Society’s reliance on artificial intelligence (AI) and machine learning (ML) applications continues to grow, redefining how information is consumed. From AI-powered chatbots to information syntheses produced from Large Language Models (LLMs), society has access to more information and deeper insights than ever before. However, as technology companies race to implement AI across their value chain, a critical question looms. Can we really trust the outputs of AI solutions?
Can we really trust AI outputs without uncertainty quantification
For a given input, a model might have generated many other equally-plausible outputs. This could be due to insufficient training data, variations in the training data, or other causes. When deploying models, organizations can leverage uncertainty quantification to provide their end-users with a clearer understanding of how much they should trust the output of an AI/ML model. Uncertainty quantification is the process of estimating what those other outputs could have been.
Imagine a model predicting tomorrow’s high temperature. The model might generate the output 21ºC, but uncertainty quantification applied to that output might indicate that the model could just as well have generated the outputs 12 ºC, 15 ºC, or 16 ºC; knowing this, how much do we now trust the simple prediction of 20 ºC? Despite its potential to engender trust or to counsel caution, many organizations are choosing to skip uncertainty quantification because of the additional work they need to do to implement it, as well as because of its demands on computing resources and inference speed.
Human-in-the-loop systems, such as medical diagnosis and prognosis systems, involve humans as part of the decision-making process. By blindly trusting the data of healthcare AI/ML solutions, healthcare professionals risk misdiagnosing a patient, potentially leading to sub-par health outcomes—or worse. Uncertainty quantification can allow healthcare professionals to see, quantitatively, when they can place more trust in the outputs of AI and when they should treat specific predictions with caution. Similarly, in a fully-automated system such as a self-driving car, the output of a model for estimating the distance of an obstacle could lead to a crash that might have been otherwise avoided in the presence of uncertainty quantification on the distance estimate.
The challenge of leveraging Monte Carlo methods to build trust in AI/ML models
Monte Carlo methods, developed during the Manhattan Project, are a robust way to perform uncertainty quantification. They involve re-running algorithms repeatedly with slightly different inputs until further iterations do not provide much more information in the outputs; when the process reaches such a state, it is said to have converged. One disadvantage of Monte Carlo methods is that they are typically slow and compute-intensive, requiring many repetitions of their constituent computations to obtain a converged output and have an inherent variability across those outputs. Because Monte Carlo methods use the outputs of random number generators as one of their key building blocks, even when you run a Monte Carlo with many internal repetitions, the results you obtain will change when you repeat the process with identical parameters.
The path forward to trustworthiness in AI/ML models
Unlike traditional servers and AI-specific accelerators, a new breed of computing platforms are being developed to directly process empirical probability distributions in the same way that traditional computing platforms process integers and floating-point values. By deploying their AI models on these platforms, organizations can automate the implementation of uncertainty quantification on their pre-trained models and can also speed up other kinds of computing tasks that have traditionally used Monte Carlo methods, such as VaR calculations in finance. In particular, for the VaR scenario, this new breed of platforms allows organizations to work with empirical distributions built directly from real market data, rather than approximating these distributions with samples generated by random number generators, for more accurate analyses and faster results.
Recent breakthroughs in computing have significantly lowered the barriers to uncertainty quantification. A recent research article published by my colleagues and I, in the Machine Learning With New Compute Paradigms workshop at NeurIPS 2024, highlights how a next-generation computation platform we developed enabled uncertainty quantification analysis to run over 100-fold faster compared to running traditional Monte-Carlo-based analyses on a high-end Intel-Xeon-based server. Advances such as these allow organizations deploying AI solutions to implement uncertainty quantification with ease and to run such uncertainty quantification with low overheads.
The future of AI/ML trustworthiness depends on advanced next-generation computation
As organizations integrate more AI solutions into society, trustworthiness in AI/ML will become a top priority. Enterprises can no longer afford to skip implementing facilities in their AI model deployments to allow consumers to know when to treat specific AI model outputs with skepticism. The demand for such explainability and uncertainty quantification is clear, with approximately three in four people indicating they would be more willing to trust an AI system if appropriate assurance mechanisms were in place.
New computing technologies are making it ever easier to implement and deploy uncertainty quantification. While industry and regulatory bodies grapple with other challenges associated with deploying AI in society, there is at least an opportunity to engender the trust humans require, by making uncertainty quantification the norm in AI deployments.
AI trust,monte carlo methods,Signaloid,trust
