Deeply learned face representations are sparse, selective, and robust
Yi Sun
1
Xiaogang Wang
2,3
Xiaoou Tang
1,3
1
Department of Information Engineering, The Chinese University of Hong Kong
2
Department of Electronic Engineering, The Chinese University of Hong Kong
3
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
sy011@ie.cuhk.edu.hk xgwang@ee.cuhk.edu.hk xtang@ie.cuhk.edu.hk
Abstract
This paper designs a high-performance deep convo-
lutional network (DeepID2+) for face recognition. It
is learned with the identification-verification supervisory
signal. By increasing the dimension of hidden repre-
sentations and adding supervision to early convolutional
layers, DeepID2+ achieves new state-of-the-art on LFW
and YouTube Faces benchmarks.
Through empirical studies, we have discovered three
properties of its deep neural activations critical for the high
performance: sparsity, selectiveness and robustness. (1) It
is observed that neural activations are moderately sparse.
Moderate sparsity maximizes the discriminative power of
the deep net as well as the distance between images. It is
surprising that DeepID2+ still can achieve high recognition
accuracy even after the neural responses are binarized. (2)
Its neurons in higher layers are highly selective to identities
and identity-related attributes. We can identify different
subsets of neurons which are either constantly excited or
inhibited when different identities or attributes are present.
Although DeepID2+ is not taught to distinguish attributes
during training, it has implicitly learned such high-level
concepts. (3) It is much more robust to occlusions, although
occlusion patterns are not included in the training set.
1. Introduction
Face recognition achieved great progress thanks to ex-
tensive research effort devoted to this area [31, 33, 6, 24,
37, 27, 25, 23, 38]. While pursuing higher performance is
a central topic, understanding the mechanisms behind it is
equally important. When deep neural networks begin to ap-
proach human on challenging face benchmarks [27, 25, 23]
such as LFW [13], people are eager to know what has been
learned by these neurons and how such high performance is
achieved. In cognitive science, there are a lot of studies [30]
on analyzing the mechanisms of face processing of neurons
in visual cortex. Inspired by those works, we analyze the
Figure 1: Left: neural responses of DeepID2+ on images
of Bush and Powell. The second face is partially occluded.
There are 512 neurons in the top hidden layer of DeepID2+.
We subsample 32 for illustration. Right: a few neurons
are selected to show their activation histograms over all
the LFW face images (as background), all the images
belonging to Bush, all the images with attribute Male, and
all the images with attribute Female. A neuron is generally
activated on about half of the face images. But it may
constantly have activations (or no activation) for all the
images belonging to a particular person or attribute. In this
sense, neurons are sparse, and selective to identities and
attributes.
behaviours of neurons in artificial neural networks in a
attempt to explain face recognition process in deep nets,
what information is encoded in neurons, and how robust
they are to corruptions.
Our study is based on a high-performance deep convo-
lutional neural network (deep ConvNet [15, 16]), referred
to as DeepID2+, proposed in this paper. It is improved
upon the state-of-the-art DeepID2 net [23] by increasing the
dimension of hidden representations and adding supervision
to early convolutional layers. The best single DeepID2+
net (taking both the original and horizontally flipped face
images as input) achieves 98.70% verification accuracy on
1
资源评论
