Although today’s artificial intelligence systems possess immense size and capability, they frequently struggle to differentiate between what is real and what is a hallucination. For instance, autonomous driving systems can fatally overlook pedestrians and emergency vehicles positioned directly in their path. Similarly, conversational AI systems confidently fabricate information and often provide inaccurate assessments of their level of uncertainty, particularly after undergoing reinforcement learning.
However, a collaboration between researchers at MIT and the University of California, Berkeley, has yielded a novel approach to constructing advanced AI inference algorithms. This method enables the algorithms to generate multiple plausible explanations for data while also accurately gauging the quality of these explanations.
The newly developed method utilises a mathematical technique called Sequential Monte Carlo (SMC). SMC algorithms are commonly used for uncertainty-calibrated AI to propose likely explanations for data and assess their plausibility as more information becomes available.
However, SMC falls short when applied to complex tasks due to the simplicity of generating probable explanations. Particularly in challenging domains like self-driving, where analysing video data, identifying objects, and predicting hidden motion paths are involved, sophisticated algorithms are required to make plausible guesses. Regular SMC cannot support such advanced algorithms.
The newly introduced SMC method with probabilistic program proposals (SMCP3) addresses these limitations. SMCP3 enables more intelligent approaches to generate probable explanations for data, update them based on new information, and accurately estimate their quality. Unlike previous versions of SMC, which only allowed simple strategies with calculable probabilities, SMCP3 allows using any probabilistic program that incorporates random choices. This flexibility enables sophisticated guessing procedures with multiple stages, overcoming the previous restriction.
The research paper on SMCP3 demonstrates that employing advanced proposal procedures can enhance the precision of AI systems in tracking 3D objects, analysing data, and improving the algorithms’ estimations of data likelihood. Previous studies conducted by MIT and other institutions have revealed that these estimations can be utilised to infer the effectiveness of an inference algorithm in explaining data compared to an idealised Bayesian reasoner.
George Matheos, the first co-author of the paper and soon-to-be MIT EECS PhD student, expressed great enthusiasm for SMCP3’s ability to enable the practical application of well-established, uncertainty-calibrated algorithms in complex problem scenarios where previous versions of SMC were ineffective.
Today, many new algorithms propose explanations based on data but often lack uncertainty calibration and fail to consider alternatives or assess their explanations. SMCP3 offers the potential to use these algorithms more effectively by incorporating uncertainty calibration, ensuring trustworthy AI systems for reliable and safe decision-making.
Vikash Mansinghka, the paper’s Senior Author, further explains, “Monte Carlo methods have been fundamental in computing and artificial intelligence since the advent of electronic computers.” However, designing and implementing them has always been challenging, requiring manual derivation of mathematical equations and awareness of intricate mathematical constraints. SMCP3 automates these complex mathematical aspects while expanding the range of possible designs.