CS Events
PhD DefenseDeep Learning with Limited Data |
|
||
Thursday, March 16, 2023, 03:00pm - 05:00pm |
|||
Abstract: The field of deep learning has made significant strides in recent years. With the help of neural network architectures, deep learning has made significant advancements in many fields by autonomously identifying patterns in raw data through learning from large-scale datasets. One reason for deep learning’s success is its over-parameterization, which allows it to be optimized to global optimality using local search heuristics. And there is a growing trend in designing larger neural network models with bigger datasets. Unfortunately, the cost of training these models scales with the product of the number of parameters and the number of data points, which can quickly become a burden. Additionally, big data with annotations is not always readily available. To address this, our research aims to tackle the challenge of learning with limited data by exploring various solutions, including transfer learning from large datasets and improving data efficiency on the existing training datasets. We begin with proposing a method based on transferable reward for query object localization to enable test-time policy adaptation in new environments with limited labeled data. We then proceed on to propose a solution to the problem of limited data for one modality in a multi-modal training set. Specifically, we introduce a multi-modal model for new product recommendations. By combining the static inherent properties of items with the limited customer engagement data (e.g., views, purchases, reviews, and likes), this approach enables efficient learning and recommendation of new products. Next, we shift our focus to real-world applications in fiber optic sensing, specifically vehicle run-off-road and manhole open/close events detection. Despite small training set sizes due to data collection constraints, we demonstrate how detecting these events can be achieved by investigating the underlying structure characterized by transformations and temporal relations, rather than simple pattern recognition. Our solutions, approaches, and findings offer valuable insights into the potential of learning from limited data. As such, these insights can play an important role in developing robust solutions for deep learning applications.
Speaker: Tingfeng Li
Location : CBIM 22
Committee:
Professor Dimitris Metaxas (Chair)
Professor HaoWang
Professor Konstantinos Michmizos
Professor Sharon Xiaolei Huang (Pennsylvania State University)
Event Type: PhD Defense
Abstract: See above.
Organization:
Rutgers University
School of Arts & Sciences
Department of Computer Science
Contact Professor Dimitris Metaxas