Speaker:Azeez Idris

Deep learning models for supervised image classification tasks have achieved human-level performance due to large labeled image datasets and increased computing power. However, a major problem of training supervised deep learning image classifiers that persist is the limited availability of domain-specific labeled image datasets. This problem is particularly pressing in areas such as medicine due to the rarity of certain diseases and the cost and expertise needed for data annotation. Data augmentation, transfer learning, few-shot learning, and improved training and scaling strategies are some approaches proposed to solve this challenge. 

This thesis proposes a new training strategy to reduce model misclassification as follows. 1) We use weakly supervised learning to understand which classes are most confused with each other. 2) We introduce a new class containing synthetic training data highlighting the class confusion. We call this class the not-sure class. Random Cropping (RCP) and Equal Mixup (EQM) are two methods used for creating the synthetic training data for the not-sure class. 3) We train the model with the expanded training data, and then 4) We fine-tune the final model using transfer learning to better classify given inputs. 

We tested our new training strategy using open medical and non-medical datasets. The proposed training strategy improves classification accuracy, precision, recall, and f-1 score. We realize that the RCP method works better for medical datasets while the EQM works for most cases of non-medical datasets. We achieve model improvements of 0.6% to 5.6% for medical datasets and 3.9% to 4.4% for non-medical datasets across all metrics.

The ablation study reveals the impact of various components of our training strategy, including using data augmentation during training and using not-sure data randomly rather than following the not-sure algorithm. We find that data augmentation helps improve the overall performance of our training pipeline. In further improving the performance of the deep learning model, we find that generating not-sure data without using the not-sure algorithm provides the best model performance in some cases.

Committee: Ying Cai (major professor), Wallapak Tavanapong, and Wensheng Zhang

Join on Zoom: Please click this URL to start or join. https://iastate.zoom.us/j/92321382236 Or, go to https://iastate.zoom.us/join and enter meeting ID: 923 2138 2236