5.4.22-Startup-Ecosystem-Themis-AI

ELAHEH AHMADI: AI models are actively used to make decisions about our loan qualification, job screening, or health care access. However, many of these models are proved to be biased towards certain demographics and women. My name is Elaheh Ahmadi, and I'm the CEO and co-founder of Themis AI. Themis was founded by a group of MIT researchers and professors to bring the solution of fair and robust artificial intelligence to the industry.

Many companies over the past couple of years have been starting to adopt AI broadly to increase their revenue and decrease their costs. However, a lot of these companies, while they've been successful in increasing the revenue, they've been exposed by having biased models and having to deal with the negative publicity as a result of that. However, in addition to the negative societal impact of biased models, there is a huge millions and trillions of untapped revenue as a result of using such models.

You may wonder, where does this bias even come from? So the bias can be within your data. You can have an imbalanced data that already has a lot of biases. It can be within your model architecture, or your data distribution can shift over time. At MIT, we developed a technology to be able to automatically detect risk and biases and mitigate them during training.

Many of the current models are designed to provide an output, which can be a prediction or classification, given an input data. With our technology, we can convert your model into a bias and risk aware model, such that, in addition to the prediction, it will provide you with a risk score that includes bias score, uncertainty, and explaining how the prediction was derived. Now let's have a deeper dive of how our solution works.

So the very first step of being able to mitigate bias is identifying and extracting hidden features within your data, so basically, understanding what your data is and what it includes. In this example, we were able to extract features such as skin color, gender, hair, and position of this data set, all automatically without any need of prior labeling. The second step is understanding how are these features distributed within your data, and what part of the data that you're working has your model already learned in certain, and what are the areas that the model is still uncertain about?

Combining these two information together, we can train, certify, and deploy AI models that are robust toward all the edge cases and difficult data samples within your use case. We applied our technology in many different scenarios, such as health care, finance, autonomous vehicle, and facial detection. In one of our research, we applied our technology on a clinical trials outcome predictor.

And we were able to improve the overall performance by nearly 20% in addition to improving the true positive rate by 50%, which resulted in a financial benefit estimation of more than $700 million per trial. In another use case, we applied our technology on predicting diseases from chest X-rays. In this use case, we were able to improve the overall performance in addition to improving performance on underrepresented diseases by 70%.

At Themis AI, our mission is to create a world where AI is fair and unbiased. And we're doing so by providing an end-to-end de-biasing platform to enable fair and robust AI solutions. At Themis, we want to work with your company to enable use of AI and automation in regulated spaces, enable explainable and fair AI, and uncover and mitigate risk and vulnerabilities within your current AI models.

Again, my name is Elaheh Ahmadi. Thank you so much for your time and attention.

[APPLAUSE]