Capstone Projects.

Utilize Semi-Supervised Consistency Regularized Representational Learning, with strategic target labelling, to significantly reduce the need for fully labelled data in domains (e.g. medical) where labeling data is extremely costly or impractical.

SpeakUpAi is an end-to-end model for speech enhancement. Users will be able to upload spoken audio recorded on consumer devices (phones, tablets, etc.), and SpeakUpAI will scale up the fidelity until that audio sounds like it was professionally recorded. The SpeakUpAI project team uses deep learning solutions to break up their results into an evaluation and speaker set.

We propose a solution that expands the capacity of a question answering system to use an entire knowledge base or text corpus as its context. This is a two-step process. The first step is to identify the best context(s) for a given question. The second step is to run the question and the context(s) chosen from the first step through the more typical BERT question-answering task.

The quality of classroom relationships is one of the strongest predictors of students’ academic success. Our system visualizes engagement metrics for discussion-based classrooms. StudentStats delivers a dashboard that tracks individual students and quantifies their engagement. Post-lesson, the team provides descriptive statistics for each class period, along with automatic article and video recommendations related to the content of the given lesson.

This is a digital gallery that can be effortlessly customized to display artwork that is apt for that moment, occasion and mood. Imagine a digital art frame at home, in a hotel, at a restaurant or hospital ward, on an airplane powered by a neural network-based recommendation engine that personalizes it. A web app assists new users to quickly build galleries to display and existing users to further refine their existing galleries. (Using Neural Collaborative Filtering Networks).

Inspired by the recent advances in self-supervised learning methods, we apply the research work SimCLR from the Google Research Team to retinal OCT image dataset. With the help of self-supervised contrastive learning, our deep learning algorithm is able to learn useful representations of the data in the absence of labeled data, and can be further fine tuned to adapt downstream classification tasks. We evaluate the training sample size threshold at which our framework outperforms a comparable supervised-learning architecture.

Depth estimation is at the core of visual perception for autonomous vehicles. Stereo cameras are crucial sensors for self-driving vehicles as they are low-cost and can be used to estimate depth. The project aims to focus on the challenges of monocular depth by stereo thus overcoming low reliability when deployed, limited field of view, expenses, real-time estimation. This project aims to use KITTI data.