Reproducing Deep Double Descent: A Beginner's Journey
2025-06-05
A machine learning novice at the Recurse Center embarked on a journey to reproduce the deep double descent phenomenon. Starting from scratch, they trained a ResNet18 model on the CIFAR-10 dataset, exploring the impact of varying model sizes and label noise on model performance. The process involved overcoming challenges such as model architecture adjustments, correct label noise application, and understanding accuracy metrics. Ultimately, they successfully reproduced the deep double descent phenomenon, observing the influence of model size and training epochs on generalization ability, and the significant role of label noise in the double descent effect.
Read more