Unsupervised Feature Learning via Non-Parametric Instance-level
Discrimination
Z. Wu, Y. Xiong, S. Yu, и D. Lin. (2018)cite arxiv:1805.01978Comment: CVPR 2018 spotlight paper. Code: https://github.com/zhirongw/lemniscate.pytorch.
Аннотация
Neural net classifiers trained on data with annotated class labels can also
capture apparent visual similarity among categories without being directed to
do so. We study whether this observation can be extended beyond the
conventional domain of supervised learning: Can we learn a good feature
representation that captures apparent similarity among instances, instead of
classes, by merely asking the feature to be discriminative of individual
instances? We formulate this intuition as a non-parametric classification
problem at the instance-level, and use noise-contrastive estimation to tackle
the computational challenges imposed by the large number of instance classes.
Our experimental results demonstrate that, under unsupervised learning
settings, our method surpasses the state-of-the-art on ImageNet classification
by a large margin. Our method is also remarkable for consistently improving
test performance with more training data and better network architectures. By
fine-tuning the learned feature, we further obtain competitive results for
semi-supervised learning and object detection tasks. Our non-parametric model
is highly compact: With 128 features per image, our method requires only 600MB
storage for a million images, enabling fast nearest neighbour retrieval at the
run time.
Описание
Unsupervised Feature Learning via Non-Parametric Instance-level
Discrimination
%0 Generic
%1 wu2018unsupervised
%A Wu, Zhirong
%A Xiong, Yuanjun
%A Yu, Stella
%A Lin, Dahua
%D 2018
%K ml
%T Unsupervised Feature Learning via Non-Parametric Instance-level
Discrimination
%U http://arxiv.org/abs/1805.01978
%X Neural net classifiers trained on data with annotated class labels can also
capture apparent visual similarity among categories without being directed to
do so. We study whether this observation can be extended beyond the
conventional domain of supervised learning: Can we learn a good feature
representation that captures apparent similarity among instances, instead of
classes, by merely asking the feature to be discriminative of individual
instances? We formulate this intuition as a non-parametric classification
problem at the instance-level, and use noise-contrastive estimation to tackle
the computational challenges imposed by the large number of instance classes.
Our experimental results demonstrate that, under unsupervised learning
settings, our method surpasses the state-of-the-art on ImageNet classification
by a large margin. Our method is also remarkable for consistently improving
test performance with more training data and better network architectures. By
fine-tuning the learned feature, we further obtain competitive results for
semi-supervised learning and object detection tasks. Our non-parametric model
is highly compact: With 128 features per image, our method requires only 600MB
storage for a million images, enabling fast nearest neighbour retrieval at the
run time.
@misc{wu2018unsupervised,
abstract = {Neural net classifiers trained on data with annotated class labels can also
capture apparent visual similarity among categories without being directed to
do so. We study whether this observation can be extended beyond the
conventional domain of supervised learning: Can we learn a good feature
representation that captures apparent similarity among instances, instead of
classes, by merely asking the feature to be discriminative of individual
instances? We formulate this intuition as a non-parametric classification
problem at the instance-level, and use noise-contrastive estimation to tackle
the computational challenges imposed by the large number of instance classes.
Our experimental results demonstrate that, under unsupervised learning
settings, our method surpasses the state-of-the-art on ImageNet classification
by a large margin. Our method is also remarkable for consistently improving
test performance with more training data and better network architectures. By
fine-tuning the learned feature, we further obtain competitive results for
semi-supervised learning and object detection tasks. Our non-parametric model
is highly compact: With 128 features per image, our method requires only 600MB
storage for a million images, enabling fast nearest neighbour retrieval at the
run time.},
added-at = {2018-05-23T20:57:14.000+0200},
author = {Wu, Zhirong and Xiong, Yuanjun and Yu, Stella and Lin, Dahua},
biburl = {https://www.bibsonomy.org/bibtex/23249efe92416bc55699c05f9a6ae6643/ste2526},
description = {Unsupervised Feature Learning via Non-Parametric Instance-level
Discrimination},
interhash = {dc5213f25c501b421da28532bc2f0af1},
intrahash = {3249efe92416bc55699c05f9a6ae6643},
keywords = {ml},
note = {cite arxiv:1805.01978Comment: CVPR 2018 spotlight paper. Code: https://github.com/zhirongw/lemniscate.pytorch},
timestamp = {2018-05-23T20:57:14.000+0200},
title = {Unsupervised Feature Learning via Non-Parametric Instance-level
Discrimination},
url = {http://arxiv.org/abs/1805.01978},
year = 2018
}