New Approaches to Combine Individual Training Losses for Machine Learning
Abstract
Machine learning (ML) has been instrumental for the recent advances in AI and big data analysis. The central task of an ML algorithm is to ÒtrainÓ a model using a large number of data, which entails seeking models that minimize certain performance metrics known as the losses. The objective of the proposed work is to deepen the understanding of the aggregate loss and its role in general ML algorithms. In particular, we will study new types of aggregate loss, applying them to different ML problems. We will study properties of general aggregate loss from the perspective of set functions, from which we aim to find new types of aggregate loss. ML algorithms with new aggregate losses. The Pl will develop new ML algorithms by combining the new aggregate losses with different types of individual loss or applied to different ML problems. In the proposed work, we will study the ATk SVM algorithm, which is from the ATk (aggregate) loss and the hinge (individual) loss, and develop new algorithms using new aggregate losses for more ML problems, including multi-label classification and pairwise learning. As an important yet often overlooked aspect in many ML algorithms, the aggregate loss is of significant theoretical and practical importance. The proposed work is the first step towards a comprehensive and systematic study of the aggregate loss and the role it plays in a wide range of ML algorithms. Focusing on devising new types of aggregate loss, developing efficient learning algorithms for their adaptation in various ML problems, and studying their theoretical properties, the proposed research will advance our understanding of existing learning algorithms from a new perspective, and contribute new principles and insights for the future development of new ML algorithms. As a ubiquitous component bridging the training data with the model to be learned, aggregate losses play an important role in essentially all ML algorithms. The proposed research will advance the Army and nationÕs knowledge and understanding of the fundamental ML principles and theories. With the improved robustness of ML algorithms, the outcome of this research will benefit intelligent systems that are based on ML algorithms and aim to facilitate decision making and situation awareness for the Army s mission. The new insights about general ML algorithms obtained from the perspective of the aggregate loss can also stimulate future research and help to keep the U.S. at the forefront in computing sciences. á
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810297
Entities
People
- Siwei Lyu
Organizations
- Army Contracting Command
- United States Army
- University at Buffalo