Blogs

Participants from the micro-internship keep journals as they progress through the two-week program. Use the drop-down menus below to read about their experiences!

Claire Camacho [2023]

Claire Camacho is a sophomore at MIT studying Computer Science and Neuroscience. She is interested in the intersection of data, software, and neural engineering, and its applications to social justice. Claire spent the summer of 2022 interning for a cybersecurity startup in Tel Aviv and teaching Python to a binational group of Israeli and Palestinian high schoolers. She is passionate about improving equitable access to technology and STEM education, which guides her work with Harvard’s Women in Technology conference, SWE, and the Women of Color Professional Initiative at MIT. In the future, Claire hopes to continue her exploration of AI and grow her interdisciplinary skills in technology through undergraduate research and summer internships.


Orientation + Meeting 1; Choosing our project

TryAI’s orientation zoom is filled with familiar faces from the undergrad meeting as well as new mentors we’re meeting for the first time. In our first official meeting after orientation, my mentor Kezia and I talk about our backgrounds and goals for the next two weeks. I share about my summer abroad interning for a startup in Tel Aviv, and Kezia tells me this will be her second time experiencing winter because she was born and raised in Indonesia. We have similar goals for the future, which are to complete our Master’s in AI/ML adjacent fields and go into industry. Kezia and I both have a background in the health sciences and hope to build tools to make healthcare more effective and equitable. This leads us to choose diabetic retinopathy medical imaging as our data set. In the next two weeks we will explore how deep learning can reveal which parts of the eye show markers of impending blindness, and build a model to diagnose this disease even when imaging is not of the highest quality.


Meeting #2; Algorithm Exploration

In our second meeting, Kezia and I dive into what AI and deep learning really mean. While I have a background in data analysis and software engineering, this is my first time gaining exposure to the backend of how these models are built and used. Thankfully Kezia has a history of mentoring others on this topic, and she presents a slide deck going over the layers of ML models, common mistakes like overfitting, and a simulation showing how layers and perceptron grouping can affect how a machine learns patterns. Drawing on my previous experience with the Pandas library from an internship, we set a goal for our next meeting to organize our dataset into a pandas DataFrame and begin applying a pre-trained CNN using PyTorch. I spend the next few days after this meeting finding that learning these new techniques is easier said than done! Getting a handle on Google Colab, Gdrive file synching, and image handling in Pandas feels like a barrier to actually learning AI at the moment. However, I am no stranger to the frustrations of a steep learning curve and recognize that these foundational skills will set me up for greater success in the future. While I wrangle these starting technicalities, I read articles and watch youtube and OpenCourseWare videos on CNNs and other deep learning techniques.


Meeting #3; The Final Push

In our third meeting, Kezia and I review code for retinal image pre-processing. We are drawing on several research papers written on deep learning for our dataset, which is helpful to me to be able to learn by example. Having such a visual project means that each image-altering success feels more concrete than most software engineering outputs. The potential real-world impact of AI in healthcare gives an extra spark of motivation to these dense two weeks of research and implementation.


Kezia and I also talk about where we see our futures, especially in the uncertainty of the tech world recently. We both have had successes and failures in our career search we can laugh over. I share about a downtrodden interview with a Brazilian tech company 10 minutes after their team lost a world cup game, and Kezia shares her internal debate of academia vs industry. Even though neither of us knows quite where we want to direct our future, our passion for using technology to better society and ourselves guides us.


Meeting #4; Presentation Rehearsals and Goodbyes


In our final meeting, I practiced my 10-minute project presentation with Kezia. My voice is barely hanging on after fighting a cold the past week, but the presentation goes well and there are only a few minor changes Kezia points out. I’ve learned dozens of new terms in the past two weeks, differentiating between types of neural networks, data processing techniques, and useful libraries. It’s a lot of information to process in such a short amount of time, which is why I’m so grateful I had Kezia as a mentor to answer my endless questions. TryAI has helped open the door for me to join an MIT Ph.D. student in conducting Machine Learning research for nanoscale transistor optimization. I also plan on bringing my new AI skills to my summer software engineering internship. I will be working with a non-profit tech company that is reforming the fragmented data of the US criminal justice system and advising policy changes for more equitable, data-backed solutions.

Hannah Kim [2023]

Hello! My name is Hannah, a sophomore at Harvard studying CS and Statistics. During the Try AI program, I explored the subfield of XAI, which is the field that studies ways to make Artificial Intelligence more interpretable and understandable. In the future, I hope to be working in the intersection of technology and social impact, whether that is through research, industry, or business -- I am honestly still figuring it out. Some social impact areas that I am passionate about right now are education, government, and climate. Here is my LinkedIn! Always happy to chat with more people. 

Meeting #1: Orientation and kickoff

Really excited and glad that I got matched with the mentor that I did. Sonja was engaged the entire time and it really felt like she was trying to give me actionable advice. We ended up talking for more than an hour. Probably the most enjoyable part for me was just learning about her personal experience with research and pursuing a STEM degree. I was surprised by how we had similar interests in education technology, but from such different experiences (I had a lot of exposure to it in middle and high school because my school was based off of an edtech platform while Sonja worked a lot with making useful edtech to improve the experience of grading CS assignments as a teacher’s assistant during her undergrad years). All in all, I learned a lot and am looking forward to the next few sessions. I’m just a little bummed because I’m getting my wisdom teeth pulled out so I really won’t be able to make substantial progress in the first half of this program. 


Meeting #2: Reviewing the Literature review 

We reviewed some of the top takeaways that I had from my literature review. I personally learned a lot and was excited because there were so many subfields of the field that I was not aware of, and because Sonja’s expertise isn’t Explainable AI, it was nice because we were learning and discovering new information together. What we did afterwards was narrow down on a problem that I wanted to solve. This was difficult because from the literature review, I created a list of project ideas that I could potentially try just based off of the methods that I thought were interesting, but as a result, it was a lot harder to ideate a project proposal because we couldn’t map one technique specifically to a domain that was interesting to me. So we started just brainstorming problems, focusing on education since we both were interested in that field, and we were able to narrow down on the problem of retention of STEM degrees in college, which was largely based on my personal experience. Now that we clarified the problem, we were able to write out the rest of the project proposal so I could start the implementation. 


Meeting #3: Implementation  

I told Sonja that I was having trouble dealing with the implementation, especially when thinking about model accuracy. My reasoning was that I couldn’t really find a good dataset online that fit the exact specifications of our project proposal, so as Sonja had suggested, using a mock dataset would be a great solution, but I worried that the model itself would not be accurate or that I wouldn’t be able to improve because it wouldn’t be trained on real data anyway. Sonja first suggested that I could tweak the data to have built in results that I’d know would likely be true, which would be with the evaluation of my model in the testing phase. She also clarified a lot of the project expectations not as something where we need to have a really polished final product, rather, just a good idea of what we would try to do for next steps, so that lessened the pressure a bit!


Meeting #4: Final Presentation Runthrough

I was a little frustrated because I basically wasn’t able to implement anything because I wasn’t able to resolve several dependency errors I was getting with Anaconda and Jupyter Notebook. Additionally, my laptop (I upgraded to a new one) was having issues with connecting to my University wifi so I had to troubleshoot that for a while. So, I opted for just creating the slides of what I was able to complete, which was primarily the literature review and project proposal ideation. Sonja and I ran through my presentation, and she gave very helpful feedback, suggesting that I expand more on providing context for my project proposal and the motivation and also providing more context for my specific interests prior to joining TryAI. After running through the presentation I also got to chat with Sonja more about why I was pursuing the research pathway in the first place, and why I got disillusioned with going into government. All in all, a very productive meeting, and I know I’ll meet with Sonja again, especially since I’ll be taking classes in the same building that she works at!

Alex Wong [2023]

Meeting 1: 

In our first meeting, Jessie and I had a discussion on how expansive our project should be, in addition to learning more about each other. For the project, the first topic was the research that Liz had sent to me over break, in addition to some papers that Jessie had sent. We tried to narrow down the scope of our project, and eventually concluded that setting up a game that simulated illegal fishing/overfishing would be doable within the two weeks of the micro-internship. It also helped that Jessie had lots of prior experience with games, and I learned that her first foray into academia as an undergraduate was studying how to reach socially efficient outcomes in games. 

One point of interest that we discussed was subsidies when I mentioned that according to the paper, a lot of fishing would be unprofitable without subsidies. We realized this could be a good avenue of further research, so my task for our next meeting was to dig deeper into the impact of subsidies on fishing. 


Meeting 2:

First, we reviewed the fishing subsidy literature and the main takeaways that I had found. I found through a few studies that large percentages of fishing would be unprofitable without subsidies, and we decided that this could be the center of our project. My main homework for the next meeting was to try to find any public documents from Taiwan or China about their subsidy policies. Specifically, if I could find an algorithm that they used, or take any public data to create an algorithm, we could have a good start for a maximization problem. 

We briefly discussed what this optimization game might look like. Jessie gave the example of choosing between fishing for tuna or another fish based on expected profits, and if we could know what the subsidies are, we could more easily set up the game. On the other hand, we decided that If we couldn’t find any information, we’d use a rough approximation instead. 

The next topic of the meeting was how to structure a game. I also had the following questions: How does AI come into it? What model would a game use? That’s when Jessie discussed gradient descent, which was about following the regularized leader.   In a game, verifying Nash equilibrium is easy, but finding it is hard. There’s a lot of guess and check involved, but when you use gradient descent and libraries that do the optimization for you, it becomes much easier to do these calculations. 


Emails:

We had scheduled our next meeting for Thursday, but unfortunately, I got sick and it made it pretty hard to talk, so we decided to discuss my findings over email. My email mainly talked about the subsidy algorithm I had found in a highly detailed paper by Oceana. It listed out all of China’s fishing subsidy policies and included how to calculate many of them. I gave a report about the most important ones: tax breaks, fuel subsidies, and vessel subsidies, as well as the algorithms involved in each. 

In response, Jessie helped to further define the maximization problem we’d use, based on two types of fish that could be caught (tuna and squid). She also made a graph that showed what the basic idea of our problem would be


Meeting 3: 

Our penultimate meeting was a review of what we had discussed in the emails, as well as a further exploration of how we’d be implementing the game. We specifically talked about the cvxpy library, and Jessie sent me some materials including some starter code that would help code the game. 

We also talked about the final presentation. We decided that I would work concurrently on the presentation and implementing the game in cvxpy, with higher priority to finishing the presentation. She sent a slideshow called “Giving a Talk” by Dan Larremore, which was about how to conduct a great presentation. We decided that I’d give a large portion (around half) of my focus to the “story” of overfishing and illegal fishing, which is the context behind the project.


Meeting 4: 

In our last meeting, I went through all the slides in my presentation. Jessie gave some advice on how to make the slideshow better. Some of the most important advice included making the transitions between slides clearer (ie how does this slide connect to this slide), and which slides to prioritize and which slides to cut down to meet the time limit (specifically, doing less on explaining the concept behind gradient descent, which would take up a lot of time that could be spent on the project). 


Final Thoughts: 

I’m confident in saying that this was a great program. I thought the balance between independent work and guidance from my mentor was perfect. This program is really about “You get what you put into it”, and I personally enjoyed the self-motivated nature of the work. I also liked how we could direct the project in any way we wanted to combine both my interests and my mentor’s expertise. Jessie is awesome, and her knowledge was super helpful to me in navigating this space, whether it was reading academic papers or setting up a game with a library I’d never used before. 

Finally, I thought it was great to know more about Jessie as a person. Talking about her journey through academia and why she decided to take the path she did was extremely interesting and valuable, and I appreciate that this program had a focus on this personal side too. 

Nathan Moreno [2023]

Journal 1 - 1/11 

In our first meeting I prepared a small slides presentation for my mentor to outline two possible problems I would be interested in tackling. One of my ideas had to do with making predictions about morphogenesis in developing organisms using AI and the other was about segmentation and tracking of cells using AI. I was unsure which would be a more realistic problem to work on for this internship, but Hyewon (my mentor) said that I was noticeably more interested in the first problem. She said she could see “my eyes light up” when I talked about it and asked if I was planning on researching this question/topic in the future. Hyewon encouraged me to work on the morphogenesis question because it would give me foundation for future research and satisfy me intellectually. She clarified concepts (convolutional neural networks, u-nets, and backpropagation) from the articles I showed her and helped me find supplementary resources on deep learning. Her encouragement excited me not just about the program but also about my future career in research! 

Journal 2 - 1/13 

For my second meeting, I prepared another small presentation about one specific article I had read (Exploring The Behavior of Bioelectric Circuits using Evolution Heuristic Search). She helped me differentiate between what aspects of the study were artificial intelligence versus simulation models and she helped me narrow down my problem statement. Additionally, we discussed how AI could be applied further to the problem, what kind of datasets could be used (synthetic vs. real), and how supervised/unsupervised learning works. She suggested I look into recurrent neural networks and generative models. For next meeting I was told to download the programs used in the study I looked at and start to play around with them just to get a better understanding of the work 

Journal 3 - 1/18 

For the third meeting, Hyewon and I met briefly to discuss what I had learned from programs I downloaded. She assured me that in research, researchers often start by downloading code from previous studies and manipulating its parameters to gain a better understanding of what has already been done. Additionally, I laid out what my presentation would go over and we discussed some final details about how recurrent neural networks and supervised learning can be used in the prediction of morphogenesis! 

Journal 4 - 1/19 

For our last meeting we looked over my presentation and Hyewon gave some final comments and suggestions. She suggested that I add slides not just about the problem statement I ended up focusing on but also about all the other topics and ideas we looked into. This way, I could “present my accomplishments” better. Hearing her say this was extremely encouraging because I felt as if I had not done enough for the final presentation without actually implementing any code, but Hyewon was able to reassure me that I had put a lot of effort into forming an idea and researching a range of topics. After the TryAI experience, I definitely feel more prepared and confident to start research projects on my own!