MIMA Group Fast Scalable Supervised Hashing Xin Luo 1 , Liqiang Nie - - PowerPoint PPT Presentation

mima group
SMART_READER_LITE
LIVE PREVIEW

MIMA Group Fast Scalable Supervised Hashing Xin Luo 1 , Liqiang Nie - - PowerPoint PPT Presentation

MIMA Group Fast Scalable Supervised Hashing Xin Luo 1 , Liqiang Nie 2 , Xiangnan He 3 Ye Wu 1 , Zhen-Duo Chen 1 , Xin-Shun Xu 1 1 Lab of M achine I ntelligence & M edia A nalysis, School of Software, Shandong University, China 2 School of


slide-1
SLIDE 1

MIMA Group

MIMA @ SDU School of Software, Shandong University

Fast Scalable Supervised Hashing

Xin Luo1, Liqiang Nie2, Xiangnan He3 Ye Wu1, Zhen-Duo Chen1, Xin-Shun Xu1

1Lab of Machine Intelligence & Media Analysis,

School of Software, Shandong University, China

2School of Computer Science and Technology,Shandong University, China 3National University of Singapore, Singapore

slide-2
SLIDE 2

MIMA

MIMA @ SDU School of Software, Shandong University 2

Outline

n Introduction n Proposed Method n Overall Objective Function n Optimization Algorithm n Out-of-Sample Extension n Experiments n FSSH_deep n Conclusion & Future Work

slide-3
SLIDE 3

MIMA

MIMA @ SDU School of Software, Shandong University 3

Introduction

n Given a query point q, NNS

returns the points closest (most similar) to q in the database.

n Underlying many machine

learning, information retrieval, and computer vision problems.

n Nearest Neighbor SearchNNS

query

slide-4
SLIDE 4

MIMA

MIMA @ SDU School of Software, Shandong University 4

Introduction

n Challenges in large-scale data applications: n Expensive storage cost n Slow query speed n data on the Internet increases explosively n curse of dimensionality problem n One popular solution is the hashing based

approximate nearest neighbor (ANN) technique.

slide-5
SLIDE 5

MIMA

MIMA @ SDU School of Software, Shandong University 5

Introduction

Similarity Preserving

Illustration comes from http://cs.nju.edu.cn/lwj/L2H.html

slide-6
SLIDE 6

MIMA

MIMA @ SDU School of Software, Shandong University 6

Introduction

Ø advantages of hashing: ü fast query speed ü low storage cost

Illustration comes from http://cs.nju.edu.cn/lwj/L2H.html

1 1 1 1 1 1 1 1 1 1 XOR Hamming distance

slide-7
SLIDE 7

MIMA

MIMA @ SDU School of Software, Shandong University 7

Introduction

n According whether to use semantic information: n unsupervised hashing n supervised hashing (better retrieval accuracy) n We propose a novel supervised hashing

method, named, Fast Scalable Supervised Hashing (FSSH).

slide-8
SLIDE 8

MIMA

MIMA @ SDU School of Software, Shandong University 8

Introduction

n Two commonly used objective functions:

instance pairwise semantic similarity instance i and instance j are semantically similar i and j are semantically dissimilar r-bit binary hash codes for n instances Frobenius norm labels for n instances instance i is in class k instance i is not in class k a projection from labels to hash codes c the number of classes

slide-9
SLIDE 9

MIMA

MIMA @ SDU School of Software, Shandong University 9

Introduction

n Two commonly used objective functions:

instance pairwise semantic similarity instance i and instance j are semantically similar i and j are semantically dissimilar r-bit binary hash codes for n instances Frobenius norm labels for n instances instance i is in class k instance i is not in class k a projection from hash codes to labels c the number of classes limitations

slide-10
SLIDE 10

MIMA

MIMA @ SDU School of Software, Shandong University 10

Proposed Method

n Motivations: n How to generate hash codes fast? n How to make the model scalable to large-scale data? n How to guarantee precise hash codes?

slide-11
SLIDE 11

MIMA

MIMA @ SDU School of Software, Shandong University 11

Proposed Method

n Motivations: n How to generate hash codes fast? n How to make the model scalable to large-scale data? n How to guarantee precise hash codes? Ø simultaneously update all bits of hash codes Ø avoid the direct use of large matrix S Ø consider both semantic and visual information

slide-12
SLIDE 12

MIMA

MIMA @ SDU School of Software, Shandong University 12

n Overall Objective Function

Proposed Method

slide-13
SLIDE 13

MIMA

MIMA @ SDU School of Software, Shandong University 13

n We present an iterative optimization algorithm, in

which each iteration contains three steps, i.e., W Step, G Step, and B Step.

n More specifically, n W step: fix G and B, update W; n G step: fix W and B, update G; n B step: fix W and G, update B.

Proposed Method - optimization

slide-14
SLIDE 14

MIMA

MIMA @ SDU School of Software, Shandong University 14

Proposed Method - optimization

n W Step

slide-15
SLIDE 15

MIMA

MIMA @ SDU School of Software, Shandong University 15

Proposed Method - optimization

n W Step

setting the derivative regarding W to zero

slide-16
SLIDE 16

MIMA

MIMA @ SDU School of Software, Shandong University 16

Proposed Method - optimization

n W Step

setting the derivative regarding W to zero

slide-17
SLIDE 17

MIMA

MIMA @ SDU School of Software, Shandong University 17

n W Step

Proposed Method - optimization

where , , .

setting the derivative regarding W to zero

m << n, c << n

slide-18
SLIDE 18

MIMA

MIMA @ SDU School of Software, Shandong University 18

Proposed Method - optimization

n G Step

setting the derivative regarding W to zero

where , , .

slide-19
SLIDE 19

MIMA

MIMA @ SDU School of Software, Shandong University 19

Proposed Method - optimization

n B Step Then, we transform the above equation into, where Tr( ) is the trace norm.

slide-20
SLIDE 20

MIMA

MIMA @ SDU School of Software, Shandong University 20

Proposed Method - optimization

n B Step Then, we transform the above equation into, where Tr( ) is the trace norm.

constants

slide-21
SLIDE 21

MIMA

MIMA @ SDU School of Software, Shandong University 21

Proposed Method - optimization

n B Step Thus, B can also be solved with a closed-form solution stated as follows,

slide-22
SLIDE 22

MIMA

MIMA @ SDU School of Software, Shandong University 22

Proposed Method - optimization

slide-23
SLIDE 23

MIMA

MIMA @ SDU School of Software, Shandong University 23

Proposed Method -

Out-of-Sample Extension

FSSH_os simultaneously learns its hash functions and hash codes FSSH_ts uses linear regression as the hash function FSSH_deep adopts deep network as the hash function

Suppose Xquery and Bquery are the original features and corresponding hash codes of the queries. FSSH_os

slide-24
SLIDE 24

MIMA

MIMA @ SDU School of Software, Shandong University 24

Proposed Method -

Out-of-Sample Extension

FSSH_ts where λe is a balance parameter, is the regularization term, and is the RBF kernel features. Then, the optimal P can be computed as,

slide-25
SLIDE 25

MIMA

MIMA @ SDU School of Software, Shandong University 25

Experiments

n Datasets n MNIST n CIFAR-10 n NUS-WIDE

MNIST CIFAR-10

slide-26
SLIDE 26

MIMA

MIMA @ SDU School of Software, Shandong University 26

Experiments

n Compared supervised methods n one-step hashing: KSH, SDH, FSDH. n two-step hashing: TSH, LFH, COSDISH. n Evaluation Metrics (accuracy) n Mean Average Precision (MAP), n Top-N Precision curves, n Precision-Recall curves. n Time cost (efficiency)

slide-27
SLIDE 27

MIMA

MIMA @ SDU School of Software, Shandong University 27

Experiments - MAP results

  • The best MAP values of each case are shown in boldface.
  • One-step hashing: KSH, SDH, FSDH, and FSSH_os.
  • Two-step hashing: TSH, LFH, COSDISH, and FSSH_ts.
slide-28
SLIDE 28

MIMA

MIMA @ SDU School of Software, Shandong University 28

Experiments - curves

slide-29
SLIDE 29

MIMA

MIMA @ SDU School of Software, Shandong University 29

Experiments - time

  • The numbers of training images on CIFAR-10 and

MNIST are 59,000 and 69,000, respectively.

  • Only 2,000 samples are used for training KSH and

TSH due to their large complexity.

slide-30
SLIDE 30

MIMA

MIMA @ SDU School of Software, Shandong University 30

FSSH_deep

n FSSH_deep is one two-step variant of FSSH:

n 1st STEP: We use features which are extracted by an off-

the-shelf deep network.

n 2nd STEP: We adopt CNN-F network as the hash function.

(We train the network by solving a multi-label classi cation problem.) n Compared deep hashing methods include DSRH,

DSCH, DRSCH, DPSH, VDSH, DTSH, and DSDH.

slide-31
SLIDE 31

MIMA

MIMA @ SDU School of Software, Shandong University 31

FSSH_deep - MAP results

  • All baselines are end-to-end methods.
  • FSSH_deep is not end-to-end.
  • For a fair comparison, the results of baselines are directly reported from

previous works.

slide-32
SLIDE 32

MIMA

MIMA @ SDU School of Software, Shandong University 32

Conclusion & Future Work

n We propose a novel supervised hashing method named

Fast Scalable Supervised Hashing.

n FSSH can be trained extremely fast. n FSSH is scalable to large-scale data. n FSSH generates precise hash codes. n Three variants of FSSH are further proposed: n two shallow variants, i.e., FSSH_os and FSSH_ts n one deep variant, i.e., FSSH_deep

slide-33
SLIDE 33

MIMA

MIMA @ SDU School of Software, Shandong University 33

Conclusion & Future Work

n Extensive experiments are conducted on three

benchmark datasets. Experimental results shows the superiority of FSSH.

n In future, we plan to realize our proposed FSSH method

in an end-to-end deep version to boost its performance.

slide-34
SLIDE 34

MIMA Group

MIMA @ SDU School of Software, Shandong University

Any Question?

Codes are available at: https://lcbwlx.wixsite.com/fssh