SydneyNear-identical training can leave AI systems behaving differently.
Some AI unpredictability is rooted in the mathematics of learning, not hardware or engineering flaws.
Starting settings form riddled basins, fractals where every point lies arbitrarily close to points leading to another outcome.
Sharper starting values may do little to predict where training ends, the study suggests.
How each outlet framed it
- Phys.org
- attributes AI training unpredictability to the fractal mathematics of starting conditions, not hardware or engineering flaws
Sources: Phys.org