《人工智能、机器学习与大数据》课程教学资源（参考文献）Semi-Supervised Deep Hashing with a Bipartite Graph

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) Semi-Supervised Deep Hashing with a Bipartite Graph Xinyu Yan,Lijun Zhang,Wu-Jun Li National Key Laboratory for Novel Software Technology,Nanjing University,Nanjing 210023,China yanxy,zhanglj@lamda.nju.edu.cn,liwujun @nju.edu.cn Abstract (LSH)[Gionis et al.,1999]and data-dependent ones like Iter- ative Quantization (ITQ)Gong et al.,2013].Spectral Hash- Recently,deep learning has been successfully ap- plied to the problem of hashing,yielding remark- ing(SH)[Weiss et al,2009],Anchor Graph Hashing (AGH) able performance compared to traditional method- [Liu et al.,2011].On the other hand,to deal with more com- plicated semantic similarity,supervised hashing methods are s with hand-crafted features.However,most of proposed to exploit label information to improve the hash- existing deep hashing methods are designed for ing quality.Representative supervised methods include La- the supervised scenario and require a large num- tent Factor Hashing (LFH)[Zhang et al.,2014],Fast Super- ber of labeled data.In this paper,we propose vised Hashing (FastH)[Lin et al.,2014],Supervised Discrete a novel semi-supervised hashing method for im- age retrieval,named Deep Hashing with a Bipar- Hashing(SDH)[Shen et al.,2015].However,labeling large- scale image dataset is inefficient and time-consuming.As a tite Graph(BGDH),to simultaneously learn em- result,Semi-Supervised Hashing (SSH)[Wang et al.,2012] beddings,features and hash codes.More specifi- has been developed to make use of labeled data as well as the cally,we construct a bipartite graph to discover the abundant unlabeled data. underlying structure of data,based on which an em- bedding is generated for each instance.Then,we In traditional hashing methods,images are represented feed raw pixels as well as embeddings to a deep by hand-crafted features such as GIST [Oliva and Torralba. neural network,and concatenate the resulting fea- 2001],and the choice of features requires heavy manual in- tures to determine the hash code.Compared to ex- terventions.Motivated from the great success of deep neu- isting methods,BGDH is a universal framework ral networks in image analysis [Krizhevsky et al.,2012],re- that is able to utilize various types of graphs and cently some deep hashing methods have been proposed to losses.Furthermore,we propose an inductive vari- learn features and hash codes simultaneously [Li et al..2016: ant of BGDH to support out-of-sample extensions Zhu et al..2016:Liu et al..2016.Although those deep hash- Experimental results on real datasets show that our ing methods yield better performance compared with the tra- BGDH outperforms state-of-the-art hashing meth- ditional methods,they usually need a large number of labeled ods instances as training data.To address this limitation,a semi- supervised deep hashing,named SSDH,have been develope- d [Zhang et al.,2016].SSDH is fundamentally built upon 1 Introduction graph-based semi-supervised learning [Zhou et al.,2004]and With the explosion in the volume of image data,it has been the loss function contains a graph regularization term which raised as a big challenge about how to index and organize involves both the labeled and unlabeled data.In theory,SS- these data efficiently and accurately.Approximate Nearest DH needs to construct a nearest neighbor graph of all the da- Neighbor (ANN)search [Indyk and Motwani.1998]has be- ta.Unfortunately,this step takes O(n2)time,where n is the come a popular way to retrieve content from images with both number of instances,and thus intractable for large scale data. computational efficiency and search quality.Among existing In this paper,we propose a novel semi-supervised hash- ANN search methods.hashing is advantageous due to its fast ing method,named Deep Hashing with a Bipartite Graph query speed and low memory complexity [Gionis et al.,1999; (BGDH),which performs graph embedding,feature learning Gong et al.,2013].It aims to transform high-dimensional im- and hash code learning in a unified framework.First,we con- ages into a set of short binary codes while maintaining simi- struct a bipartite graph to capture the information hidden in larity of the original data. the labeled and unlabeled data.The bipartite graph could be Generally speaking,hashing methods can be divided into a semantic graph that describes relationships between images two categories:unsupervised and supervised.Unsupervised and concepts,an anchor graph that describes similarities be- methods utilize some kinds of distance metrics to learn a hash tween images and landmarks [Liu et al.,2010],or a tradition- function from unlabeled data.Methods in this category in- al nearest neighbor graph.Then,inspired by the recent work clude data-independent ones like Locally Sensitive Hashing on graph embedding [Yang et al.,20161,we learn an embed- 3238

Semi-Supervised Deep Hashing with a Bipartite Graph Xinyu Yan, Lijun Zhang, Wu-Jun Li National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China {yanxy, zhanglj}@lamda.nju.edu.cn, liwujun@nju.edu.cn Abstract Recently, deep learning has been successfully ap￾plied to the problem of hashing, yielding remark￾able performance compared to traditional method￾s with hand-crafted features. However, most of existing deep hashing methods are designed for the supervised scenario and require a large num￾ber of labeled data. In this paper, we propose a novel semi-supervised hashing method for im￾age retrieval, named Deep Hashing with a Bipar￾tite Graph (BGDH), to simultaneously learn em￾beddings, features and hash codes. More specifi- cally, we construct a bipartite graph to discover the underlying structure of data, based on which an em￾bedding is generated for each instance. Then, we feed raw pixels as well as embeddings to a deep neural network, and concatenate the resulting fea￾tures to determine the hash code. Compared to ex￾isting methods, BGDH is a universal framework that is able to utilize various types of graphs and losses. Furthermore, we propose an inductive vari￾ant of BGDH to support out-of-sample extensions. Experimental results on real datasets show that our BGDH outperforms state-of-the-art hashing meth￾ods. 1 Introduction With the explosion in the volume of image data, it has been raised as a big challenge about how to index and organize these data efficiently and accurately. Approximate Nearest Neighbor (ANN) search [Indyk and Motwani, 1998] has be￾come a popular way to retrieve content from images with both computational efficiency and search quality. Among existing ANN search methods, hashing is advantageous due to its fast query speed and low memory complexity [Gionis et al., 1999; Gong et al., 2013]. It aims to transform high-dimensional im￾ages into a set of short binary codes while maintaining simi￾larity of the original data. Generally speaking, hashing methods can be divided into two categories: unsupervised and supervised. Unsupervised methods utilize some kinds of distance metrics to learn a hash function from unlabeled data. Methods in this category in￾clude data-independent ones like Locally Sensitive Hashing (LSH) [Gionis et al., 1999] and data-dependent ones like Iter￾ative Quantization (ITQ) [Gong et al., 2013], Spectral Hash￾ing (SH) [Weiss et al., 2009], Anchor Graph Hashing (AGH) [Liu et al., 2011]. On the other hand, to deal with more com￾plicated semantic similarity, supervised hashing methods are proposed to exploit label information to improve the hash￾ing quality. Representative supervised methods include La￾tent Factor Hashing (LFH) [Zhang et al., 2014], Fast Super￾vised Hashing (FastH) [Lin et al., 2014], Supervised Discrete Hashing (SDH) [Shen et al., 2015]. However, labeling large￾scale image dataset is inefficient and time-consuming. As a result, Semi-Supervised Hashing (SSH) [Wang et al., 2012] has been developed to make use of labeled data as well as the abundant unlabeled data. In traditional hashing methods, images are represented by hand-crafted features such as GIST [Oliva and Torralba, 2001], and the choice of features requires heavy manual in￾terventions. Motivated from the great success of deep neu￾ral networks in image analysis [Krizhevsky et al., 2012], re￾cently some deep hashing methods have been proposed to learn features and hash codes simultaneously [Li et al., 2016; Zhu et al., 2016; Liu et al., 2016]. Although those deep hash￾ing methods yield better performance compared with the tra￾ditional methods, they usually need a large number of labeled instances as training data. To address this limitation, a semi￾supervised deep hashing, named SSDH, have been develope￾d [Zhang et al., 2016]. SSDH is fundamentally built upon graph-based semi-supervised learning [Zhou et al., 2004] and the loss function contains a graph regularization term which involves both the labeled and unlabeled data. In theory, SS￾DH needs to construct a nearest neighbor graph of all the da￾ta. Unfortunately, this step takes O(n 2 ) time, where n is the number of instances, and thus intractable for large scale data. In this paper, we propose a novel semi-supervised hash￾ing method, named Deep Hashing with a Bipartite Graph (BGDH), which performs graph embedding, feature learning and hash code learning in a unified framework. First, we con￾struct a bipartite graph to capture the information hidden in the labeled and unlabeled data. The bipartite graph could be a semantic graph that describes relationships between images and concepts, an anchor graph that describes similarities be￾tween images and landmarks [Liu et al., 2010], or a tradition￾al nearest neighbor graph. Then, inspired by the recent work on graph embedding [Yang et al., 2016], we learn an embed￾Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) 3238

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence(IJCAI-17) ding for each instance to predict the neighborhood context in point xi into a c-dimensional binary code bi=H(xi)= the graph.Finally,we feed both raw pixels and embeddings [hi(xi),h2(xi),...,he(xi)]TE{-1,1)c.The binary codes to a deep neural network,and concatenate the corresponding B=[bshould preserve the semantic similarity and hidden layers when producing binary codes.BGDH is a gen- structure similarity in the Hamming space. eral learning framework in the sense that any loss function of hashing and any type of graph can be incorporated. 3 Semi-Supervised Deep Hashing with a Graph-based methods are usually transductive,because they can only handle instances that are already appeared in Bipartite Graph the graph.Since embeddings of instances are learnt from In this section,we first present the details of our semi- the graph,the basic BGDH is also transductive.To address supervised Deep Hashing with a Bipartite Graph (BGDH), the out-of-sample problem,we further propose an inductive then introduce an inductive variant and finally discuss the variant of BGDH,in which the embeddings are defined as learning procedure. a parametric function of the raw features.In this way,we can produce hash codes for new instances that have not seen 3.1 The Proposed BGDH Framework during training.To demonstrate the effectiveness of our ap- The end-to-end deep learning architecture of our BGDH is proach,we conduct extensive experiments on two large-scale shown in Figure 1,which includes three main components: datasets:CIFAR-10 and NUS-WIDE.Experimental result- graph embedding,feature learning,and hash code learning. s show that BGDH outperforms other methods and achieves Similar to other semi-supervised learning methods [Yang et the state-of-the-art performance in image retrieval. al.,2016],the loss function of BGDH can be expressed as Finally,we emphasize that although both BGDH and SS- DH are semi-supervised deep hashing methods,the proposed C.+λCg (1) BGDH differs from SSDH in the following two aspects: where Cs is a supervised loss designed to preserve the simi- a. While SSDH is built upon graph regularization,our larity between pairwise instances,and Co is an unsupervised BGDH relies on graph embedding. loss of predicting the graph context.In the following,we b.SSDH uses graphs to exploit the unlabeled data,in con- first introduce Co which aims to learn embeddings from the tract BGDH makes use of bipartite graphs,which can bipartite graph g,then formulate Cs which is used to leam be constructed more efficiently since building an anchor both features and binary codes from hidden layers of deep graph only costs O(n)time. networks. 2 Notations and Problem Definitions Graph embedding We propose to use a bipartite graph g=(,O,E)to cap- In this section,we introduce notations and problem defini- ture the information hidden in the unlabeled data.It can be tions. constructed in different ways as stated below. 2.1 Notations An anchor graph constructed from the dataset t.In this case,O contains m landmarks and the construction of G We use script letters like to denote sets,boldface lowercase takes O(n)time [Liu et al.,20101. letters like e to denote vectors and boldface uppercase letters .A nearest neighbor graph.In this case,=Y and the like E to denote matrices.We denote the element at the i-th construction of G takes O(n2)time. row and j-th column of E by Eii.ET is the transpose of E. A semantic graph constructed from external data.In this l2 denotes the Euclidean norm of a vector and F denotes case,O may contain m concepts,styles,or owners. the Frobenius norm of a matrix.sgn()is an element-wise In the following,we briefly introduce one way to construct an sign function and o()is the sigmoid function.u;v]denotes anchor graph.We first randomly sample m instances from the concatenation of two vectors u and v. as landmarks,denoted by =fo1,...,om}.Then,we put an edge between xi and o;if o;is among k nearest landmarks 2.2 Problem Definitions of x;,or if the distance between them is smaller than some Given a set of n instances/images ={xi where xi is threshold e.Let A E Rxm be the similarity matrix of g, the feature vector of the i-th instance.Without loss of gener- where A:;denotes the weight of the edge between xi and oj. ality,we assume the first l instances [x1,...,x}are labeled The value of Aij may be binary,that is,Ai=1 if there is and the rest are unlabeled.We assume the supervised infor- an edge,otherwise 0.If a real value is preferred,Aij can be mation is given in term of pairwise labels,though our method set according to the heat kernel:Aij=e/,where can also support other kinds of labels.Specifically,for the P>0 is a parameter. first I instances,we have a set of pairwise labels S={sij} The goal of graph embedding is to learn an embedding for with sij E[0,1,where sij =1 means that xi and xj are each instance that predicts the context in the graph.Given an similar,sij=0 means that xi and xj are dissimilar.In ad- instance and its context,the objective of graph embedding is dition,a bipartite graph g=(,O,E)between n instances usually formulated as minimizing certain loss of predicting and m objects are given,where is a set of objects such as the context using the embedding of an instance as input fea- concepts or landmarks,and is the set of edges. ture [Weston et al.,2012;Mikolov et al.,2013].The context The goal of our semi-supervised hashing method is to learn of an instance can be simply defined as its neighbors in the a mapping function H:xi-1,1c,which encodes each graph,or generated by sophisticated methods such as random 3239

ding for each instance to predict the neighborhood context in the graph. Finally, we feed both raw pixels and embeddings to a deep neural network, and concatenate the corresponding hidden layers when producing binary codes. BGDH is a gen￾eral learning framework in the sense that any loss function of hashing and any type of graph can be incorporated. Graph-based methods are usually transductive, because they can only handle instances that are already appeared in the graph. Since embeddings of instances are learnt from the graph, the basic BGDH is also transductive. To address the out-of-sample problem, we further propose an inductive variant of BGDH, in which the embeddings are defined as a parametric function of the raw features. In this way, we can produce hash codes for new instances that have not seen during training. To demonstrate the effectiveness of our ap￾proach, we conduct extensive experiments on two large-scale datasets: CIFAR-10 and NUS-WIDE. Experimental result￾s show that BGDH outperforms other methods and achieves the state-of-the-art performance in image retrieval. Finally, we emphasize that although both BGDH and SS￾DH are semi-supervised deep hashing methods, the proposed BGDH differs from SSDH in the following two aspects: a. While SSDH is built upon graph regularization, our BGDH relies on graph embedding. b. SSDH uses graphs to exploit the unlabeled data, in con￾tract BGDH makes use of bipartite graphs, which can be constructed more efficiently since building an anchor graph only costs O(n) time. 2 Notations and Problem Definitions In this section, we introduce notations and problem defini￾tions. 2.1 Notations We use script letters like X to denote sets, boldface lowercase letters like e to denote vectors and boldface uppercase letters like E to denote matrices. We denote the element at the i-th row and j-th column of E by Eij . ET is the transpose of E. k·k2 denotes the Euclidean norm of a vector and k·kF denotes the Frobenius norm of a matrix. sgn(·) is an element-wise sign function and σ(·) is the sigmoid function. [u; v] denotes the concatenation of two vectors u and v. 2.2 Problem Definitions Given a set of n instances/images X = {xi} n i=1 where xi is the feature vector of the i-th instance. Without loss of gener￾ality, we assume the first l instances {x1, . . . , xl} are labeled and the rest are unlabeled. We assume the supervised infor￾mation is given in term of pairwise labels, though our method can also support other kinds of labels. Specifically, for the first l instances, we have a set of pairwise labels S = {sij} with sij ∈ {0, 1}, where sij = 1 means that xi and xj are similar, sij = 0 means that xi and xj are dissimilar. In ad￾dition, a bipartite graph G = (X , O, E) between n instances and m objects are given, where O is a set of objects such as concepts or landmarks, and E is the set of edges. The goal of our semi-supervised hashing method is to learn a mapping function H : xi → {−1, 1} c , which encodes each point xi into a c-dimensional binary code bi = H(xi) = [h1(xi), h2(xi), · · · , hc(xi)]T ∈ {−1, 1} c . The binary codes B = {bi} n i=1 should preserve the semantic similarity and structure similarity in the Hamming space. 3 Semi-Supervised Deep Hashing with a Bipartite Graph In this section, we first present the details of our semi￾supervised Deep Hashing with a Bipartite Graph (BGDH), then introduce an inductive variant and finally discuss the learning procedure. 3.1 The Proposed BGDH Framework The end-to-end deep learning architecture of our BGDH is shown in Figure 1, which includes three main components: graph embedding, feature learning, and hash code learning. Similar to other semi-supervised learning methods [Yang et al., 2016], the loss function of BGDH can be expressed as Ls + λLg (1) where Ls is a supervised loss designed to preserve the simi￾larity between pairwise instances, and Lg is an unsupervised loss of predicting the graph context. In the following, we first introduce Lg which aims to learn embeddings from the bipartite graph G, then formulate Ls which is used to learn both features and binary codes from hidden layers of deep networks. Graph embedding We propose to use a bipartite graph G = (X , O, E) to cap￾ture the information hidden in the unlabeled data. It can be constructed in different ways as stated below. • An anchor graph constructed from the dataset X . In this case, O contains m landmarks and the construction of G takes O(n) time [Liu et al., 2010]. • A nearest neighbor graph. In this case, O = X and the construction of G takes O(n 2 ) time. • A semantic graph constructed from external data. In this case, O may contain m concepts, styles, or owners. In the following, we briefly introduce one way to construct an anchor graph. We first randomly sample m instances from X as landmarks, denoted by O = {o1, . . . , om}. Then, we put an edge between xi and oj if oj is among k nearest landmarks of xi , or if the distance between them is smaller than some threshold . Let A ∈ R n×m be the similarity matrix of G, where Aij denotes the weight of the edge between xi and oj . The value of Aij may be binary, that is, Aij = 1 if there is an edge, otherwise 0. If a real value is preferred, Aij can be set according to the heat kernel: Aij = e −kxi−oj k 2 2/ρ, where ρ > 0 is a parameter. The goal of graph embedding is to learn an embedding for each instance that predicts the context in the graph. Given an instance and its context, the objective of graph embedding is usually formulated as minimizing certain loss of predicting the context using the embedding of an instance as input fea￾ture [Weston et al., 2012; Mikolov et al., 2013]. The context of an instance can be simply defined as its neighbors in the graph, or generated by sophisticated methods such as random Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) 3239

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) Feature Learning Hash Code Learning labels Predict pairwise Labeled Training Set Graph Embedding Graph context Predict graph Embeddings Prediction Loss Figure 1:The end-to-end deep learning architecture of BGDH. walk [Perozzi et al.,2014].Following previous studies [Yang Algorithm 1 Context generation based on random walk et al.,2016],we present a simple procedure for context gen- 1:Input:A bipartite graph g=(,O,E),parameters r eration in Algorithm 1,where the parameter d is the length and d of the random walk and r E(0,1)determines the ratio of positive contexts. 2:Uniformly sample an instance i from 3:Uniformly sample a random variable u from [0,1] By invoking Algorithm 1 t times,we obtain a set of triples 4:if u<r then (ijcwhere 1 indicates node cj is a context y←+1 of node ij,and =-1 indicates cj is not a context.Let 6 ei,be the embedding of node ij,and we,be the parameters Uniformly sample a random walk sequence S of length d started from i for predicting node c;as a context.We define the objective 7: Uniformly sample a context c from S except i function of graph embedding 8:else t 9: y←-1 Cg= (ei,we,:7j) (2) 10: Uniformly sample context c from&' 11 11:end if 12:return(i,c,y） where e(,)is a loss that measures the discrepancy between ei We,and Y.In machine learning,the following losses are commonly used. network is presented in Table 1,and a detailed explanation ●The square loss can be found in [Li et al.,2016].The output of the last fea- ewg)=(-ewg)月 ture learning layer(full7)of labeled instance xi is represent- ed by (xi;0),where 6 denotes all the parameters in the seven layers of feature learning part.In contrast with exist- ·The logistic loss ing supervised hashing methods,we also learn features from embeddings of labeled instances.The output associated with (ewe,)=log(1+e-yeiWey） embedding ei is denoted by (ei;0e),where be contains all the parameters in hidden layers.In this paper,we only add To constrain the solution space,we may further impose sparse one fully-connected layer for embeddings,of which the size constraints or nonnegative constraints [Lee and Seung,1999]. is determined by the dimension of embeddings. Feature learning and hash code learning We concatenate (xi;0)and (ei;be)as a new feature We utilize deep neural network model to learn features from for instance i,then send it to a hash code learning layer as: the raw pixels and embeddings of labeled instances,and then ui M [o(xi;0z);(ei;0)]+v (3) combine them together to learn the hash codes.BGDH con- tains a CNN model to learn features from raw image pixels, where M E R(4096+d)xe denotes a weight matrix,d is the and the model has seven layers as those of CNN-F [Chatfield dimension ofembedding,and vERex is a bias vector.Note et al.,2014]while other networks like AlexNet [Krizhevsky that any supervised loss function of hashing can be used in and Hinton,2009]can be used too.The configuration of the our framework to learn parameters M and v.In this paper,we 3240

Labeled Training Set Graph Feature Learning Hash Code Learning Predict pairwise labels Embeddings Prediction Loss Predict graph context labeled unlabeled {ui} ψ(e, θe) Graph Embedding φ(x,θx) Figure 1: The end-to-end deep learning architecture of BGDH. walk [Perozzi et al., 2014]. Following previous studies [Yang et al., 2016], we present a simple procedure for context gen￾eration in Algorithm 1, where the parameter d is the length of the random walk and r ∈ (0, 1) determines the ratio of positive contexts. By invoking Algorithm 1 t times, we obtain a set of triples {(ij , cj , γj )} t j=1 where γj = 1 indicates node cj is a context of node ij , and γj = −1 indicates cj is not a context. Let eij be the embedding of node ij , and wcj be the parameters for predicting node cj as a context. We define the objective function of graph embedding Lg = 1 t Xt j=1 `(e > ijwcj , γj ) (2) where `(·, ·) is a loss that measures the discrepancy between e > ijwcj and γj . In machine learning, the following losses are commonly used. • The square loss `(e > ijwcj , γj ) =  γj − e > ijwcj 2 • The logistic loss `(e > ijwcj , γj ) = log  1 + e −γje > ij wcj  To constrain the solution space, we may further impose sparse constraints or nonnegative constraints [Lee and Seung, 1999]. Feature learning and hash code learning We utilize deep neural network model to learn features from the raw pixels and embeddings of labeled instances, and then combine them together to learn the hash codes. BGDH con￾tains a CNN model to learn features from raw image pixels, and the model has seven layers as those of CNN-F [Chatfield et al., 2014] while other networks like AlexNet [Krizhevsky and Hinton, 2009] can be used too. The configuration of the Algorithm 1 Context generation based on random walk 1: Input: A bipartite graph G = (X , O, E), parameters r and d 2: Uniformly sample an instance i from X 3: Uniformly sample a random variable u from [0, 1] 4: if u < r then 5: γ ← +1 6: Uniformly sample a random walk sequence S of length d started from i 7: Uniformly sample a context c from S except i 8: else 9: γ ← −1 10: Uniformly sample context c from X 11: end if 12: return (i, c, γ) network is presented in Table 1, and a detailed explanation can be found in [Li et al., 2016]. The output of the last fea￾ture learning layer (full7) of labeled instance xi is represent￾ed by φ(xi ; θx), where θx denotes all the parameters in the seven layers of feature learning part. In contrast with exist￾ing supervised hashing methods, we also learn features from embeddings of labeled instances. The output associated with embedding ei is denoted by ψ(ei ; θe), where θe contains all the parameters in hidden layers. In this paper, we only add one fully-connected layer for embeddings, of which the size is determined by the dimension of embeddings. We concatenate φ(xi ; θx) and ψ(ei ; θe) as a new feature for instance i, then send it to a hash code learning layer as: ui = MT [φ(xi ; θx); ψ(ei ; θe)] + v (3) where M ∈ R (4096+d)×c denotes a weight matrix, d is the dimension of embedding, and v ∈ R c×1 is a bias vector. Note that any supervised loss function of hashing can be used in our framework to learn parameters M and v. In this paper, we Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) 3240

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence(IJCAI-17) Table 1:Configuration of the feature learning network inductive hashing model can be written as: Layer Configuration convl filter64×11×1l,stride4×4,pad0,LRN,pool2×2 conv2 filter 256x5x5,stride 1x1,pad 2,LRN,pool 2x2 c。+λcg=-∑(8,6j-log(1+e9)月 conv3 filter256×3×3,stride1×l,padl 845∈S conv4 filter256x3×3,stride1×1,padI conv5 filter256x3×3,stride1x1,pad1,pool2×2 +n∑Ilb,-(MP[(x;0)片(5(x;0z):0e月+v)l full6 4096 i=1 full7 4096 ∑ewl choose the loss function of deep pairwise-supervised hashing (DPSH)[Li et al.,2016],and Cs is given by (7) We can predict the hash code of any point xas: C。=-∑(s,6-log(1+e9)+n∑Ib-ul3 ba sgn(M [o(xg:0);(E(xg;v);0e)]+v). (8) 8y∈S i=l (4) 3.3 Learning where j=uuj and n>0 is a regularization parameter. In the transductive version of BGDH,the optimization vari- By substituting Eq.(3)into Eg.(4),we obtain the final loss ables include M,v.[bi},0r,0e,fei}and {we}.We adopt function of the supervised part: stochastic gradient descent(SGD)[Bottou,2010]to train our model. c=-∑ (s6-log(1+e9)】 First,we sample a batch of labeled instances of which set 8时∈S I contains indexes.A gradient step is then taken to optimize the supervised loss Cs.For all i Z1,bi can be directly +b:-(MT[(x:0):(e::0c)]+v)2 optimized as follow: i=1 (5) bi=sgn(ui)=sgn (MT [o(xi;0);u(e;:0e)]+v)(9) The objective of BGDH For other parameters in Cs,i.e.,M,v,0r,e,and fei:iE We combine the supervised part and unsupervised part to for- I1},we use back propagation(BP)to optimize them.Deriva- m the transductive version of BGDH.From(2)and(5),the tives of Cs w.r.t.ui are presented as follows: loss function of BGDH is OC.=1 1 C.+=->e-log(1+e0u)) Oui 2 ∑(a-s4+2∑(a-s叫 j:84y∈S j:8∈S 84∈S +2m(u-b) (10) +n∑lb,-(MP(x;0z(e:9e】+v)l where aij=a(uuj).We can then update other parame- i=1 ters according to +∑em OC (11) j=1 i=[(x;0)片(e:0e】 (6) where A >0 is a constant weighting factor.The first t- ot._OC. wo terms are the loss of predicting pairwise labels and the (12) third one is the loss of predicting context.As a result,our BGDH can simultaneously learn embeddings,features,and 8Cs OCs hash codes.During the training phase,semantic similarity 3o(x:0)Ov(ei;0e) =MOC, u (13) can affect graph embeddings,at the same time structure of Then,we perform a gradient step to optimize the unsuper- data also influence the prediction of pairwise labels. vised graph embedding loss Ca calculated by sampling triples 3.2 An Inductive Variant generated in Algorithm 1.In this case parameters fei,we (i,c)E Z2}will be updated where Z2 denotes the set con- Note that the basic BGDH is transductive,because the em- taining indexes of sampled instances and contexts.The above beddings of instances are learnt from the graph.Since the procedures are repeated for T1 and T2 iterations respectively. hash code of an instance depends on both the raw pixels and its embedding,we need to design a way to infer the embed- The whole learning algorithm of BGDH is summarized in ding of a unseen instance.To this end,we insert hidden layers Algorithm 2.Notice that before training jointly,we first train to connect the raw pixels and embedding [Yang et al.,2016] unsupervised part for a number of iterations to learn the ini- and in this way,the embedding e;becomes a parameterized tialization embeddings [e).For the inductive variant,we function of xi,denoted by (xi;).The loss function of will update parameters instead of embeddings [ei}. 3241

Table 1: Configuration of the feature learning network Layer Configuration conv1 filter 64×11×11, stride 4×4, pad 0, LRN, pool 2×2 conv2 filter 256×5×5, stride 1×1, pad 2, LRN, pool 2×2 conv3 filter 256×3×3, stride 1×1, pad 1 conv4 filter 256×3×3, stride 1×1, pad 1 conv5 filter 256×3×3, stride 1×1, pad 1, pool 2×2 full6 4096 full7 4096 choose the loss function of deep pairwise-supervised hashing (DPSH) [Li et al., 2016], and Ls is given by Ls = − X sij∈S ￾ sijΘij − log(1 + e Θij )
+ η X l i=1 kbi − uik 2 2 (4) where Θij = 1 2 u T i uj and η > 0 is a regularization parameter. By substituting Eq. (3) into Eq. (4), we obtain the final loss function of the supervised part: Ls = − X sij∈S ￾ sijΘij − log(1 + e Θij )
+ η X l i=1 bi − (MT [φ(xi ; θx); ψ(ei ; θe)] + v) 2 2 . (5) The objective of BGDH We combine the supervised part and unsupervised part to for￾m the transductive version of BGDH. From (2) and (5), the loss function of BGDH is Ls + λLg = − X sij∈S ￾ sijΘij − log(1 + e Θij )
+ η X l i=1 bi − (MT [φ(xi ; θx); ψ(ei ; θe)] + v) 2 2 + λ t Xt j=1 `(e > ijwcj , γj ) (6) where λ > 0 is a constant weighting factor. The first t￾wo terms are the loss of predicting pairwise labels and the third one is the loss of predicting context. As a result, our BGDH can simultaneously learn embeddings, features, and hash codes. During the training phase, semantic similarity can affect graph embeddings, at the same time structure of data also influence the prediction of pairwise labels. 3.2 An Inductive Variant Note that the basic BGDH is transductive, because the em￾beddings of instances are learnt from the graph. Since the hash code of an instance depends on both the raw pixels and its embedding, we need to design a way to infer the embed￾ding of a unseen instance. To this end, we insert hidden layers to connect the raw pixels and embedding [Yang et al., 2016], and in this way, the embedding ei becomes a parameterized function of xi , denoted by ξ(xi ; ϑx). The loss function of inductive hashing model can be written as: Ls + λLg = − X sij∈S ￾ sijΘij − log(1 + e Θij )
+ η X l i=1 bi − ￾MT [φ(xi ; θx); ψ(ξ(xi ; ϑx); θe)] + v
2 2 + λ t Xt j=1 `(e > ijwcj , γj ) (7) We can predict the hash code of any point xq ∈ X/ as: bq = sgn(MT [φ(xq; θx); ψ(ξ(xq; ϑx); θe)] + v). (8) 3.3 Learning In the transductive version of BGDH, the optimization vari￾ables include M, v, {bi}, θx, θe, {ei} and {wc}. We adopt stochastic gradient descent (SGD) [Bottou, 2010] to train our model. First, we sample a batch of labeled instances of which set I1 contains indexes. A gradient step is then taken to optimize the supervised loss Ls. For all i ∈ I1, bi can be directly optimized as follow: bi = sgn(ui) = sgn ￾MT [φ(xi ; θx); ψ(ei ; θe)] + v
(9) For other parameters in Ls, i.e., M, v, θx, θe, and {ei : i ∈ I1}, we use back propagation (BP) to optimize them. Deriva￾tives of Ls w. r. t. ui are presented as follows: ∂Ls ∂ui = 1 2 X j:sij∈S (aij − sij )uj + 1 2 X j:sji∈S (aji − sji)uj + 2η(ui − bi) (10) where aij = σ( 1 2 u T i uj ). We can then update other parame￾ters according to ∂Ls ∂M = [φ(xi ; θx); ψ(ei ; θe)]  ∂Ls ∂ui T , (11) ∂Ls ∂v = ∂Ls ∂ui , (12) ∂Ls ∂φ(xi ; θx) = ∂Ls ∂ψ(ei ; θe) = M ∂Ls ∂ui . (13) Then, we perform a gradient step to optimize the unsuper￾vised graph embedding loss Lg calculated by sampling triples generated in Algorithm 1. In this case parameters {ei , wc : (i, c) ∈ I2} will be updated where I2 denotes the set con￾taining indexes of sampled instances and contexts. The above procedures are repeated for T1 and T2 iterations respectively. The whole learning algorithm of BGDH is summarized in Algorithm 2. Notice that before training jointly, we first train unsupervised part for a number of iterations to learn the ini￾tialization embeddings {ei}. For the inductive variant, we will update parameters ϑx instead of embeddings {ei}. Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) 3241

Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence(IJCAI-17) Algorithm 2 Learning algorithm for BGDH [Lin et al.,20141.SDH [Shen et al.,2015]and COSDISH 1:Input:A bipartite graph G=(.O.)images [Kang et al.,2016];supervised deep hashing methods:DPSH {xi}是l,pairwise labels S={si},parametersn,入， [Li et al.,2016],DHN [Zhu et al.,2016]and DSH [Liu et al., batch iterations T1,T and sizes N,N2 2016];semi-supervised deep methods SSDH [Zhang et al., 2:Output:Binary codes B=(bi} 2016]. 3:Initialization:Initialize 0r with the pre-trained CNN-F For hashing methods using hand-crafted features,we rep- model on ImageNet;Initialize each entry of M,v,and e resent each image in CIFAR-10 by a 512-dimensional GIST by randomly sampling from a Gaussian distribution with vector.In NUS-WIDE,an image is represented by a 1134- mean 0 and variance 0.01. dimensional feature vector,including 500-D bag-of-words 4:REPEAT features,64-D color histogram,144-D color correlogram,73- 5:fort←1toT1do D edge direction histogram,128-D wavelet texture,225-D block-wise color moments.For deep methods,we use the 6: Randomly sample Ni labeled images,and let ZI be the set containing indexes of sampled instances raw image pixels as input.The network architectures of these 7: Calculate o(xi;0)and (ei:0e)for all i E I by deep models are different from each other,for fair compar- forward propagation ison,we adopt the same deep network architecture CNN-F 8: Compute ui M[o(xi;0);(e:;0e)]+v and the with the same initialization parameters pre-trained on Im- binary code of xi with bi=sgn(ui)for all i Z ageNet [Deng et al.,2009]for all deep hashing methods. 9: Compute the derivative of Cs w.r.t.fui:iZ1} The bipartite graph of BGDH is constructed based on hand- 10: Update parameters M,v,0,0e,and [e;:iE T1} crafted features with heat kemnel,where the hyper-parameter by back propagation p is set as 1 for CIFAR-10 and 10 for NUS-WIDE.The hyper- 11:end for parameter n in BGDH is set as 10 for CIFAR-10 and 100 for 12:fort←-1toT2do NUS-WIDE similar to DPSH [Li et al.,2016].We simply set 13: Randomly generate a batch of triples by invoking Al- T1 =10,T2 =5,0.1 in all the experiments. gorithm 1 N2 times,and let Z2 be the set containing in- 4.2 Experimental Results dexes of sampled instances and contexts 14: Compute the derivative of o w.r.t.e;,we:(i,c)E To evaluate the retrieval accuracy,we use the Mean Average Z Precision (MAP)as the evaluation metric.Following Lai et 15: Update parameters [ei,we:(i,c)EZ2) al.,2015;Xia et al.,2014],1000 images (100 images per 16:end for class)are randomly selected as the query set in CIFAR-10. 17:UNTIL stopping We use all the rest of images as training set for the unsu- 18:Calculate bi sgn (MT[o(xi;0);(e:;0e)]+v)for pervised methods and for building a neighbour graph g in i=1,...,n BGDH.We randomly select 5000(500 images per class)im- ages as the training set for supervised methods. In NUS-WIDE,we randomly sample 2100 query images 4 Experiments from 21 most frequent labels following the strategy in [Lai et al.,2015;Xia et al.,2014].We use all the rest of images as In this section,we present experimental results.All the ex- training set for the unsupervised methods.To reduce the time periments are performed on a NVIDIA K80 GPU server with complexity,we only use 5000 landmarks sampled randomly MatConvNet [Vedaldi and Lenc,2014]. to build a bipartite graph for BGDH.For supervised methods, we randomly select 10500 images from dataset.The MAP 4.1 Datasets and Setting values are calculated within the top 5000 returned neighbors We conduct experiments on two widely used benchmark We present the MAP results in Table 2 where BGDH.SS- datasets:CIFAR-10 and NUS-WIDE.The CIFAR-10 dataset DH,DSH,DHN,DPSH are deep methods.Except for SSDH consists of 60,000 images from 10 classes (6000 images per which uses triplet labels,all deep methods are trained with class).It is a single-label dataset.The NUS-WIDE dataset2 pairwise labels.To be fair,parameters of these methods are contains 269,648 images from Flickr.Following the settings set according to the suggestions of authors.The results show in [Lai et al.,2015],we use 19,5834 images belonging to the that in most cases our proposed BGDH method substantially 21 most frequent classes and each class consists at least 5000 outperforms other baselines including unsupervised methods, images.Additionally,two images are defined as a ground- supervised methods and semi-supervised methods truth neighbor when they share at least one common label. In real applications,the number of labeled instances is usu- In our experiments,BGDH-T denotes transductive mod- ally far less than that of unlabeled instances.To further verify el while BGDH-I denotes inductive one.We compare our the effectiveness of leveraging both pairwise labels and un- method with several state-of-the-art methods including unsu- label data.we reduce the number of labeled instances while pervised methods:ITQ [Gong et al.,2013],LSH [Gionis et other settings remain the same.For supervised learning.in al.,1999],IsoH [Kong and Li,2012]and SpH [Heo et al., CIFAR-10,we randomly select 2500(250 images per class) 2012];supervised methods:LFH [Zhang et al.,2014],FastH images and in NUS-WIDE,we randomly select 5000 images. The MAP results with this experiment setting are listed in Ta- https://www.cs.toronto.edu/kriz/cifar.html ble 3.Note that the results of unlisted unsupervised methods http://lms.comp.nus.edu.sg/research/NUS-WIDE.htm are the same as those in Table 2.We observe that our BGDH 3242

Algorithm 2 Learning algorithm for BGDH 1: Input: A bipartite graph G = (X , O, E), images X = {xi} n i=1, pairwise labels S = {sij}, parameters η, λ, batch iterations T1, T2 and sizes N1, N2 2: Output: Binary codes B = {bi} n i=1 3: Initialization: Initialize θx with the pre-trained CNN-F model on ImageNet; Initialize each entry of M, v, and θe by randomly sampling from a Gaussian distribution with mean 0 and variance 0.01. 4: REPEAT 5: for t ← 1 to T1 do 6: Randomly sample N1 labeled images, and let I1 be the set containing indexes of sampled instances 7: Calculate φ(xi ; θx) and ψ(ei ; θe) for all i ∈ I1 by forward propagation 8: Compute ui = MT [φ(xi ; θx); ψ(ei ; θe)] + v and the binary code of xi with bi = sgn(ui) for all i ∈ I1 9: Compute the derivative of Ls w. r. t. {ui : i ∈ I1} 10: Update parameters M, v, θx, θe, and {ei : i ∈ I1} by back propagation 11: end for 12: for t ← 1 to T2 do 13: Randomly generate a batch of triples by invoking Al￾gorithm 1 N2 times, and let I2 be the set containing in￾dexes of sampled instances and contexts 14: Compute the derivative of Lg w. r. t. {ei , wc : (i, c) ∈ I2} 15: Update parameters {ei , wc : (i, c) ∈ I2} 16: end for 17: UNTIL stopping 18: Calculate bi = sgn ￾MT [φ(xi ; θx); ψ(ei ; θe)] + v
for i = 1, . . . , n 4 Experiments In this section, we present experimental results. All the ex￾periments are performed on a NVIDIA K80 GPU server with MatConvNet [Vedaldi and Lenc, 2014]. 4.1 Datasets and Setting We conduct experiments on two widely used benchmark datasets: CIFAR-10 and NUS-WIDE. The CIFAR-10 dataset1 consists of 60, 000 images from 10 classes (6000 images per class). It is a single-label dataset. The NUS-WIDE dataset2 contains 269, 648 images from Flickr. Following the settings in [Lai et al., 2015], we use 19,5834 images belonging to the 21 most frequent classes and each class consists at least 5000 images. Additionally, two images are defined as a ground￾truth neighbor when they share at least one common label. In our experiments, BGDH-T denotes transductive mod￾el while BGDH-I denotes inductive one. We compare our method with several state-of-the-art methods including unsu￾pervised methods: ITQ [Gong et al., 2013], LSH [Gionis et al., 1999], IsoH [Kong and Li, 2012] and SpH [Heo et al., 2012]; supervised methods: LFH [Zhang et al., 2014], FastH 1 https://www.cs.toronto.edu/ kriz/cifar.html 2 http://lms.comp.nus.edu.sg/research/NUS-WIDE.htm [Lin et al., 2014], SDH [Shen et al., 2015] and COSDISH [Kang et al., 2016]; supervised deep hashing methods: DPSH [Li et al., 2016], DHN [Zhu et al., 2016] and DSH [Liu et al., 2016]; semi-supervised deep methods SSDH [Zhang et al., 2016]. For hashing methods using hand-crafted features, we rep￾resent each image in CIFAR-10 by a 512-dimensional GIST vector. In NUS-WIDE, an image is represented by a 1134- dimensional feature vector, including 500-D bag-of-words features, 64-D color histogram, 144-D color correlogram, 73- D edge direction histogram, 128-D wavelet texture, 225-D block-wise color moments. For deep methods, we use the raw image pixels as input. The network architectures of these deep models are different from each other, for fair compar￾ison, we adopt the same deep network architecture CNN-F with the same initialization parameters pre-trained on Im￾ageNet [Deng et al., 2009] for all deep hashing methods. The bipartite graph of BGDH is constructed based on hand￾crafted features with heat kernel, where the hyper-parameter ρ is set as 1 for CIFAR-10 and 10 for NUS-WIDE. The hyper￾parameter η in BGDH is set as 10 for CIFAR-10 and 100 for NUS-WIDE similar to DPSH [Li et al., 2016]. We simply set T1 = 10, T2 = 5, λ = 0.1 in all the experiments. 4.2 Experimental Results To evaluate the retrieval accuracy, we use the Mean Average Precision (MAP) as the evaluation metric. Following [Lai et al., 2015; Xia et al., 2014], 1000 images (100 images per class) are randomly selected as the query set in CIFAR-10. We use all the rest of images as training set for the unsu￾pervised methods and for building a neighbour graph G in BGDH. We randomly select 5000 (500 images per class) im￾ages as the training set for supervised methods. In NUS-WIDE, we randomly sample 2100 query images from 21 most frequent labels following the strategy in [Lai et al., 2015; Xia et al., 2014]. We use all the rest of images as training set for the unsupervised methods. To reduce the time complexity, we only use 5000 landmarks sampled randomly to build a bipartite graph for BGDH. For supervised methods, we randomly select 10500 images from dataset. The MAP values are calculated within the top 5000 returned neighbors. We present the MAP results in Table 2 where BGDH, SS￾DH, DSH, DHN, DPSH are deep methods. Except for SSDH which uses triplet labels, all deep methods are trained with pairwise labels. To be fair, parameters of these methods are set according to the suggestions of authors. The results show that in most cases our proposed BGDH method substantially outperforms other baselines including unsupervised methods, supervised methods and semi-supervised methods. In real applications, the number of labeled instances is usu￾ally far less than that of unlabeled instances. To further verify the effectiveness of leveraging both pairwise labels and un￾label data, we reduce the number of labeled instances while other settings remain the same. For supervised learning, in CIFAR-10, we randomly select 2500 (250 images per class) images and in NUS-WIDE, we randomly select 5000 images. The MAP results with this experiment setting are listed in Ta￾ble 3. Note that the results of unlisted unsupervised methods are the same as those in Table 2. We observe that our BGDH Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) 3242