An AI startup needs to build a model that classifies rare medical skin lesions. Because they only have a few hundred labeled images, training a deep convolutional neural network from scratch would lead to severe overfitting. Instead, they choose to take a network pre-trained on the massive ImageNet dataset, freeze its early layers, and fine-tune the final layer on their skin lesion images. What is this machine learning practice called?

Transfer Learning is correct because it is a machine learning technique where a model developed for a source task is reused as the starting point for a model on a second, related target task. This is highly beneficial in deep learning because the early layers of neural networks trained on large datasets, such as ImageNet for vision or Wikipedia corpora for NLP, learn general features like edges, textures, shapes, or basic grammar. By leveraging these pre-trained weights and fine-tuning only the later layers on a smaller, task-specific dataset, developers can achieve high accuracy with significantly less training data and computational time. In this scenario, the startup freezes early ImageNet layers and fine-tunes the final layer on a few hundred skin lesion images, which is exactly transfer learning and avoids severe overfitting from training a deep convolutional neural network from scratch. The Model Distillation choice is incorrect because distillation, or knowledge distillation, is the process of transferring knowledge from a large, complex model, often called a teacher, to a smaller, more efficient model, often called a student, to reduce computational footprint. The Data Transfer Protocol choice is incorrect because it refers to networking protocols, such as HTTP or FTP, used to transmit data across systems, not a machine learning training methodology. The Zero-shot Inference choice is incorrect because zero-shot learning or inference involves using a model to classify or predict on tasks it has never seen without any gradient updates or fine-tuning of its weights.