Randomness In Neural Network Training: Characterizing The Impact of Tooling

Overall accuracy can mask performance disparities: While standard metrics like top-1 accuracy remain largely unaffected by randomness in neural network training, the model's performance on specific subgroups within the data distribution is "far more sensitive to the introduction of randomness," suggesting that relying solely on aggregate metrics may hide important fairness and safety issues.

Tooling choices matter as much as algorithmic design: The paper shifts focus from algorithmic sources of randomness to an understudied area—how software tools and hardware accelerators introduce unintended randomness into training. This suggests that responsible AI requires attention to implementation details, not just theoretical design choices.

Determinism has significant but variable costs: Achieving deterministic training (necessary for reproducibility and safety) carries substantial computational overhead, ranging from 196% to 746% depending on the GPU architecture used, meaning the practical feasibility of deterministic AI systems varies dramatically based on hardware choices.

Deterministic tooling is critical for AI safety: The authors argue that deterministic training is essential for responsible AI development, likely because non-determinism makes it harder to identify, reproduce, and fix problematic model behaviors—a key requirement for safe AI deployment.

Hardware and architecture choices have outsized impacts: The dramatic variation in determinism costs across different GPU types and neural network architectures suggests that responsible AI governance may need to consider hardware and tooling standardization, not just model design.