How does transfer learning work for regression tasks? Can someone point to an application where transfer learning has been successfully applied for regression tasks.
Consider the denoising autoencoder. The input features to a neural network are contaminated by a small amount of noise, sent through one or more intermediate layers, and then through a final layer of the same size as the input layer. This network is optimized to reconstruct the original data which can be seen as a form of regularization.
The resulting network weights (minus the final layer) are frozen and transferred to a supervised learning task (perhaps added to another neural network). It has been found that the features constructed by this process are quite valuable. This applies to regression or classification.