Yu-Ting Chen

Google (United States)
3 previous affiliations
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Yu-Ting Chen focuses on computer vision, domain adaptation for semantic segmentation, and language modeling optimization. Early work addressed geometric image reconstruction and panoramic stitching for planar scenes. More recently, the focus has shifted toward unsupervised learning techniques for cross-domain visual recognition and context-aware scaling methods in natural language processing.

Research Trajectory

The available publication record for Yu-Ting Chen spans from the mid-2000s to 2020, reflecting an evolution from classical geometric computer vision to modern deep learning applications in both vision and natural language processing.

Initial research focused on image integration and geometric reconstruction. In 2004, Chen co-authored Stitching and Reconstruction of Linear-Pushbroom Panoramic Images for Planar Scenes, which proposed a real-time method for integrating multiple linear-pushbroom panoramic images. This work was particularly applicable to planar environments, such as large-scale paintings or aerial and satellite imagery. The approach divided the integration process into stitching and Euclidean reconstruction, demonstrating that under specific configurations—such as column-to-column image correspondence—the number of required non-collinear corresponding points could be reduced, optimizing the computational requirements for panoramic reconstruction.

Following a gap in the observed publication record, Chen's trajectory resumed in 2017 with a shift toward deep learning, specifically addressing the vulnerabilities of semantic segmentation models when deployed in novel environments. Chen participated as a middle author in No More Discrimination: Cross City Adaptation of Road Scene Segmenters. This collaborative work tackled dataset bias in urban scene understanding, contributing to the broader area of Geographic Domain Adaptation for Semantic Segmentation. Rather than relying on expensive manual annotations for every new city, the research team developed an unsupervised learning approach to adapt pre-trained segmenters. By leveraging time-lapse imagery from Google Street View to extract static-object priors, the work aligned with the research direction of Learning Invariant Representations from Temporal Revisits, demonstrating how temporal data can help models overcome visual noise and domain shifts.

By 2020, Chen's collaborative efforts expanded beyond computer vision into natural language processing. As a middle author on Context-Aware Temperature for Language Modeling, Chen contributed to improving the calibration and performance of language models. The research challenged the standard practice of using fixed or manually crafted schedules for temperature scaling. Instead, the team proposed a generalized, context-aware approach that allowed individual optimal temperature trajectories to be learned dynamically during training, resulting in measurable improvements in perplexity across standard benchmarks like the Penn Treebank and WikiText-2.

Current Directions

Based on the most recent papers in the evidence, Chen's latest collaborative work investigates optimization techniques for deep neural networks, specifically in the domain of natural language processing. The 2020 research highlights an interest in dynamic, learnable parameters—such as context-aware temperature scaling—that adapt to specific inputs rather than relying on global heuristics. This suggests a broader focus on improving the efficiency, calibration, and contextual sensitivity of large-scale sequence models, allowing internal model parameters to adjust dynamically during training to better capture the underlying distribution of the data.

Profile updated

See something wrong? Report a mistake