Deeply learned face reprentations are spar,lective,and robust
Yi Sun1,Xiaogang Wang2,3,Xiaoou Tang1,3
1Department of Information Engineering,The Chine University of Hong Kong.2Department of Electronic Engineering,The Chine University of Hong Kong.3Shenzhen Institute of Advanced T echnology,Chine Academy of Sciences.
Face recognition achieved great progress thanks to extensive rearch effort devoted to this area.While pursuing higher performance is a central topic, understanding the mechanisms behind it is equally important.When deep neural networks begin to approach human on challenging face benchmark-s[3,4,5]such as LFW[2],people are eager to know what has been learned by the neurons and how such high performance is achieved.In cognitive science,there are a lot of studies[7]on analyzing the mechanisms of face processing of neurons in visual cortex.Inspired by tho works,we analyze the behaviours of neurons in artificial neural networks in a attempt to ex-plain face recognition process in deep nets,what information is encoded in neurons,and how robust they are to corruptions.形容冷的词
Our study is bad on a high-performance deep convolutional neural network(deep ConvNet),referred to as DeepID2+,propod in this paper. It is improved upon the state-of-the-art DeepID2net[3]by increa
我国世界遗产>鸡肉的英语sing the dimension of hidden reprentations and adding supervision to early con-volutional layers.The best single DeepID2+net(taking both the original and horizontallyflipped face images as input)achieves98.70%verification accuracy on LFW(vs.96.72%by DeepID2).Combining25DeepID2+nets ts new state-of-the-art on multiple benchmarks:99.47%on LFW for face verification(vs.99.15%by DeepID2[3]),95.0%and80.7%on LFW for clod-and open-t face identification,respectively(vs.82.5%and61.9% by Web-Scale Training(WST)[6]),and93.2%on YouTubeFaces[8]for face verification(vs.91.4%by DeepFace[5]).
With the state-of-the-art deep ConvNets and through extensive empiri-cal evaluation,we investigate three properties of neural activations crucial for the high performance:sparsity,lectiveness,and robustness.They are naturally owned by DeepID2+after large scale training on face data,and we did NOT enforce any extra regularization to the model and training process to achieve them.Therefore,the results are valuable for understanding the intrinsic properties of deep networks.
It is obrved that the neural activations of DeepID2+are moderately spar.As examples shown in Fig.1,for an input face image,around half of the neurons in the top hidden layer are activated.On the other hand,each neuron is activated on roughly half of the face images.Such sparsity dis-tributions c
an maximize the discriminative power of the deep net as well as the distance between images.Different identities have different subts of neurons activated.Two images of the same identity have similar activation patterns.This motivates us to binarize the neural respons in the top hid-den layer and u the binary code for recognition.Its result is surprisingly good.Its verification accuracy on LFW only slightly drops by1%or less. It has significant impact on large-scale face arch since huge storage and computation time is saved.This also implies that binary activation patterns are more important than activation magnitudes in deep neural networks.
童年作文450字
Related to sparness,it is also obrved that neurons in higher layers are highly lective to identities and identity-related attributes.When an i-dentity(who can be outside the training data)or attribute is prented,we can identify a subt of neurons which are constantly excited and also can find another subt of neurons which are constantly inhibited.A neuron from any of the two subts has strong indication on the existence/non-existence of this identity or attribute,and our experiment show that the s-ingle neuron alone has high recognition accuracy for a particular identity or attribute.In other words,neural activations have sparsity on identities and attributes,as examples shown in Fig.1.Although DeepID2+is not taught to distinguish attributes during training,it has implicitly learned such high-level concepts.Directly employing the face reprentation learned by De
epID2+leads to much higher classification accuracy on identity-related attributes than widely ud handcrafted features such as high-dimensional LBP[1].
This is an extended abstract.The full paper is available at the Computer Vision Foundation webpage
.Figure1:Left:neural respons of DeepID2+on images of Bush and Pow-ell.The cond face is partially occluded.There are512neurons in the top hidden layer of DeepID2+.We subsample32for illustration.Right:a few neurons are lected to show their activation histograms over all the LFW face images(as background),all the images belonging to Bush,all the im-ages 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 partic-ular person or attribute.In this n,neurons are spar,and lective to identities and attributes.
Our empirical study shows that neurons in higher layers are much more robust to image corruption in face recognition than handcrafted features such as high-dimensional LBP or neurons in lower layers.As an example shown in Fig.1,when a face image is partially occluded,its binary acti-vation patterns remain stable,although the magnitudes could change.We conjecture the reason might be that neurons in higher layers capture global features and are less nsitive to local variations.DeepID2+is trained by natural web face images and no artificial occlusion patterns were added to the training t.
炭炉
圆体积计算公式[1]Dong Chen,Xudong Cao,Fang Wen,and Jian Sun.Blessing of di-
一脉相承的意思mensionality:High-dimensional feature and its efficient compression for face verification.In Proc.CVPR,2013.
[2]Gary B.Huang,Manu Ramesh,Tamara Berg,and Erik Learned-Miller.
Labeled Faces in the Wild:A databa for studying face recognition in unconstrained environments.Technical Report07-49,University of Massachutts,Amherst,2007.
[3]Yi Sun,Yuheng Chen,Xiaogang Wang,and Xiaoou Tang.Deep learn-
ing face reprentation by joint identification-verification.In Proc.NIP-S,2014.
[4]Yi Sun,Xiaogang Wang,and Xiaoou Tang.Deep learning face repre-
ntation from predicting10,000class.In Proc.CVPR,2014.
[5]Y.Taigman,M.Yang,M.Ranzato,and L.Wolf.DeepFace:Closing the
gap to human-level performance in face verification.In Proc.CVPR, 2014.
呼风唤雨的世纪教案
[6]Y.Taigman,M.Yang,M.Ranzato,and L.Wolf.Web-scale training for
face identification.Technical report,arXiv:1406.5266,2014.
[7]Doris Y.Tsao and Margaret S.Livingstone.Neural mechanisms for face
perception.Annu Rev Neurosci,31:411–438,2008.
[8]Lior Wolf,Tal Hassner,and Itay Maoz.Face recognition in uncon-
strained videos with matched background similarity.In Proc.CVPR, 2011.
