Ph.D. DISSERTATION DEFENSE - Zhenzhu Zheng
Knowledge Distillation for Food Image Classification on Mobile Devices
Abstract:
The use of mobile devices to analyze images and videos of food, both pre-and post-cooking, holds the promise of automated inference of nutritional information and meal preparation assistance through visual identification and portion measurement of single ingredients and mixed dishes. Performing such analysis robustly poses a number of challenges due to a large number of categories, high visual similarity between many different foods, and a relative lack of datasets for training state-of-the-art deep models. Furthermore, the best-performing neural networks usually have massive numbers of parameters that prevent them from being deployed directly on mobile devices with limited GPU resources. Finally, as new dishes/food are invented or encountered, the system should have a lifelong learning ability to adapt to a continuously changing environment, even when such examples of new items are scarce.
Deep Neural Networks are typically large and deep. How to extract knowledge from a large model to a small model becomes an interesting problem. One popular paradigm is Knowledge Distillation (KD), which is a technique to transfer knowledge from a Teacher (large) network to a Student (small) network. Most of the approaches are developed within the "Teacher-Student" framework.
We present a simple but effective approach for online knowledge distillation. Comprehensive experiments on Food-101, CIFAR10/100, and SVHN shows that it is possible to achieve competitive performance without introducing any extra model parameters. This yields an efficient and general solution, and easy for implementation. This work helps to enable more possibilities for on-device training.
Committee members:
Prof. Christopher Rasmussen (University of Delaware, Computer and Information Science)
Prof. Xi Peng (University of Delaware, Computer and Information Science)
Prof. Chandra Kambhamettu (University of Delaware, Computer and Information Science)
Prof. Liao Li (University of Delaware, Computer and Information Science)
Prof. Austin Brockmeier (University of Delaware, Electrical and Computer Engineering)
Prof. Brandon McFadden (University of Delaware, Applied Economics and Statistics)
Dial-In Information
Zoom link: https://udel.zoom.us/j/2305165881
Monday, November 30, 2020 at 2:00pm to 3:00pm
Virtual Event- Event Type
- Calendar
- Departments
- Website
- Group
- ENGR - Computer & Information Sciences
- Contact Email
- Contact Name
-
Kristin Halberg
- Contact Phone
-
3028312711
- Subscribe
Recent Activity
No recent activity