“Fine-tuning existing neural network models on new data causes them to suffer from "catastrophic forgetting," where they lose knowledge from their original training.”