Reproduce “Subsidiary Prototype Alignment for Universal Domain Adaptation” 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
Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, R. Venkatesh Babu
Subsidiary Prototype Alignment (SPA) enhances Universal Domain Adaptation (UniDA) by integrating a Bag-of-visual-Words inspired architecture and a self-supervised pretext task at a mid-level network layer. The method achieves state-of-the-art H-scores across various UniDA and Open-Set DA benchmarks, effectively mitigating negative transfer.
Key insights
- An inherent tradeoff exists between negative-transfer-risk and domain-invariance-score across network layers, suggesting that adaptation at a mid-level layer is optimal for UniDA.
- Applying Bag-of-visual-Words concepts at an intermediate feature level can create more stable and universal representations for all categories, including unknown ones, robust to domain and category shifts.
- A grid-shuffling based self-supervised pretext task can implicitly enforce discriminative word-prototype learning and alignment by correlating word-histogram entropy with the number of contributing instances, thereby reducing word-level misalignment and negative transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.