Artificial intelligence systems like Chatgpt provide plausible answers to any question you may ask. But they do not always reveal gaps in their knowledge or areas where they are uncertain. This problem can have enormous consequences because AI systems are increasingly used to do things such as developing drugs, synthesizing information and driving autonomous cars.
Now MIT Spinout Themis AI helps to quantify the uncertainty of the model and correct outings before causing greater problems. The company's CAPSA platform can work with any automatic learning model to detect and correct unreliable outings in seconds. It works by modifying AI models to allow them to detect models in their processing of data which indicate ambiguity, incompleteness or bias.
“The idea is to take a model, to wrap it in CAPSA, to identify the uncertainties and the failure methods of the model, then to improve the model,” explains the co -founder of Themis Ai and professor of MIT Daniela Rus, who is also director of IT and artificial intelligence laboratory (CSAIL). “We are delighted to offer a solution that can improve models and offer guarantees that the model is working properly.”
Rus founded Themis Ai in 2021 with Alexander Amini '17, SM '18, Phd '22 and Elaheh Ahmadi '20, Meng '21, two former affiliates of research in his laboratory. Since then, they have helped telecommunications companies with the planning and automation of networks, helped oil and gas companies to use AI to understand seismic imagery and have published articles on the development of more reliable and trustworthy chatbots.
“We want to activate AI in the highest applications in each industry,” says Amini. “We have all seen examples of hallucinating AI or making mistakes. As AI is deployed more widely, these errors could cause devastating consequences. Our software can make these systems more transparent. ”
Help models to know what they don't know
The RUS laboratory has been looking for the uncertainty of the model for years. In 2018, she received funding from Toyota to study the reliability of an autonomous driving solution based on automatic learning.
“It is a critical context of security where understanding the reliability of the model is very important,” explains Rus.
In separation workRUS, AIni, and their collaborators have built an algorithm that could detect racial and gender biases in facial recognition systems and automatically re-use the model training data, showing that it has eliminated biases. The algorithm worked by identifying the non-representative parts of the underlying training data and generating new samples of similar data to rebalance them.
In 2021, any co-founders showed a similar approach could be used to help pharmaceutical companies to use AI models to predict the properties of medicines. They founded Themis later that year.
“Guide the discovery of drugs could potentially save a lot of money,” said Rus. “This is the case of use that made us realize how powerful this tool could be.”
Today, Themis works with companies in a wide variety of industries, and many of these companies build large -language models. Using CAPSA, the models are able to quantify their own uncertainty for each output.
“Many companies are interested in using LLM that is based on their data, but they are concerned about reliability,” observes Stewart Jamieson SM '20, Phd '24, Themis of Themis AI technology. “We help LLM self-assess their confidence and uncertainty, which makes it possible to answer more reliable questions and report unreliable outings.”
Themis AI is also under discussion with semiconductor companies that build AI solutions on their chips that can operate outside cloud environments.
“Normally, these small models that work on phones or integrated systems are not very precise compared to what you could execute on a server, but we can make the most of the two worlds: low latency, efficient edge computer without sacrificing quality,” explains Jamieson. “We see a future where Edge devices do most of the work, but each time they are not sure of their release, they can transmit these tasks to a central server.”
Pharmaceutical companies can also use CAPSA to improve AI models used to identify drugs and predict their performance in clinical trials.
“The predictions and outings of these models are very complex and difficult to interpret – the experts spend a lot of time and efforts to try to give them meaning,” says Amini. “The CAPSA can give information as soon as the door is left to understand if the predictions are supported by evidence in the training set or are only speculations without much earthing.
Research for impact
Themis Ai's team believes that the company is well placed to improve the cutting -edge of AI technology constantly evolving. For example, the company explores the capacity of CAPSA to improve precision in a technique of AI known as the reasoning in chain of thoughts, in which the LLM explain the steps they take to obtain an answer.
“We have seen signs that Capsa could help guide these reasoning processes to identify the most confidence reasoning chains,” said Amini. “We believe that this has enormous implications in terms of improving LLM experience, reduction in latencies and reduction of calculation requirements. It is an extremely impact extremely an opportunity for us. ”
For Rus, who has co -founded several companies since his arrival at the MIT, Themis Ai is an opportunity to make sure that his research on MIT has an impact.
“My students and I have become more and more passionate about going to the additional step to make our work relevant to the world,” explains Rus. “AI has enormous potential to transform industries, but AI also raises concerns. What excites me is an opportunity to help develop technical solutions that raise these challenges and also strengthen trust and understanding between people and technologies that are part of their daily life. ”
