Deep autoencoder based domain adaptation for transfer learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Transfer learning is crucial for leveraging knowledge across domains.
• Minimizing domain divergence is a key challenge in transfer learning.
• Domain adaptation enhances feature robustness for transfer learning.

Purpose of the Study:

• To present a novel transfer learning framework using marginal probability-based domain adaptation and deep autoencoders.
• To reduce distribution deviation between source and target domain features.
• To propose two variants for classification: D-TLDA-1 (linear regression) and D-TLDA-2 (softmax regression).

Main Methods:

• Employs marginal probability-based domain adaptation to align source and target domains.
• Utilizes a deep autoencoder for feature representation learning.
• Implements a supervised learning algorithm with encoding-decoding architecture.

Main Results:

• The proposed framework successfully adapts source and target domains by minimizing feature distribution deviation.
• Experimental results on ImageNet and 20_Newsgroups datasets demonstrate significant accuracy improvements.
• The D-TLDA framework outperforms prominent state-of-the-art machine learning and transfer learning methods.

Conclusions:

• The novel transfer learning framework effectively addresses domain divergence.
• Deep autoencoders combined with domain adaptation offer a powerful approach for cross-domain learning.
• The proposed D-TLDA variants show superior performance in classification tasks.