Bridging the Theoretical Bound and Deep Algorithms for Open Set Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised Open Set Domain Adaptation (UOSDA) aims to classify data from a target domain containing unknown classes not present in the source domain.
• Existing methods struggle with highly flexible classifiers like deep neural networks (DNNs), leading to misclassification of known target data as unknown.
• A key challenge is minimizing the 'open set difference' term in risk upper bounds, which can become negative with DNNs, causing most target data to be incorrectly identified as unknown.

Purpose of the Study:

• To propose a new, more robust upper bound for target-domain risk in UOSDA that accommodates flexible classifiers like DNNs.
• To develop a principle-guided deep UOSDA method that effectively minimizes this new upper bound.
• To enhance the accurate recognition and classification of both known and unknown data in the target domain.

Main Methods:

• Introduced a novel upper bound for target-domain risk, incorporating source-domain risk, an 'ϵ-open set difference' (∆ϵ), distributional discrepancy, and a constant.
• Developed a principle-guided deep UOSDA method utilizing DNNs, minimizing the new upper bound.
• Employed gradient descent to minimize source-domain risk and ∆ϵ, and adversarial training to reduce distributional discrepancy.

Main Results:

• The proposed ∆ϵ is more robust than the traditional open set difference when minimized, preventing issues with DNNs.
• The principle-guided deep UOSDA method demonstrated state-of-the-art performance across various benchmark datasets.
• Achieved superior results in digit recognition (MNIST, SVHN, USPS), object recognition (Office-31, Office-Home), and face recognition (PIE).

Conclusions:

• The novel upper bound and principle-guided deep UOSDA method effectively address limitations of previous approaches.
• The method enables the successful application of flexible DNNs in UOSDA tasks.
• Achieved state-of-the-art performance, validating the proposed approach for robust domain adaptation with unknown classes.