Teaching AI to handle uncertainty more effectively
Gennaro 骋补濒补鈥檚 research improved AI鈥檚 ability to handle uncertainty, making models more accurate, efficient, and reliable for critical applications like healthcare and finance.
Artificial Intelligence (AI) is increasingly being used in critical fields such as healthcare, finance, and science, where quick and accurate decisions are essential. However, many AI systems struggle with uncertainty, often making decisions based on rough estimates rather than fully understanding all possible outcomes. This can result in mistakes, inefficiencies, or even serious issues, particularly in high-risk scenarios like diagnosing illnesses or predicting financial trends.
Recognizing this challenge, PhD researcher Gennaro Gala aimed to improve how AI models reasoned about uncertainty, making them more precise, scalable, and trustworthy. He defended his thesis on Monday, March 24.
More precise and scalable AI models
骋补濒补鈥檚 work focused on a special class of AI models known as probabilistic circuits (PCs), which were designed to calculate probabilities with both accuracy and efficiency.
Unlike many existing AI techniques that relied on approximations, probabilistic circuits provided exact answers, making them particularly useful in fields where precision is essential.
Smarter learning with Bayesian scores
One of the key challenges in using probabilistic circuits was designing them in a way that balanced accuracy and efficiency.
Gala tackled this by introducing Bayesian scores, a new method that helped these models learn their structure in a more principled way.
This reduced the risk of overfitting, a situation where a model memorized data instead of learning meaningful patterns, leading to AI that generalized better and required less manual fine-tuning.
Scaling up with tensorized computation
Another breakthrough in his work was linking probabilistic circuits with a mathematical concept called tensor factorizations. This connection allowed these models to scale up, handling larger and more complex datasets without becoming too slow or inefficient.
By using modern computing techniques, probabilistic circuits could fully leverage powerful hardware like GPUs, making AI both faster and more capable.
Bridging probabilistic models and deep learning
A major contribution of this research was the development of Probabilistic Integral Circuits (PICs), an advanced extension of probabilistic circuits.
PICs introduced continuous variables, a key feature in deep learning models like variational autoencoders, and allowed probabilistic circuits to handle even more complex problems.
By bridging the gap between probabilistic models and deep learning, PICs enabled the training of large-scale AI models that retained the benefits of precise probability calculations.
A step toward more trustworthy AI
In summary, 骋补濒补鈥檚 research provided a solution to AI鈥檚 uncertainty problem by making probabilistic models more powerful, scalable, and reliable.
By combining structured probability models with deep learning techniques, his work paved the way for AI systems that were not only more efficient but also more trustworthy, a crucial step for AI in real-world, high-stakes applications.
Title of PhD thesis:
Supervisors: prof.dr. C. De Campos,
Media contact
More on AI and Data Science
