Transfer learning reuses a pre-trained network as a feature extractor to train a classifier from few examples

Transfer learning reuses a neural network already trained on a large dataset (e.g. ImageNet) as a fixed feature extractor, then trains only a small classifier on top of it using new, task-specific examples. For artists this is practical: you need tens of examples per class rather than millions, and a standard laptop can run inference fast enough for real-time performance. ml4a’s openFrameworks tools operationalize this — ConvnetOSC sends a frozen convnet’s activations as a 4096-value vector over OSC, and a lightweight classifier (trained interactively, e.g. via Wekinator) maps that vector to a class label or regression value streamed to any creative environment. The key misconception: fine-tuning ALL weights is not required; freezing the backbone and training only the head is almost always sufficient for domain-adapted classification.

Examples

Use ml4a’s ConvnetOSC to stream webcam convnet activations over OSC, train a classifier in Wekinator with 20 examples per gesture, and route the predicted label to Hydra to switch a visual scene.

Assessment

Explain why transfer learning needs far fewer examples than training from scratch, and identify which part of the network you retrain versus freeze.