My Experience At The AI Latin America SumMIT
A few weeks ago, I had the opportunity to participate in a hackathon at MIT, and for my first ever hackathon, I must say I thoroughly enjoyed it. I got to meet some super interesting and smart people who were all enthusiastic about using AI to improve the world. From meeting a computer scientist in Brazil to a professor of Ethics, I was immersed in diversity.
Within the five focus areas of the hackathon(Poverty, Education, Justice, Climate Change and Health) I chose Climate Change because I believe there is huge potential in using AI to help solve the issue of climate change. I joined a group of 4 other people who were equally passionate about climate change, and we named ourselves the Climate Hackers.
Hacking(Our solution)
The majority of the first day was spent finding and organizing data sets. First, we had to find the data set. We were given two data sets which consisted of monthly climate data in Brazil and dates on which there were forest fires. The two data sets did not match up, one had a monthly frequency, while the other had a daily frequency. Mentors at the hackathon suggested to find data with daily, and maybe even hourly frequency. This meant that we had to find two data sets with hourly frequency. After several hours, I found a data set with daily accuracy for climate in Brazil with 18 different parameters and a data set provided by NASA which gave hourly frequency of the forest fires.
Now came data organization. Being a novice at coding made data organization a bit more challenging. However, with the help of some of the more experienced coders on my team I was able to organize the NASA data set. Some of our other group members organized and visualized the climate and NASA data set, and prepared them for the neural net.
After a day of data, then came the fun part; the neural net. Unfortunately, I was unable to participate in the development of it but I do know we built a Light GBM. A light GBM is a gradient boosting algorithm which is based on a decision tree, allowing for a ranking of variables to be created. This is especially helpful for our situation because if we are able to visualize the importance of climate data (max temp, altitude, etc.) it can be used to help identify the signs of a potential forest fire. Light GBMs are also useful in that they have faster training speeds and higher accuracy due to their unique method of decision tree building, leaf wise growth instead of level wise.
After training, we found that the most important variable is altitude followed by the date and max temperature.
We then used this data to estimate the probability of a forest fire starting with 83% accuracy.
What we built has huge potential for countries around the world who struggle with the issue of forest fires. Even though only a small fraction of forest fires are started by abnormal weather conditions, algorithms like this can help estimate forest fires caused by such weather conditions.
There were many other teams at the hackathon, all of which had really cool ideas. Here are a few that stood out:
- The Health Hackers built a variational auto-encoder to identify the symptoms of hypertension and help a patient change their diet to reduce symptoms
- Anglue, a team that built an NLP chat bot which was able to give speakers a score out of 10 rating how well they spoke English, and use that to help teach them English.
Takeaways
As this was my first hackathon, I have some key takeaways:
- My biggest takeaways from this experience is that 80% of creating and training a model is finding the right data. Without the right data properly formatted there is no model
- There is no right answer, and don’t be afraid to take risks in your solution
- You have to figure it out and push yourself in order to finish with a successful product
Finally, I would like to end by saying thanks to my group members Issac(for helping with the code), Laurens(for also helping with the code), Ricardo(for putting together the presentation and helping with big picture issues), and Biva(for also putting together the presentation and finding data sets).
Also thank you to Luis for giving me and seven other high schoolers the opportunity to participate in the first every Latin AI hackathon at MIT!
Connect with me:
Linkedin: https://www.linkedin.com/in/vikram-menon-986a67193
Email: vikrammenon03@gmail.com
