Speaker:Azeez Idris

Training Deep Learning Models With Limited Labeled Data By Learning from Confusion

In deep learning, large labeled datasets and computing resources help improve performance. However, access to large labeled datasets in some domains, like medicine, can be difficult to gather. This work provides some approaches for training deep learning models for improved performance when limited data is available. The motivation is to understand the model confusion and improve model performance through understanding this confusion. 

We introduce a training framework that synthesizes new images based on confusion and train models using these confused images. The second use of confusion is in active learning, where we use confusion to select images to be annotated from the first epoch of training. Rather than spending multiple epochs for training, we use only one epoch based on model confusion.

Finally, our proposed approaches have provided experimental results that show their viability in both medical and non-medical domains.

Committee: Ying Cai (major professor), Sigurdur Olafsson, Soumik Sarkar, Wallapak Tavanapong, and Wensheng Zhang

Join on Zoom: Please click this URL to start or join. https://iastate.zoom.us/j/98179732980 
    Or, go to https://iastate.zoom.us/join and enter meeting ID: 981 7973 2980