1 of 8

1

Michael Vanden Heuvel

Department of Materials Science and Engineering �University of Wisconsin-Madison

Workshop: Enhancing Undergraduate Research with Machine Learning

10/22/2022

Impact of Machine Learning Focused Research Projects on My Career Path: An Undergraduate Experience

2 of 8

2

Motivation: Properties of metallic glasses

Introduction

University of Wisconsin – Madison

Graduating December 2022
Computer Science, B.S.

Skunkworks

Education Group – Fall ‘19
Pancreatic Cysts – Spring ‘20 – Fall ‘20
Renal AMLs – Spring 22 - Present

Microsoft

Intern – Summer 2022
Fulltime – March 2023

3 of 8

3

Motivation: Properties of metallic glasses

Project Overview

Pancreatic Cysts

Nonmalignant

Malignant

Nonmucinous

Mucinous

Been a part of a few projects
The project I will be focusing on today, is the Project I spent the most time on
The motivation for this project was that pancreatic cysts may develop into pancreatic cancer, which has one of the lowest survivability rates among cancers.
However, earlier diagnosis and treatment of cysts results in substantially better patient outcomes. Surgery is one of the primary treatments. But, due to the general risks involved with surgery, only patients who are at serious risk should be recommended for surgery
Multiple categories: mucinous and Nonmucinous. The mucinous cysts can then be classified as either malignant, which are the dangerous ones we care most about, or non malignant
So, if we could develop a model that can discriminate between the harmful and not as harmful cysts, it could be extremely valuable to healthcare professionals

4 of 8

4

Motivation: Properties of metallic glasses

Project Structure

Dr. Mingren Shen

Dr. Meghan Lubner

Slides from weekly progress reports

Skunkworks all-hands meeting fall 2019

Skunkworks all-hands meeting Spring 2021

A little bit about the structure of our group,
I worked on most of the day-to-day tasks with the other undergraduate students
Grad student mentor, Mingren, and Dr. Morgan guided us on the Machine Learning side of things

Give us a topic or two to research and apply to the project

We also drew on the expertise of Dr. Megan Lubner, who was our domain expert for the medical side of the project

An example of something that Dr. Lubner would help us with would be giving us insights into whether the things that our model was picking up on were medically sound

As a full group, we had weekly progress meetings where the other undergraduates and I would create a slide deck and update the group with what we’ve been working on
Initially these weekly meetings were in person, but this project started during Spring of 2020, so we only had a few of these weekly meetings before having to move everything online

There were definitely some adjustments that needed to be made, but the format that we had, really translated well to moving online. I was still able to communicate with the other undergrad students through virtual meetings and our slack channel and the weekly full group meetings were moved to zoom. In fact, because we moved online, the other undergrad students and I were able to meet more often
So, while some other labs on campus had to stop abruptly, we were able to continue relatively seamlessly

As a result of our weekly progress report meetings, I very quickly got better at communicating what I’ve worked on and what I was struggling with
This communication and being able to give project reports were very relevant to my experience during my internships, where I would have to summarize my short-term work during twice weekly standup meetings and broadly during my mid point and final presentations
One of the greatest skills I’ve taken away from skunkworks is being able to identify what I need help with, and coming up with very focused questions to help get me unstuck which was very helpful during my internship

In addition, because of the success we had in this project and the great support system that I had, in future skunkworks projects, in addition to the usual research work, I was able to take on a role similar to Mingren, where I would direct some of the other students to investigate certain topics and figure out how to utilize it for their part of the project

5 of 8

5

Motivation: Properties of metallic glasses

Research Tasks

Learn relevant technologies
Cleaning Data
Fit Model
Metrics and Imbalanced Data

Dataset before any processing

In the figure, yellow represents data that is missing

Dataset after processing

On to some of the research tasks,
Our first task was to learn the relevant technologies for the project

First experience with python, using machine learning libraries

Had to learn how to learn something with more independence than I saw in my classes

For this project, we worked with patient data and data that was extracted from patient scans
Handling data and cleaning it to prepare to be used by a model

The two images represent the data before and after processing. The yellow represented missing data where we either needed to remove the column or fill in the missing data

Once the data was cleaned, we could fit the model to the data
From there, we Obtain metrics of the model’s performance
Initially it seemed like our model was working very well right out of the box
However, after doing some investigating, we realized our model was predicting every cyst as mucinous due to the composition of the dataset

Logically, it makes sense that the hospital would primarily be treating patients with the more serious type of cyst, and therefore have more data for it
The model realized that because our data was primarily for mucinous cysts, if the model predicted everything as mucinous, it would seem to do very well using accuracy as the metric

This made us realize that not only did we have to find the solution to the research problem, but we also had to make sure that we understood the problem first

We also had to reassess the metric that we were using to determine success

6 of 8

6

Motivation: Properties of metallic glasses

Source: https://doi.org/10.1007/s00261-021-03289-0

7 of 8

Key Takeaways

7

Motivation: Properties of metallic glasses

Key Takeaways

Technical Knowledge
Communication
Research Experience
Industry-relevant Experiences

Lastly, to wrap up, I just wanted to go over some of my key takeaways from my time with skunkworks
Technical knowledge

Learn python and machine learning and statistic principles
Learning with more independence than I was used to

Communication

I became comfortable presenting progress reports and learned how to ask the right questions to continually make progress
This helped me during my internship when I communicated with my manager and other coworkers

Research Experience

Skunkworks was my first experience with research and trying to solve these research problems that had a lot of uncertainty
The support system and education group really helped make research approachable for me

Industry-relevant experiences

By far, this is the Experience I get asked about most during interviews
Before I had internship experience, this is what I could point to to show that I could work as a part of a team to solve problems
Most employers understand that research is very different than classwork, employers want to see that you are able to work on these open-ended problems with more independence than you see in classes

Really enjoyed skunkworks

Through this experience I learned that I really enjoyed working on research and I hope to incorporate it into my future in some way

Thank you all for listening, please let me know if you have any questions

8 of 8

Template slide

8

Motivation: Properties of metallic glasses

[Title]

items