Many people are surprised to learn that Unlearn’s founding team didn’t set out on a mission to revolutionize clinical trials or even other areas of medicine. Rather, we started the company in 2017 on the belief that generative models were the future of machine learning and artificial intelligence. That’s how we ended up with the name Unlearn.AI—a portmanteau of Unsupervised Learning—instead of a name that signals a focus on medical applications. I don’t think we even solidified around a focus on medicine until 6 months or so after we started. Perhaps that’s why I find it so strange when people think that Unlearn is a biotech company when that is definitely not the company we started—we started an AI company.

To clear things up, here’s a very short history of how Unlearn was founded (probably with important details missing; it’s been a few years).

Me and my co-founders, Aaron and Jon, had been working together as part of the machine learning team at a virtual reality company and we each decided to leave to pursue other opportunities around the same time. Prior to that, we had been academic researchers in math and physics, and I had a short stint as a computational scientist at Pfizer. Over the next couple months, we’d get together once or twice a week with another friend of ours Cami (who became an advisor to Unlearn) to throw around some ideas. These ideas usually focused on the intersection of two themes:

• We believed that generative models offered a more powerful framework for AI than discriminative models; yet, few real applications of generative models had made it to market even though they, particularly Generative Adversarial Networks (GANs), had become a popular research topic.
• Almost all of the research on generative modeling focused on images and text and we believed we would be able to discover some cool stuff if we bucked the trend and simply focused on some other problem like medical data or something.

One of the key differences between images and many real world systems is the presence of noise. Humans prefer to look at crystal clear images, as evidenced by the ever increasing number of pixels on our phone screens and televisions. We like sharp lines and high contrast. But, most real world applications aren’t like that at all. The world is noisy. Perhaps the types of algorithms that are great for generating crystal clear images and those that are great at modeling noisy, real world systems are different? We thought they might be.

We got to work resurrecting a class of generative models that had fallen out of favor—Restricted Boltzmann Machines (RBMs). RBMs are interesting generative models for real world systems, in part, because the last step in generation involves adding noise and because they are undirected which allows one to naturally integrate over missing variables. That is, they provide simple ways to model noisy systems with missing data; and it turns out that medicine is full of problems like that. We wrote some software that combined modern concepts in deep learning like training on GPUs and using adaptive optimizers, but we ended up going quite a bit farther and created an architecture that crossed RBMs and GANs into something new—at the time we called these Boltzmann Encoded Adversarial Machines (BEAMs) but it was a missed opportunity to call them BoltGAN Machines (which I now prefer by a large margin).

Now, here’s the crazy part, at this point we thought “let’s go try to raise some venture capital to see if anyone else thinks that generative modeling is going to be a big deal?” So we did! We went out and raised a small pre-seed round from DCVC based entirely on this thesis and on tech. We did not have a business plan, or even a business person (Aaron, Jon, and I are mathematicians and physicists by training).

Our pitch went something like this. There are lots of things for which we wish we could create highly accurate computer simulations, but many of these systems are too complex to handle with mechanistic models. Rather than trying to work out detailed mechanistic models for these incredibly complex systems, we thought that we could use generative AI to learn simulators for them. We talked about simulating health outcomes for patients in clinical trials (so the seed of the idea was in the original pitch, inspired by problems I saw at Pfizer), simulating the effects of perturbations on gene expression, and even simulating the weather.

Even though our pitch was all about generative models, we knew that we’d have to find a specific application in order to create a viable business. I recruited a friend of mine, Graham, to come out from Boston and join us as our business person. With the team in place, we decided to take a lean startup approach and try to figure out the right vertical by getting a few engagements with potential customers and trying out different applications. After working on some projects focused on simulating gene expression and others focused on simulating longitudinal health outcomes, we found that we did a lot better at the latter. We never did try to build a deep learning weather simulator, but that did end up becoming a real thing. So we decided to focus on solving the problem we were best equipped to solve, creating digital twins of people to predict how their health may change over time.

Anyway, that’s how we started Unlearn and how we ended up going down the path towards medicine. To this day, we still believe that generative models are the best framework for advancing AI, and that fascinating discoveries await in the unexplored areas of AI in medicine. We intend to find them.