (TNT), for effective HDA.
• To develop a unified architecture that jointly handles cross-domain feature mapping, adaptation, and classification.
• To extend TNT for semi-supervised HDA and zero-shot learning.

Main Methods:

• Developed Transfer Neural Trees (TNT), a deep learning architecture for HDA.
• Introduced Transfer Neural Decision Forest (Transfer-NDF) as the prediction layer for adaptation via stochastic pruning.
• Incorporated a unique embedding loss term for semi-supervised HDA, ensuring prediction and structural consistency.
• Extended TNT for zero-shot learning by associating image and attribute data.

Main Results:

• TNT demonstrated effectiveness in jointly solving feature mapping, adaptation, and classification within a unified framework.
• The embedding loss term successfully preserved consistency for semi-supervised HDA.
• The extended TNT achieved promising performance in zero-shot learning for image-attribute association.
• Experiments across diverse tasks, datasets, and modalities validated TNT's effectiveness.

Conclusions:

• Transfer Neural Trees (TNT) provide a powerful and unified approach to heterogeneous domain adaptation.
• The model's flexibility extends to semi-supervised and zero-shot learning scenarios.
• TNT offers a robust solution for cross-domain data association and classification across various data types and modalities.