The Unlearn Story

There are two important things to know about Unlearn.

Unlearn revels in being different. Most notably, its location is rather unusual for a technology start-up. It may seem out of place in San Francisco’s theater district, but it is quite comfortable among diverse eateries, hotels, and streams of tourists.

It also has a simple motto: do good science and damn the hype. 

That comes from Unlearn’s founding team: Charles Fisher, Aaron Smith, and Jon Walsh. A trio of physicists turned founders, they often describe themselves as anti-hype. In fact, they seem almost too anti-hype to have started a technology company in the most hyped place on Earth.   

They enjoy this twist of irony and have used it to create a unique company ethos. Unlearn’s founders don’t just build machine-learning technology. They want to be a vibrant team who value camaraderie and living life fully. The three founders see themselves as friends first and co-founders second. They also actively pursue diverse personal interests, including boxing, cycling, and woodworking. 

Unlearn is a portmanteau of “unsupervised learning,” the company’s specialty. But the name isn’t merely clever wordplay and thought-provoking branding. It shows that the founders want to create lasting, impactful solutions to challenging problems. They’ve chosen to do this in drug development, a field that desperately needs innovation. 

Developing and selling a new medicine takes almost $2 billion and over 10 years [1]. Much of that cost and time is spent recruiting people to participate in clinical trials. To solve this problem, Unlearn uses AI to change how clinical trials are run [2]. In a standard clinical trial, subjects are assigned to two different groups: the control (or placebo) group and the treatment group. Unlearn has created machine-learning technology to replace some of the human subjects in the control group with AI-generated subjects—thus allowing these human subjects to receive active treatments instead. 

As the pharmaceutical industry embraces this new way of running clinical trials, Unlearn’s founders are eager for the turning point that’s coming. They believe that moment will happen when AI becomes an integral part—not just of clinical trials—but of drug development as a whole.  



Charles, Aaron, and Jon all enjoy using physics-inspired approaches to solve problems. But beneath this shared enthusiasm, each founder has unique qualities that perfectly complement one another. For example, the three founders each worked in a different area of physics. Charles used statistical physics to explore biological problems. Aaron specialized in mathematical physics. Jon focused on theoretical particle physics. Although they gravitate towards different approaches, each approach is an essential ingredient for creating Unlearn’s rigorous machine-learning models. 

Unlearn’s founders leverage their diverse strengths in other ways as well. Charles is the CEO and product visionary. Aaron enjoys constructing and experimenting—always with an eye for optimizing performance—and has a keen sense of aesthetics. Jon has excellent analytical skills, patiently poring over datasets and probing models to understand their inner workings. 

As Charles, Aaron, and Jon work together, they share the same goal and vision. They want to develop the best machine-learning tools to extract valuable information from clinical data. And to them, Unlearn is and will always be a technology company first. 


Even from his days in biophysics, Charles had always wanted to build rigorous computational methods for biology. Following postdoctoral training in Boston and Paris, he worked as a computational biologist at Pfizer. After moving to the Bay Area, he began exploring ways he could use machine learning for clinical data. When he realized that there weren’t any good machine-learning tools to model such complex, challenging data, he determined to find the right tool. 

In the machine-learning world, models are not “one size fits all.” In other words, models that solve one type of problem may not work for another. Thus, finding the right machine-learning model to solve the right problem is crucial [3].

For example, the latest machine-learning technologies (such as this) are predictive models designed to answer a predefined question. This class of models couldn’t answer the kinds of questions Charles was interested in. If he wanted to understand what was happening in the control group of a clinical trial, for instance, he would first ask how a disease affects human subjects. This was a complex, seldom predefined question. To answer it, Charles needed to model all of a person’s characteristics from clinical data. So he had to find, or even build, a special type of model with this capability. 

Going back to his roots in physics, Charles believed restricted Boltzmann machines (RBMs) were up to the job. Inspired by statistical mechanics, RBMs have two advantages. They can model all of an individual’s characteristics. They’re also well-suited to the challenges of clinical data—varied data types, missing data, and the need to examine many different aspects of the data.  

Over the last decade, machine learning has progressed because of high-level machine-learning frameworks—such as Keras, Tensorflow, and PyTorch. When Unlearn was founded, however, similar frameworks for RBMs didn’t exist. So Charles, Aaron, and Jon set out to build one. They created an open source package called Paysage, which implemented several state-of-the-art unsupervised learning algorithms.

Impressed by the Unlearn team and their work on Paysage, investors at Data Collective led a small round of investment in the company in 2017. The capital enabled Unlearn to further develop both its software and business model for clinical trial modeling. Shortly after this investment, Unlearn’s founders built their first machine-learning models with electronic health data obtained from MIT’s eICU project, with the help of scientific advisor Dr. Collin Stultz.

The eICU research led to a critical insight: standard RBMs performed poorly on modeling data from different groups. Rather than capturing distinct groups of patients, RBMs produced distributions that were too smooth, blending together these patient groups. After much theoretical and experimental work, Charles, Jon, and Aaron solved this problem by inventing a new kind of neural network—a hybrid of an RBM and another popular machine-learning model called a generative adversarial network (GAN).

Around the same time, the founders also solved another key problem. They figured out how to train RBMs to model time-series data. This breakthrough happened following a research collaboration. Working with a major pharmaceutical company, the founders used clinical data to model disease progression in Alzheimer’s disease (AD) patients. By the spring of 2018, Unlearn’s training methods and ability to model AD had improved substantially. What began as challenges became Unlearn’s greatest strengths. Having handled difficult types of data, the founders now achieved their goal of generating realistic clinical records. 

Unlearn’s founders then needed to find the right business model within the clinical trials space. Although they had pitched the idea of optimizing clinical trial design to investors at first, they wanted to be more than a statistical consulting firm. So they came up with another idea. Unlearn already knew how to simulate people’s clinical records with machine-learning models. What if they used these models to replace significant parts of clinical trials? 

Thus, Unlearn’s business model—creating synthetic subjects for clinical trials—was born. 

Unlearn is currently building what Charles calls “cohorts of AI-generated subjects” for clinical trials testing different complex diseases. But the company’s long-term vision is much larger: a future when predictive models guide precision therapies. Charles hopes that one day every patient will have a digital twin, which will help physicians evaluate the risks each patient faces and develop the best course of treatment for that patient. 

Unlearn will become known as a pioneer for machine-learning tools in clinical trials. But it should also be known for the kind of technology the founders aspire to create. 

Aaron and Jon once described Charles as “compassionate,” a striking characteristic for a Bay Area start-up founder. So not surprisingly, Unlearn’s founders have built their technology with a compassionate mission.

Above all else, Unlearn works to give patients the treatments they need. 

Other Articles That Might Interest You