5.5.22-Efficient-AI-Interpretable-AI

DAISY ZHUO: Hi, everyone. My name is Daisy Zhuo. I'm the co-founding partner at Interpretable AI. I started the company a few years ago with my doctoral thesis advisor, Professor Dimitris Bertsimas and a classmate of mine, Jack Dunn, was a shared passion for analytics and a vision for the future of AI.

We all know AI has made tremendous progress in the past decade, especially in areas like vision and in natural language processing. But this did not come at no cost. The models are just getting bigger and bigger. Take the state of the art natural language processing model, for example. It has billions of parameters and costs over millions of to train. And this poses significant challenges both in terms of the cost of training the model, the actual deployment, because there are some speed and size constraint for the devices, and perhaps more importantly, some models need to be interpretable. People have to be able to inspect the models to make sure that the loans are not discriminating against certain socioeconomic groups, for example.

So this makes us wonder, is the bigger the better really the future of AI? And that was the core theme of our research and of the company as well. To provide a little bit more background, these deep learning models belong to a certain class of machine learning called black box models. So in this example, you can see that you have some input. And through a giant model after you crunch the data, it gives you a single point estimate. In this case, the mortality rate for the patient.

In contrast, the interpretable models and the core design principle behind our product takes the input and tells you because this person is male and over the age 30, that's why the mortality rate is a certain number. So it may not even be accurate, but at least it is transparent and someone can see and inspect and agree or disagree.

And because these interpretable models are hugely popular, they've actually been around since the '70s. However, there is a trade off between the performance and the interpretability. And our goal was to really bridge the gap. So we used some of the modern advances in mixed integer optimization, global optimization, and convex programming to really push the performance of these models to the next level where it's comparable, perhaps sometimes even better than their black box counterparts.

And not only that, that we focus on the prediction task, which a lot of the deep learning methods do, we cover the entire spectrum of the data analytics process going from data cleaning, including the module optimal imputation, which was the topic of my thesis, to predictions, very important, and to prescriptions where you can actually make data driven best decisions simply based on observational data alone. You don't need to run A/B test. You can just learn what action works well in certain controlled environments.

So to give you a case that we've worked on, this major European car manufacturer came to us and wanted to learn their machine failures. What is driving the machine failure and can we detect and also improve on them? So a traditional black box model can have a very good prediction of that and give you alert, but the issue is that's not going to help you solve the problem. The machine failures are still going to be there.

So we trained our optimal decision tree with an example outlet here and it shows you some paths to failure so that the engineers can actually when they get such alert actually inspected and see, oh, it's because the pressure is too high or the temperature is too low so that they can make such adjustment and actually avoid these failures. So it was deployed in several plants and we've calculated that a few of hundred of downtimes, 100 hours of downtime were avoided.

And because of interpretability is needed in many industries, we have had partnerships and clients in various areas, including finance, insurance, health care, retail. So we're really looking for anyone from a variety of industries to either license our software to become a client of prebuilt solutions or to develop new solutions together. That can be something that your whole company can trust with that. So I hope to speak with you at the lunch time. Thank you.