API
< Back to all clusters
[TECHNOLOGY] · Australia · 5 sources

started · updated

University of Sydney researchers identify mathematical cause of AI unpredictability

Researchers from the University of Sydney have identified a mathematical reason why artificial intelligence systems can produce different behaviors even when trained under nearly identical conditions. The study, published in Nature Machine Intelligence by Professor Pulin Gong and PhD student Andrew Li, suggests that the unpredictability stems from the mathematical structure of the learning process itself rather than mere random factors.

By training approximately 500,000 neural networks, including those for text processing and image recognition, the team discovered that minute changes in initial parameters can lead to vastly different outcomes. In one instance involving a large image recognition network, changing a single bit in the initial settings resulted in a different internal configuration in roughly 48 percent of cases.

The researchers found that learning outcomes are distributed according to complex mathematical patterns known as fractals. This structure means that extremely similar starting conditions do not guarantee identical results. Notably, increasing the precision of initial settings did not resolve the issue, as even settings ten billion times more precise failed to measurably improve reproducibility.

Entities

Andrew Li · Nature Machine Intelligence · Pulin Gong · University of Sydney