Independent Study Defense – Suraj Acharya

On Thursday, September 6 in Cramer 221 at 1:30 pm, Suraj Acharya will defend his independent study “Sub-Clustering Fine Grained Images Classes using Pre-trained model and ensemble learning with voting.”

“In the process of cancer diagnosis, researchers get large classes of similar

malignant cells images, yet for different cancer types. In general, large classes with such property are called Fine Grained Images classes (FGIC’s). Members cells (images) of FGIC’s have very small intra-class differences and much larger inter-class differences. Hence, for better accuracy of detecting the correct type of cancer, the need arises for sub-clustering FGIC’s. Yet, two major challenges are present in the sub-clustering process, namely the close similarity of the FGIC’s member’s images, and the typical lack of rich training set of FGIC’s images. Thus, we need to utilize an already pre-trained deep learning (DL) model, intercepting feature vectors (FV’s) just before the DL’s output layer and using them to advance the training of another untrained machine learning classifiers (MLC’s). Such FV’s will be used to train, in parallel, the selected MLC’s to implement our FGIC’s sub-clustering model. Meanwhile, voting mechanism is utilized among the MLC’s to aid in the clustering accuracy of our sub-clustering model. Our approach

achieved a very promising improvement in images clustering accuracy of FGIC’s over the competitive peer classifiers that deal with FGIC’s.”