Instructing Artificial Intelligence systems on unfamiliar matters
Rebooting AI Reliability: Meet Themis AI's Capsa Platform
Imagine a world where artificial intelligence (AI) systems could admit when they don't know something, or flag their own potential mistakes before they cause damage. Welcome to the future Themis AI is building with their Capsa platform. This ingenious system wraps around any machine-learning model, identifying uncertainties and fixing them in a blink of an eye. Let's dive into the fascinating world of Themis AI!
Their secret sauce? Recognizing the AI's weak points. By modifying models to detect patterns that hint at ambiguity, incompleteness, or bias in the data processing, Capsa efficiently corrects unreliable outputs. The genius behind this innovation is none other than Themis AI's co-founder and MIT Professor Daniela Rus, alongside Alexander Amini, Elaheh Ahmadi, and a team of research affiliates from her lab.
"We aim to take a model, dress it in Capsa, identify the uncertainties and failure modes, and then enhance the model," Rus explains. "Our solution is designed to improve AI models and offer guarantees, ensuring they're working correctly."
This duo of technology and academia has already made tremendous strides, aiding telecom companies in network planning and automation, and helping oil and gas companies understand seismic imagery with AI. And, let's face it, we've all seen AI going astray, making mistakes with devastating consequences. With Capsa, these "AI hallucinations" become a thing of the past, allowing any AI to forecast and predict its own failures before they happen!
Bridging the Knowledge Gap
Rus and her lab have been tackling model uncertainty since 2018, thanks to funding from Toyota. Their research focused on enhancing the reliability of an autonomous driving solution, a safety-critical context where understanding model reliability is crucial. In a separate project, they also developed an algorithm capable of detecting—and eliminating—race and gender bias in facial recognition systems.
Cleverly addressing drug discovery, they showed in 2021 that this approach could help pharmaceutical companies use AI to predict the properties of drug candidates, forming the basis for Themis AI. The vision? To empower AI in high-stakes applications across industries, ensuring safety and efficiency for all.
Today, Themis AI is partnering with enterprises from diverse sectors, including those developing large language models. The magic of Capsa? It enables these models to quantify their own uncertainty for each output, enhancing the quality and reliability of their responses.
Stepping Up for AI's Growing Pains
Stewart Jamieson, Head of Technology at Themis AI, explains the company's mission as addressing the growing concerns around AI reliability. "We are in discussions with semiconductor companies building AI solutions on their chips, helping them deliver AI solutions outside of cloud environments. Now, smaller models can work on phones or embedded systems, providing low latency while retaining quality. When they're unsure about their output, they can simply pass it to a central server."
Beyond improving AI systems, Themis AI's technology also aids pharmaceutical companies in identifying drug candidates and predicting their performance in clinical trials. By providing insights into the evidence backing a model's predictions or spotting speculative outputs with little grounding, Capsa can accelerate the identification of the strongest predictions and unlock new possibilities for societal good.
On the Cutting Edge
The Themis AI team is eager to expand Capsa's reach and impact. For example, they're exploring its capabilities in improving the accuracy of chain-of-thought reasoning, a technique used by large language models to explain the steps they take to reach an answer. By guiding these reasoning processes, the team anticipates significant improvements in latencies, computation requirements, and user experience.
For Rus, Themis AI is a beacon for her MIT research, bridging the gap between academia and practical applications. "I am excited about leveraging AI to solve real-world challenges and build trust between people and evolving technologies in their daily lives."
- In the world Themis AI is creating, AI systems would not only admit their limitations but also correct them, thus minimizing potential harm.
- The heart of Themis AI's innovation lies in identifying the AI's vulnerabilities by modifying models to detect patterns of ambiguity, incompleteness, or bias during data processing.
- The technology pioneered by Themis AI does not just foresee predicted failures in AI, but prevents them from happening by flagging uncertainties and enhancing model reliability.
- With Capsa, Themis AI has made significant strides in industries such as telecom, oil and gas, and drug discovery, helping AI systems make less erroneous predictions.
- Collaborating with MIT's Daniela Rus, MIT Professor, and her research affiliates, Themis AI is funded by entities like Toyota, focusing on enhancing model reliability in various high-stakes applications across industries.
- The Themis AI team is aggressively working on expanding Capsa's capabilities, aiming to enhance the accuracy of chain-of-thought reasoning in large language models, ultimately improving user experience and reducing computation requirements.
