Sydney researchers uncover mathematical cause of AI training inconsistency

Physicists at the University of Sydney have uncovered a mathematical explanation for why artificial intelligence models often produce different results even when trained on identical data under matching conditions. The study, conducted by Professor Pulin Gong and doctoral researcher Andrew Ly from the School of Physics, was published in Nature Machine Intelligence.

For years, software teams have struggled with training inconsistency. Most engineers attributed the problem to small fluctuations in hardware, floating-point arithmetic or the order of training data. The Sydney team wanted to see what happened when those external sources of randomness were removed entirely.

Fractal patterns in neural networks

The researchers ran approximately 500,000 training experiments across different network architectures and tasks, covering image classification and language processing. They found that starting configurations lead to complex structures known in physics as riddled basins. In these mathematical setups, starting points that produce one specific outcome sit right alongside points that lead to entirely different behaviours.

In testing an image recognition network, the team discovered that changing just one bit of data at the start changed the internal outcome roughly 48 per cent of the time. Increasing the precision of the starting inputs ten billion times failed to fix the divergence. While slowing down the learning rate did improve consistency, it reduced how effectively the model learned.

The authors stated that the findings do not mean AI outputs are completely random or that models cannot be evaluated. The results instead offer a mathematical foundation for understanding instability, which could help developers design architectures that depend less on exact starting conditions.

What this means for Australian businesses

Many Australian businesses treat artificial intelligence models like standard deterministic software, expecting the same training process to yield identical behaviour. This research shows that fine-tuning or training internal models comes with unpredictability built into the underlying maths.

Small and medium business operators can adapt by taking several practical steps:

• Do not rely on a single training run when customising a model for business operations.
• Build systematic test suites to benchmark fine-tuned models against set business tasks before rolling them into production.
• Avoid paying for excessive computational precision during fine-tuning, as extra precision alone will not eliminate training drift.
• Keep human approval checks in place for critical workflows such as finance, contracts or medical records where unpredictable decisions create real liability.