Learning from pairwise similarity for visual categorization

Transfer learning;Pairwise similarity;Clustering;Deep learning;Neural networks;Classification;Out-of-distribution detection;Instance segmentation;Lane detection

Hsu, Yen-Chang ; Kira, Zsolt Electrical and Computer Engineering Vela, Patricio Batra, Dhruv Hoffman, Judy Odom, Phillip ; Kira, Zsolt

University:Georgia Institute of Technology

Department:Electrical and Computer Engineering

关键词: Transfer learning;    Pairwise similarity;    Clustering;    Deep learning;    Neural networks;    Classification;    Out-of-distribution detection;    Instance segmentation;    Lane detection;   

Others  :  https://smartech.gatech.edu/bitstream/1853/62814/1/HSU-DISSERTATION-2020.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

Learning high-capacity machine learning models for perception, especially for high-dimensional inputs such as in computer vision, requires a large amount of human-annotated data. Many efforts have been made to construct such large-scale, annotated datasets. However, there are not many options for transferring knowledge from those datasets to other tasks with different categories, limiting the value of these efforts. While one common option for transfer is reusing a learned feature representation, other options for reusing supervision across tasks are generally not considered due to the tight association between labels and tasks. This thesis proposes to use an intermediate form of supervision, pairwise similarity, for enabling the transferability of supervision across different categorization tasks that have different sets of classes. We show that pairwise similarity, defined as whether two pieces of data have the same semantic meaning or not, is sufficient as the primary supervision for learning categorization problems such as clustering and classification. We investigate this idea by answering two transfer learning questions: how and when to transfer. We develop two loss functions for answering how to transfer and show the same framework can support supervised, unsupervised, and semi-supervised learning paradigms, demonstrating better performance over previous methods. This result makes discovering unseen categories in unlabeled data possible by transferring a learned pairwise similarity prediction function. Additionally, we provide a decomposed confidence strategy for answering when to transfer, achieving state-of-the-art results on out-of-distribution data detection. Lastly, we apply our loss function to the application of instance segmentation, demonstrating the scalability of our method in utilizing pairwise similarity within a real-world problem.