|
PROJECT
SUPERVISORS
| EECE 597 | Prof. C. Thrampoulidis - ECE | ||
| To Apply: | For information & availability of specific projects. | ||
| ID | Status | Name | |
| CT-1 | Available | Learning Imbalanced Datasets with Overparameterized DNNs | |
| Label-imbalanced and group-sensitive classification seeks to appropriately modify standard training algorithms to optimize relevant metrics such as balanced error and/or equal opportunity. In a recent work, we have shown that existing loss-adjustment techniques are not inefficient when very large DNNs are trained to zero training error. Instead, we proposed a new loss function, which we call the VS-loss. Thus far, we have shown that the VS-loss has superior performance in simple settings (eg linear and random feature models, small datasets such as MNIST). This project aims to empirically demonstrate the advantage of using the VS-loss by extensive experiments on large state-of-the-art data sets (eg CIFAR, Imagenet, etc) and DNNs. Very good knowledge of Python and prior experience in using ML packages such as PyTorch, Tensorflow are required. Please contact me for details and to set up a meeting. | |||
