Artificial intelligence systems like ChatGPT provide plausible sounding answers to any questions you may ask. However, they do not always reveal uncertain knowledge or gaps in domain. AI systems are increasingly used to develop drugs, integrate information, drive self-driving cars, and more, so they can have great results.
Currently, MIT spinout Themis AI is quantifying model uncertainty and correcting the output before causing more problems. The company's CAPSA platform works with any machine learning model to detect and correct unreliable output in seconds. Modify the AI model to allow detection of patterns of data processing that indicate ambiguity, imperfection, or bias.
“The idea is to take the model, wrap it in a capsa, identify the model's uncertainty and failure modes, and enhance the model,” says Daniela Rus, MIT AI co-founder and professor of MIT AI, who is also director of the MIT Computer Science and Artificial Intelligence Institute (CSAIL). “We are excited to provide a solution that can improve our model and ensure that it is functioning properly.”
Rus founded Themis AI in 2021 along with Alexander Amini'17, SM'18, PhD'22, and Elaheh Ahmadi'20, Meng'21. Since then, they have supported telecom companies with network planning and automation, helped oil and gas companies use AI to understand earthquake images, and published papers on the development of more reliable and reliable chatbots.
“We want to enable AI in all the finest applications in the industry,” says Amini. “We have all seen examples of hallucinations and mistakes in AI, and these mistakes can have catastrophic consequences as AI is deployed more widely.
Help the model to know what they don't know
RUS labs have been investigating model uncertainty over the years. In 2018, she received funding from Toyota to study the reliability of machine learning-based autonomous driving solutions.
“It's a safety critical context where understanding the reliability of the model is very important,” Rus says.
In separate studies, RUS, Amini, and their collaborators detected racial and gender bias in facial recognition systems, automatically remeasured model training data, and created an algorithm that eliminated bias. This algorithm worked by identifying non-representative portions of the underlying training data and generating and rebalancing new similar data samples.
In 2021, the final co-founders showed that similar approaches could be used to help pharmaceutical companies use AI models to predict the characteristics of drug candidates. They founded Themis Ai later that year.
“Led drug discovery can potentially save a lot of money,” Rus says. “That was a use case that made me realize how powerful this tool is.”
Today, Themis AI works with companies from a variety of industries, many of which are building large language models. By using CAPSA, these models can quantify the unique uncertainty of each output.
“Many companies are interested in using data-based LLM, but are concerned about reliability,” said the technology director at Stewart Jamieson SM '20, PhD '24, Themis Ai. “We help LLMS self-report their confidence and uncertainty, which allows us to flag reliable questions and unreliable output.”
Themis AI is also discussing with semiconductor companies that build AI solutions on chips that could run outside of cloud environments.
“These small models, which normally run on mobile phones and embedded systems, are less accurate compared to what you can run on a server, but they can make the most of both worlds. Low latency, efficient edge computing without sacrificing quality,” explains Jamieson. “We're looking at the future where Edge devices do most of their work, but whenever we don't know the output, we can transfer those tasks to a central server.”
Pharmaceutical companies can also use CAPSA to improve AI models used to identify drug candidates and predict performance in clinical trials.
“The predictions and outputs of these models are extremely complex and difficult to interpret. Experts spend a lot of time and effort trying to understand them,” Amini said. “Capsa can quickly give insight from the gate to understand whether predictions are supported by evidence from the training set, or whether they are just a lot of unfounded guesses.
Research on impact
The Themis AI team believes the company is well positioned to improve the cutting edge of ever-evolving AI technology. For example, the company is investigating Capsa's ability to improve the accuracy of AI technology known as mindset inference that explains the steps LLM will take to reach an answer.
“We've seen that the indications that Capsa helps us guide these inference processes help us identify the best chain of confidence in reasoning,” says Jamieson. “I think it's a big deal in terms of improving the LLM experience, reducing latency, and reducing computational requirements. I think this is a very shocking opportunity for us.”
For RUS, who has co-founded several companies since coming to MIT, Themis AI is an opportunity to ensure that MIT research will have an impact.
“My students and I have become more and more passionate about taking the extra steps to relate our work to the world,” Rus says. “AI has a huge potential to transform industry, but AI also raises concerns. What excites me is the opportunity to develop technical solutions to address these challenges and build trust and understanding between people and technologies that are becoming part of everyday life.”
