5.5.22-Efficient-AI-Themis-AI

ELAHEH AHMADI: AI models are actively making decisions about our health care access, loan qualification, or job screening. However, many of these models have been proven to be biased against racial minorities and certain genders.

Hi. My name is Elaheh Ahmadi, and I'm the CEO and co-founder of Themis AI. Themis was founded by a group of researchers and professors at MIT to bring the solution of fair artificial intelligence to the industry.

Over the past decade, a lot of companies have started to adopt AI to improve their decision-making and their products. And many have seen great results by using AI in increasing their revenue and lowering their costs. However, a lot of these companies have also suffered financial loss and negative publicity after being exposed of using biased model that are biased towards certain racial minorities or women.

However, that's not the only negative impact of using biased models. There are millions and trillions of untapped revenue and lost revenue as a result of using biased model. For example, by expanding credit, that could have added 13 trillion in income.

So now you may wonder, where does this bias even come from? So the first one is it could be in your data. Your data could be imbalanced, or if you're using historical data, most of these historical data are portraying these biases that we've had in our society over in the past couple of years. Or it could be even in your model architecture, or simply your data distribution can shift over time. For example, COVID can happen.

So at MIT, we developed a technology to be able to automatically uncover biases within models and data and mitigate them. So a lot of current machine learning algorithms produce an output, which can be a prediction or classification for a certain input. With our technology, we can convert current machine learning algorithms to bias-aware ones. So now, in addition to having the prediction for a certain input, you can also understand what are some of the risk metrics associated with this prediction, such as the bias score, [? uncertainty ?] [? score, ?] and also explaining how was the decision even derived.

So let's dive into how our solution actually does this. So the very first step in designing a machine learning algorithm that is not biased is understanding how is your data distributed and what is in your data. As a result of that, we developed a technology that can automatically identify and extract hidden features within a data set. In this example, we were able to extract features such as skin color, gender, hair, or position of the face, all automatically, without any prior need for labeling.

The second step is understanding how are these features distributed within your data, and what are the areas within your data that your model is certain and has already learned, and what are the areas that the model is uncertain and still needs to learn. Combining these two pieces of information, so better understanding of your data and understanding of the certainty of your model, we can actually train, certify, and deploy AI models that are robust and fair. We applied our technology in many different industries and sectors, and we were able to achieve great results.

In one of our case studies, we applied our technology on a clinical trials outcome predictor. We were able to achieve an overall improvement in performance by 20% and improvement in performance of predicting true positive cases by 50%, which added in an increased revenue of more than $700 million per trial. In another case study, we applied our technology on a predictor of disease based on chest X-rays. And we were able to improve the overall performance, in addition to 70% performance on underrepresented disease.

At Themis, our vision is to create a world where AI is fair and unbiased. And we're doing so by bringing an end-to-end de-biasing platform to the industry. We want to work with your company to enable the use of AI and automation in highly regulated spaces, enable use of explainable and AI at your company, and also be able to uncover and mitigate risk and vulnerabilities within your current models.

Again, my name is Elaheh Ahmadi. I look forward to talking to you in the other room. Thank you so much.