“Deep learning models often exhibit catastrophic interference, where training on new data degrades performance on previously learned tasks, which is a form of poor generalization.”