Despite the immense size and power of present-day artificial intelligence systems, they often struggle to differentiate between hallucination and reality. Autonomous driving systems can fail to recognize pedestrians and emergency vehicles right in front of them, leading to fatal consequences. Conversational AI systems confidently fabricate facts and, even after reinforcement learning, frequently provide inaccurate estimates of their own uncertainty.
To address these challenges, researchers from MIT and the University of California, Berkeley collaborated to develop a novel method for constructing advanced AI inference algorithms. This method simultaneously generates collections of probable explanations for data and accurately estimates the quality of these explanations.
The approach is based on a mathematical technique called sequential Monte Carlo (SMC), which has been widely used in uncertainty-calibrated AI. SMC algorithms propose probable explanations for data and track the likelihood of these explanations as more information becomes available. However, SMC has limitations when applied to complex tasks. One of the main issues is the process of generating guesses for probable explanations, which had to be overly simplified in previous versions of SMC. In intricate applications like self-driving, the ability to make plausible guesses based on raw data can be a challenging problem, requiring sophisticated algorithms beyond the capabilities of regular SMC.
This is where the new method, known as SMC with probabilistic program proposals (SMCP3), comes into play. SMCP3 enables the use of more intelligent ways to guess probable explanations, update those explanations with new information, and estimate their quality in sophisticated ways. It achieves this by allowing the utilization of any probabilistic program—a computer program that can make random choices—as a strategy for proposing intelligent guesses of data explanations. Unlike previous versions of SMC, which only allowed simple strategies with calculable probabilities, SMCP3 removes these restrictions and enables the use of multi-stage guessing procedures.
The researchers' SMCP3 paper demonstrates that by employing more sophisticated proposal procedures, SMCP3 improves the accuracy of AI systems in tracking 3D objects, analyzing data, and estimating the likelihood of the data itself. Previous research by MIT and others has shown that these estimates can be used to infer how well an inference algorithm explains data compared to an idealized Bayesian reasoner.
George Matheos, co-first author of the paper, expresses enthusiasm about SMCP3's potential to apply well-understood, uncertainty-calibrated algorithms in complex problem settings where previous versions of SMC were ineffective. He highlights the importance of building AI systems that are aware of their uncertainty as they make decisions across various domains to ensure reliability and safety.
Vikash Mansinghka, senior author of the paper, adds that Monte Carlo methods have been integral to computing and artificial intelligence since the inception of electronic computers. SMCP3 automates complex mathematics and expands the design space, enabling the creation of new AI algorithms that were previously unattainable.
The paper's other authors include co-first author Alex Lew, along with MIT PhD students Nishad Gothoskar, Matin Ghavamizadeh, and Tan Zhi-Xuan, and Stuart Russell, a professor at UC Berkeley. The research was presented at the AISTATS conference in Valencia, Spain, in April.