Generalization
Generalization is a model's ability to perform well on unseen data, not just the training set. It is the ultimate goal of ML and depends on data diversity, model capacity, and regularization.
What is Generalization?
Generalization is a model's ability to perform well on unseen data, not just the training set. It is the ultimate goal of ML and depends on data diversity, model capacity, and regularization.
Generalization is a model's ability to perform well on unseen data, not just the training set. It is the ultimate goal of ML and depends on data diversity, model capacity, and regularization.
Where is it used?
GPT-4 generalizes to novel reasoning tasks because it was trained on diverse data at scale; benchmarks like MMLU and GPQA measure generalization to held-out distributions.
How to build it
Compare test-set accuracy across models trained with different dropout and data augmentation levels to empirically observe which regularizers improve generalization.