How transferable are features in deep neural networks?
J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. (2014)cite arxiv:1411.1792Comment: To appear in Advances in Neural Information Processing Systems 27 (NIPS 2014).
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.
Description
How transferable are features in deep neural networks?
%0 Generic
%1 yosinski2014transferable
%A Yosinski, Jason
%A Clune, Jeff
%A Bengio, Yoshua
%A Lipson, Hod
%D 2014
%K dl
%T How transferable are features in deep neural networks?
%U http://arxiv.org/abs/1411.1792
%X 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.
@misc{yosinski2014transferable,
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.},
added-at = {2017-11-23T14:24:01.000+0100},
author = {Yosinski, Jason and Clune, Jeff and Bengio, Yoshua and Lipson, Hod},
biburl = {https://www.bibsonomy.org/bibtex/2241698deb82ab3c20a2f8180a250cecd/kmilian},
description = {How transferable are features in deep neural networks?},
interhash = {56e235c78f36ae27560500c4887cc4e8},
intrahash = {241698deb82ab3c20a2f8180a250cecd},
keywords = {dl},
note = {cite arxiv:1411.1792Comment: To appear in Advances in Neural Information Processing Systems 27 (NIPS 2014)},
timestamp = {2017-11-23T14:24:01.000+0100},
title = {How transferable are features in deep neural networks?},
url = {http://arxiv.org/abs/1411.1792},
year = 2014
}