Transfer Learning

Using knowledge learned on one task or domain to help learn another, related task.

Transfer learning means applying a model or representation learned on a source task (usually data-rich) to a target task (usually data-scarce), most commonly by pretraining and then fine-tuning. Related research traces back to Bozinovski and colleagues' 1976 neural-network work; Yosinski and colleagues (2014) found that a network's lower layers learn more general features while higher layers grow more task-specific, and that initializing with transferred weights improves generalization even after further fine-tuning. Transfer isn't automatically beneficial: when the source and target are too dissimilar, it can actually hurt performance, called negative transfer. Embodied AI relies on transfer almost everywhere: turning an internet-pretrained VLM into a VLA, sim-to-real transfer, cross-embodiment transfer, and transferring from human video to a robot; domain adaptation is one sub-area of it.

ExampleGoogle DeepMind's RT-2 fine-tunes a vision-language model pretrained on web-scale image-text data together with robot trajectory data, writing actions as text tokens; the model can carry out instructions absent from its training data, such as “pick up the extinct animal,” performing about twice as well as baselines like RT-1 on generalization tests.

Also called

Knowledge Transfer

Related

Pre-training · Fine-tuning · Domain Adaptation · Sim-to-Real Transfer · Positive / Negative Transfer · RT-2

Sources

Transfer learning (Wikipedia)
How transferable are features in deep neural networks? (Yosinski et al., arXiv 1411.1792)
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control (project page)