How transferable are features in deep neural networks?

TLDR

Deep neural networks trained on natural images learn first‑layer features resembling Gabor filters and color blobs that are general across datasets and tasks, yet the transition to task‑specific features in deeper layers has been little studied. This study experimentally quantifies the trade‑off between generality and specificity of neurons in each layer of a deep convolutional network and reports several surprising findings. The authors evaluate transferability by measuring performance when transferring features from different layers of an ImageNet‑trained network to target tasks, analyzing how specialization and optimization affect transfer. They find that higher‑layer specialization and optimization difficulties reduce transferability, that the effect depends on the layer transferred, that transferability declines with task distance but can still beat random features, and that initializing with transferred features from many layers yields a lasting generalization boost even after fine‑tuning.

Abstract

Many deep neural networks trained on natural images exhibit a curious phenomenon in common: on the first layer they learn features similar to Gabor filters and color blobs. Such first-layer features appear not to be specific to a particular dataset or task, but general in that they are applicable to many datasets and tasks. Features must eventually transition from general to specific by the last layer of the network, but this transition has not been studied extensively. In this paper we experimentally quantify the generality versus specificity of neurons in each layer of a deep convolutional neural network and report a few surprising results. Transferability is negatively affected by two distinct issues: (1) the specialization of higher layer neurons to their original task at the expense of performance on the target task, which was expected, and (2) optimization difficulties related to splitting networks between co-adapted neurons, which was not expected. In an example network trained on ImageNet, we demonstrate that either of these two issues may dominate, depending on whether features are transferred from the bottom, middle, or top of the network. We also document that the transferability of features decreases as the distance between the base task and target task increases, but that transferring features even from distant tasks can be better than using random features. A final surprising result is that initializing a network with transferred features from almost any number of layers can produce a boost to generalization that lingers even after fine-tuning to the target dataset.