Artificial intelligence platforms like ChatGPT deliver answers that sound convincing, but they often fail to communicate their uncertainties or gaps in knowledge. This oversight can lead to significant issues, especially as AI becomes integral to processes like drug development, information synthesis, and autonomous driving.
Enter Themis AI, an MIT spinout dedicated to measuring model uncertainty and refining outputs to prevent larger complications. Their Capsa platform is versatile, compatible with any machine-learning model, and can swiftly identify and rectify unreliable outputs. By adjusting AI models to recognize patterns in data processing, Capsa helps uncover ambiguity, incompleteness, or bias.
“Our aim is to encapsulate a model, identify its uncertainties and potential failure points, and then enhance its functionality,” explains Themis AI co-founder and MIT Professor Daniela Rus, who leads the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). “We’re thrilled to provide solutions that can elevate model performance and ensure their correctness.”
Founded in 2021 by Rus along with Alexander Amini and Elaheh Ahmadi, former research affiliates in her lab, Themis AI has supported telecommunications companies with network automation and assisted oil and gas firms in interpreting seismic data. Their research has also focused on creating reliable chatbots.
“Our goal is to facilitate AI in high-stakes scenarios across various industries,” says Amini. “We’ve seen AI fail or hallucinate, and as its applications widen, these errors can have severe repercussions. Our software enhances transparency in these systems.”
Enhancing Model Confidence
For years, Rus’s lab has explored model uncertainty. A 2018 partnership with Toyota aimed to assess the reliability of AI in autonomous driving—a critical area where understanding model reliability is crucial.
“In safety-critical contexts, model reliability is paramount,” Rus states.
In another significant study, Rus, Amini, and collaborators created an algorithm to detect racial and gender bias in facial recognition systems, automatically adjusting the model’s training data to eliminate bias. This process involved identifying unrepresentative data segments and generating analogous samples to rebalance the dataset.
By 2021, Amini and his co-founders demonstrated that a similar technique could aid pharmaceutical companies in using AI for drug property predictions, leading to the establishment of Themis AI later that year.
“Optimizing drug discovery has the potential to drive substantial savings,” notes Rus. “That realization highlighted the power of this tool.”
Currently, Themis collaborates with various companies, many focusing on large language models (LLMs). Capsa empowers these models to assess their own uncertainty with each output.
“With many companies eager to harness LLMs tailored to their data, concerns about reliability loom large,” observes Stewart Jamieson, Themis AI’s head of technology. “We enable LLMs to self-report their confidence and uncertainties, which enhances reliable responses and identifies unreliable outputs.”
Themis AI is also engaging semiconductor companies developing AI solutions for off-cloud environments.
“Typically, smaller models used on devices may lack the accuracy of their server counterparts. However, we can provide efficient edge computing with low latency, without sacrificing quality,” Jamieson explains. “We envision a future where edge devices handle most processing, sending uncertain tasks to a central server.”
Pharmaceutical firms can leverage Capsa to refine AI models aimed at identifying drug candidates and predicting their clinical trial performance.
“These models’ predictions can be intricate and challenging to interpret, requiring significant effort from experts,” Amini remarks. “Capsa offers immediate insights to assess whether predictions are grounded in training data or merely speculative. This can expedite the identification of the most promising predictions, creating substantial societal benefits.”
Research with Purpose
The team at Themis AI believes they are ideally situated to enhance leading-edge AI technologies. They are currently investigating Capsa’s potential to improve accuracy in AI methodologies known as chain-of-thought reasoning, where LLMs elucidate the logic behind their answers.
“We’ve observed that Capsa may guide reasoning processes, identifying the most reliable chains of reasoning,” Amini asserts. “This holds significant potential for enhancing LLM experiences, reducing latency, and minimizing computational demands—an impactful opportunity for us.”
For Rus, who has co-founded several ventures during her tenure at MIT, Themis AI represents a chance to translate her academic research into real-world solutions.
“My students and I are increasingly committed to ensuring our work positively impacts the world,” Rus emphasizes. “AI has incredible potential to revolutionize industries, yet it also comes with challenges. What excites me is developing technical solutions to address these issues while building trust and understanding between people and the technology that permeates their daily lives.”
Photo credit & article inspired by: Massachusetts Institute of Technology
