Link to Code

At Northwestern University, I worked with four other students and delved into the realm of computer vision and privacy protection with the project CLOAK: Computer Learning for Obfuscating Automobile Knowledge. As the prevalence of online image sharing continues to surge, our focus was on safeguarding individuals’ privacy by addressing the potential risks posed by publicly available vehicle license plate information. By leveraging machine learning techniques, particularly convolutional neural networks and object detection models like YOLOv5, we aimed to automatically detect license plates in images and apply appropriate levels of blur to conceal sensitive information while preserving image utility. This endeavor not only emphasized the importance of balancing privacy and utility but also showcased the potential of machine learning in addressing real-world privacy concerns in the digital age.

Poster:

Paper:

Let’s tackle the problem!

• The world is dominated by data, especially textual data. Effectively summarizing this overwhelming amount of information quickly and accurately is a notable task.
• A common challenge is that certain summarization models perform well within specific domains, but their effectiveness degrades when dealing with a mix of different domains.
• As the volume and diversity of textual data increases constantly, the need for a versatile summarizing model is increasingly evident.

To tackle these challenges, our team built an innovative solution.

CRAFTing a Solution

Our contribution, CRAFT (Cross-domain Abstractive Summarization through Incremental Fine-Tuning), is a model built to overcome the barriers of domain dependency in summarization.

The idea behind CRAFT is optimizing a pre-existing language model to be more effective across domains. We hypothesized that models could be “trained” using a mixed collection of similar types of texts - effectively adapting its capabilities to better suit the styles and structures usual to these texts.

Using this approach, coupled with incremental fine-tuning processes, the CRAFT model enhanced its comprehension of different styles and authoring traits in the supplied data - leading to a better summarization irrespective of the domain.

Testing showed that CRAFT was capable of handling a wide range of textual content - from news articles to scientific papers, from conversational dialogues to legislative documents - maintaining its consistency in summarization quality. Our findings and developments were put together in a paper we called “CRAFT: CRoss-domain Abstractive Summarization through Incremental Fine-Tuning”. Upon drafting our initial version of the research, we were keen to receive feedback from the academic faculty, and look forward to refining it further.

Paper:

During my first year as part of Northwestern’s IEEE Student Chapter I helped create a web app that diagnoses brain tumors from MRI scans with a classification ML model.

The Problem

• Brain cancer is the 10th leading cause of death, over 250,000 die from brain tumors every year
• Despite decades of research in the field, medical diagnostic errors have been at 5%, and 40 million errors in diagnostic imaging are made annually
• Radiologist burnout hovers around 35%, and the average number of images requiring interpretation per minute has increased sevenfold from 1999 to 2009. Despite increases in radiologist staffing, radiologists in 2015 had to interpret images every three to four seconds; all of that can be attributed to poor diagnoses.

Machine assisted medical imaging hsa a large potential to save lives, as early diagnosis of tumors can improve cancer patient rates by up to 400%. Existing solutions are underutilized in hospitals due to factors including inaccessibility and cost.

Our Solution

sumMRI is a web platform that makes it simpler and easier for anyone to diagnose brain tumors from MRI scans with a classification ML model. Using two convolutional neural networks (CNNs), sumMRI can detect if a brain tumor is present in an MRI scan and where the specific brain tumor is located. This tool can help neurologists short-circuit human error while examining MRI scans.

AI Demo + Technology:

This is the part I created! As part of the team, I focused on building the ML Image Segmentation Model, which detects the location of a brain tumor.

Project Slides:

Website Demo:

Here’s a demo that other team members worked on of the program running on a website.

Here’s a video of the first time I got animating the model in OnShape to work. (Well, sort of…)

Back to the “The Zoomer” Main Article here.

The Class

At Northwestern, I attended the DTC class. I worked in a team of 4 to create a solution for a client. Nancy Cowles of non-profit organization Kids in Danger asked our group to design a device that would enhance a child’s tablet, facilitating remote learning (due to Covid-19) and also making tablets safer for children to use.

Our group learned to follow the design process, following the main steps of of researching, ideating, prototyping, testing, iteration, and presentation. We came up with multiple mockups, and through further user testing/observation with families of preschool children and a design review with our classmates, we chose one design that most effectively fit our requirements, and we called it the Zoomer.

The Zoomer

The Zoomer is a tablet case and remote control that aims to make video conferencing applications, specifically Zoom, more intuitive and easier to use. The case consists of three parts: a protective case, a kickstand, and a remote control. The Zoomer helps meet the needs of our users by making the tablet more accessible, making participation easier, and making the tablet safer for a child to use. Even though the Zoomer is primarily designed for use with Zoom, which our research shows is nearly ubiquitous at the moment, it can be easily adapted to other software and hardware as needed, and it’s design is multipurpose, not limited to just be used by children.

I led the creation of our team’s design mockup, using OnShape to create a 3D model of what our tablet would look like. I also learned how to use mate connectors to animate our 3d model, revolving our prototype’s kickstand.

Initial Prototype

Updated Prototype

Updated Prototype 2 - (Design Freeze Document)

Final Prototype

Final Prototype Hinge Animation

Link to College Rankings Visualized:

Link to College Enrollment Visualized:

CART = encapsulates 2 types of trees: Classification Trees, Regression Trees

Ok so a lot of the time if you see a decision tree thing looking like this. (This is an example with 2 independent variables and the machine predicting 1 variable) The diagram will be split!

	Splits are determined by asking “does this split of data increase the amount of information we have about these points” and will stop once it cannot add more information from splitting.

How is the machine going to predict using a Decision Tree?

Let’s say the next point is x=30 y=30. It’d fall in the middle-bottom area (box), terminal leaf, of that decision tree. In each terminal leaf, the points have an average value.

R-Squared

Evaluates how good regression is going using Sum of Squares.

R^2=1-(SS/TSS)

(used in regression) Sum of Squares → Y1 - Y(predicted line)

Total Sum of Squares → Y1 - Y(average)