Why is it possible to use an AI trained on millions of photos to diagnose diseases it was never taught about?

The answer

Transfer learning is the technique of taking a model trained on one large task and repurposing it for a different, often smaller task. The key insight, demonstrated convincingly by Jason Yosinski and colleagues in 2014, is that early layers of deep neural networks learn general features (edges, textures, shapes) that are useful across many domains. By keeping these pre-trained layers and fine-tuning only the final layers on a new dataset, researchers can achieve excellent results with far less data and training time. Transfer learning democratized deep learning, allowing researchers with limited data and compute to build powerful models by standing on the shoulders of ImageNet-trained giants.