A Review of Randomness Techniques in Deep Neural Networks

This paper investigates the effects of various randomization techniques on Deep Neural Networks (DNNs) learning performance. We categorize the existing randomness techniques into four key types: injection of noise/randomness at the data, model structure, optimization or learning stage. We use this classification to identify gaps in the current coverage of potential mechanisms for the introduction of randomness, leading to proposing two new techniques: adding noise to the loss function and random masking of the gradient updates. We use a Particle Swarm Optimizer (PSO) for hyperparameter optimization and evaluate over 30,000 configurations across standard computer vision benchmarks. Our study reveals that data augmentation and weight initialization randomness significantly improve performance, and different optimizers prefer distinct randomization types. The complete implementation and dataset are available on GitHub¹.

This paper for the Hot-off-the-Press track at GECCO 2024 summarizes the original work published at [2].