My 🤯 Experience at the AI Latin America SumMIT

24 Hours of Hacking with the Smartest and Most Diverse Groups of People

Imagine traveling to ten different countries 🌍 and talking to people about exponential technologies like artificial intelligence, gene editing, and quantum computing, that will change the world.

That’s what my experience was like when I attended the AI Latin America SumMIT hackathon this past Friday and Saturday.

Not only did I learn about how Artificial Intelligence and Machine Learning can be used to solve world problems like climate change, healthcare, crime, education, and poverty, but I broke the barrier between myself and the rest of the world in doing so. It’s not everyday that you get to speak to researchers in Spanish at 100mph on climate change and artificial intelligence. 😂

From talking to a biology researcher from Argentina about her views on the use of cellular agriculture to discussing with another researcher from Peru the impact that the large scale use of AI and massive computing power will have on exacerbating and accelerating climate change, I was greatly able to expand my world view.

At the end of the day, I was able to sit down with such a diverse, talented group of individuals to solve some of the world’s most challenging, pressing problems.

Our Problem, Idea, and Approach

My team Health Hackers decided to apply artificial intelligence to a common healthcare problem.

Normally, when you go to the doctor and get diagnosed with a particular disease, the doctor instructs you to change many different aspects of your lifestyle including your habits, diets, sleeping practices, etc. However, these combined instructions are often very hard to implement. What if we could use Artificial Intelligence and Machine learning to evaluate the minimum change in diet or behavioral practices that would lessen one’s chances of acquiring a disease?

Our team proposed the use of a conditional variational auto-encoder (CVAE) model to generate distributions of symptoms for a particular disease and understand how to alleviate symptoms through modifications of elements in the latent space.

Quick Crash Course on VAEs:

Variational Auto-encoders are a class of machine learning models that take multi-feature multi-dimensional input, use an encoder to reduce the dimensionality of the data and compress it into an encoded latent space, and create variations of the original, multidimensional input by sampling a distribution of the latent space and using a decoder to produce the final output.

A common Variational Autoencoder (VAE) where mu and sigma are used for backpropagation and epsilon is sampled from a normal distribution to generate variations.

For example, a VAE could be used to encode a complex structure such as a face, with independent features for eye color, facial lines, mouth, etc. into a vector with 10 elements! The VAE is essentially finding an efficient way to compress the data and store it in less space without losing too much information.

A CVAE is similar to a VAE in that it’s a generative model that outputs variations of the inputted features. However, it allows us to specify a class from which we want variations to be sampled. For example, if we used MNIST datasets for numerical digits, we could specify in a class variable c the digit for which we want to create variations.

We used an NHANES dataset with questionnaire information regarding symptoms and behavior, lab examination results, and medications to select relevant features to the disease that we chose to focus on: hypertension. Hypertension results from an abnormally high blood pressure and affects more than 40 percent of the population living in Latin America.

After selecting the input features relevant to hypertension such as carbohydrate and fat intake, we wrote code for the conditional VAE. Unfortunately, we weren’t able to get code up and running by the end of the hackathon, but we did receive an honorary mention from the judges for the merit of our idea and the complexity of the problem.

Future Directions

In the long term, with a conditional VAE, we can potentially generalize the model to a larger amount of diseases (by specifying a certain disease, we will be able to generate variations specific to that disease).

Our end goal for this project is allow for the input data/features to be encoded into interpretable latent/salient variables in the latent encoded space. This would allow us to manipulate interpretable elements of one’s diet such as carbs to infer the minimum change that is needed in order to get them out of the latent space of a particular disease.

In the future, we plan to look into other different types of Variational Auto-encoders (other than conditional ones) such as contrastive VAEs and disentangled VAEs for this main purpose.

Contrastive VAEs:

A Contrastive VAE works by taking the background noise/variation out of a dataset.

The cVAE essentially enhances the “salient” or important features of a dataset during reconstruction with inputs for both the target and background datasets.

This could potentially help us to eliminate background variation in diet/nutrition within a human population, thereby allowing us to solely focus on latent variation that stems from a particular disease.

Disentangled VAEs:

Disentangled VAEs would allow us to craft a latent representation where each element is uncorrelated with each other. In other words, linear combinations would not overlap and each encoded feature would be saying something different about the data.

In order to understand how to do this, we have to understand how the loss functions for VAEs work.

The loss function for VAEs consists of two main elements:

• Reconstruction loss: This measures and ultimately minimizes the difference between the decoded output and the initial input.
• KL(Kullback–Leibler) loss: This loss function ensures that that the decoded probability distribution is relatively similar to the distribution of the encoded latent space.

As part of the loss function, a constant ß would be added to KL element loss function in order to weight each of the latent variables differently (essentially a larger beta value would select for the latent variables that result in a final decoded distribution that is closer to the prior distribution).

Each element in the encoded space wouldn’t be an abstract random variable: rather it would be a variable attributed to a particular aspect of one’s diet, biomarkers, or behavioral practices.

Amazing Projects and Innovation in AI

I saw amazing other projects outside of the AI x Healthcare space.

Some included…

• Anglüe — a chatbot that could test for English proficiency in Latin American countries and help educate students to become a part of the larger global community.
• AI x Wildfire Prediction — To teams focused on using machine learning to identify potential future sites for wildfires in brazil based on independent variables such as humidity and temperature.

The most mind blowing project I heard about was from James Weis, a researcher at MIT. With the Scaling Science Project, he’s constricting a trillion node graph network with institutions, researchers, and publications. What’s the use of doing this?…

We could, with the help of AI, predict the impact of scientific topics and publications decades before the inflection point. 🤯

Shoutouts!

Shoutout to Ivan (practicing physician in Germany) for helping us with the more medical aspects of hyper tension and helping us select specific features for the input.

Shoutout to James for helping us finally get the dataloader to work!

Shoutout to Luis (PhD researcher at MIT) and Camilo (researcher at CSAIL) for guiding us along our journey during the hackathon, bouncing off ideas, and helping us fix really annoying bugs in the code.

Thanks for reading! Feel free to check out my other articles on Medium and connect with me on LinkedIn!

If you’d like to discuss any of the topics above, I’d love to get in touch with you! (Send me an email at mukundh.murthy@icloud.com or message me on LinkedIn)

Please sign up for my monthly newsletter here if you’re interested in following my progress :)