AI4All Day 11: Improving the Accuracy of our Model, Robust Learning, and Industry Panel

Nidhi Parthasarathy
3 min readAug 20, 2022

--

Nidhi Parthasarathy, Tuesday, July 12th, 2022

One Rollercoaster of a Project

In today’s morning session, we worked on our projects again with our cohorts. We spent some time debugging why our accuracy rates showed incorrect values. We tried implementing the code in scikit on our own.

Our accuracy rates went up to 92% but we realized that it was due to another bug in our code. We were not checking the entire matrix but instead were checking only one result in the matrix.

We ran out of time to debug this problem, but after class, during the break, I was able to rewrite the code and it worked, so I was very happy! :D

The working accuracy code

Learning Robustly from Limited Samples to Make Good Decisions

In the afternoon session, we had a talk from Emma Brunskill (professor at Stanford) on learning robustly from limited samples to make good decisions.

She talked about the Markov decision process which provides a mathematical framework to model decisions. This method is very popular in Artificial Intelligence to model sequential decision-making scenarios.

She also talked about learning from batch samples to robustly make decisions and went through a really nice diagram that explained this. Apart from this, she also talked about counterfactuals which allow people to explain outcomes of the past and also predict future outcomes. Moreover, she explained concepts like probability of intervention, overlap requirement, accounting for modeling errors with limited data, etc.

One part that I really liked was when she showed us an example of an AI math tutor that helps with solving math problems. This really helped emphasize a cool example in the real world that I could relate to.

She also talked about other examples like diabetes insulin management. Towards the end, she talked about policy optimization with constraints, competing priorities, principled tradeoffs, and learning the best policy.

This was a very advanced lecture but it was also very interesting.

An Industry Panel

The next session was an industry panel. The people in the panel were Tamara Schmitz (From Micron technologies), Caroline Duffy (Consulting and Management Consulting), and Montserrat González (Google Robotics). All three panelists talked about their journey and about their career in industry. Even though they all had different day-to-day schedules, they emphasized that communication was key in all their jobs.

Caroline explained how she talks to many different start-ups day after day and determines their fit with her investment criteria. This requires a lot of research, and she likes to see if they have a solution to a fundamental problem.

Montserrat, on the other hand, doesn’t talk to many external people in her day-to-day job, but she does have a lot of meetings to coordinate connections across Google. For her to start something new, she has to talk to other areas in Google to make things work.

Tamara talked about how she had a talent for explaining complicated things in a simple way. Her day consists of talks to executives, making the complicated stuff more understandable, and then talking with her team to test products before they are launched.

One of the most important things I learned from this presentation was that you don’t need to have all the answers and that you should be comfortable asking for help. This was one of my favorite panels in all the sessions so far.

Read on for Day 12.

--

--

Nidhi Parthasarathy
Nidhi Parthasarathy

Written by Nidhi Parthasarathy

Highschooler from San Jose, CA passionate about STEM/technology and applications for societal good. Avid coder and dancer.

No responses yet