I WAS TURNED ONTO AI IN 2018
and having been learning and hacking it. I tinker with machine learning models to crunch COVID-19 data, generate voices, race autonomous model cars (and won!)
However, my attempts to have AI write my essays and papers have produced gibberish so far. Sooner or later I will figure it out...
AI LEARNING TIPS
Having tried a bunch of learning options over a few years - some clicked better than the others - I created this AI learning roadmap for self-learners and AI Clubs.
It starts with fun and gentle introductions of technical concepts, and gets more ambitious. My plan to start a high school AI Club was derailed by COVID. I would like to pick it up in college next year.
HACKATHONS ARE FUN!
I won #1 at Amazon Web Service DeepRacer contest in NY in 02/2020.
DeepRacer autonomous model car uses Reinforcement Learning (RL) to learn complex behaviors by interacting with its environment. Through reward functions, the model learns to navigate the tracks without being explicitly programmed. RL is the technology behind AlphaGo, the AI that beat human Go champion.
In racing, different algorithms optimize for competing objectives: steering accuracy vs speed. My winning approach was to research past algorithms and, through trial and error, choose compatible ones to tweak, simplify, and synthesize into my own new algorithm that struck the delicate balance between speed and accuracy. See DeepRacer Technology stack below:
Today's AI is mostly "narrow AI" that can only perform specialized tasks. Models require massive data to train in a specialized domain. Lacking such data is a hinderance to AI.
Transfer Learning applies knowledge learned from one domain to the others. much like how humans learn. Some believe this is a key to Artificial General Intelligence.
I validated the efficiency of using smaller datasets to achieve satisfactory model accuracy with transfer learning on Natural Language Processing using ULMFit. This is a technique learned from online deep learning class fast.ai. My paper also explained the basic concepts of deep learning and transfer learning.
MODEL Transfers Knowledge from one domain to another
A survey of Privacy-preserving techniques that I co-authored.
Research proposal and early findings for a National Institute of Health project. I assisted in background research and findings review.
My friends at Johns Hopkins got me into their project to create and detect fake voices using Generative Adversarial Networks. I helped set up the AWS cloud environment and machine learning models.
As a research assistant at A3.AI - a nonprofit Applied R&D organization that I cofounded, I work with expert data scientists to fight COVID-19.
I contributed to statistical methods design, data collection, papers editing and review.
Advisor: Dr. Mahesh Shukla
This study focuses on Social Determinants of Health. For example: COVID-19 Disease Burden by Age, Occupations, Race, Access to Care, etc.
Advisor: Professor Mark Cullen, Director of Stanford Population Health Program.
