Ivana Milovanovic

I developed PerfectShot, an interactive web-app that uses AI to help the users select the best photo from a bunch they took.

These days we all take tons of photos. With high quality digital cameras on our phones and endless cloud storage, we rarely ever take a single shot of anything anymore. But, every time we do that, we end up having multiple, redundant shots and going through photos and choosing which to keep is a time-consuming, tedious process. What if we could use ML to automatically select your best photos from the bunch you took?

I have developed a SVM ranker which uses both general image quality and content quality features to find the best photo. The app was deployed on AWS using Python, Flask, HTML, and CSS.

I built a multi-label model that can automatically tag research papers with several appropriate labels, e.g., math, physics, finance, and so on. In multi-label classification, classes are not mutually exclusive. Every data set sample can be assigned to multiple labels. For example, one physics research paper may contain advanced mathematical or statistical analysis, therefore it should be labeled with physics, math, and stat labels.

This problem is a supervised text classification problem, and I have explored several different classifiers and problem transformation methods to see which one works best for the given scenario.

I've implemented a feature extraction algorithm for NLP that uses TF-IDF of word n-grams, the appearance of words in a title, and so on. I have explored problem transformations such as binary relevance, classifier chain,s and label powerset.

I have explored the possibility of using neural activity from a single brain hemisphere to control both sides of the body in a brain-computer interface paradigm.

This is incredibly important for people who've had a stroke, where one side of the brain is partially or fully damaged. To answer this question, I analyzed terabytes of neural activity data to provide insight into the brain's function. I then used advanced statistical methods to model this activity and relate it to behavioral data.