Click to freeze and unfreeze CNN layers. See how pretrained features transfer across tasks with dramatically less training data.
Sign in to save progress
Early layers of a CNN learn universal features useful for any vision task. These can be reused directly; only the final layers need retraining for your specific problem.
Choose a transfer learning strategy. It will configure the layers on the right.
Click any layer to toggle its state. The Classification Head is always a new layer.
Click frozen / fine-tuning layers to toggle state. The Classification Head is always new.
Transfer learning typically outperforms training from scratch, especially when labelled data is scarce, and the gap narrows as data grows.
Illustrative numbers showing the typical pattern, not measurements from a specific published benchmark. Exact accuracies depend on the task, model, and datasets.
Very few teams train large models from scratch. Foundation models like GPT-4 and CLIP are pretrained once, at enormous cost, and everyone else adapts them: fine-tuning them or reusing their learned representations for downstream tasks. ChatGPT itself is a pretrained base model adapted with instruction tuning and human feedback (RLHF). Reusing a pretrained model is not a shortcut; it is the standard practice.