1. Jinhui Tang †, Shuicheng Yan †, Richang Hong †, Guo-Jun Qi ‡, Tat-Seng Chua † † National University of Singapore ‡ University of Illinois at Urbana-Champaign

2.  Motivation  Sparse-Graph based Semi-supervised Learning  Handling of Noisy Tags  Inferring Concepts in Semantic Concept Space  Experiments  Summarization and Future Work

4. No manual annotation are required.

5.  With models:  SVM  GMM  …  Infer labels directly:  k-NN  Graph-based semi-supervised methods

6.  A common disadvantage:  Have certain parameters that require manual tuning  Performance is sensitive to parameter tuning  The graphs are constructed based on visual distance  Many links between samples with unrelated-concepts  The label information will be propagated incorrectly.  Locally linear reconstruction:  Still needs to select neighbors based on visual distance

7.  Sparse Graph based Learning  Noisy Tag Handling  Inferring Concepts in the Concept Space

8.  Human vision system seeks a sparse representation for the incoming image using a few visual words in a feature vocabulary. ( Neural Science )  Advantages:  Reduce the concept-unrelated links to avoid the propagation of incorrect information;  Practical for large-scale applications, since the sparse representation can reduce the storage requirement and is feasible for large-scale numerical computation.

9. Normal Graph Construction. Sparse Graph Construction.

10.  The ℓ 1 -norm based linear reconstruction error minimization can naturally lead to a sparse representation for the images *. * J. Wright, A. Yang, A. Ganesh, S. Sastry, and Y. Ma. Robust face recognition via sparse representation. IEEE Transaction on Pattern Analysis and Machine Intelligence, 31(2):210– 227, Feb. 2009  The sparse reconstruction can be obtained by solving the following convex optimization problem: min w ||w|| 1, s.t. x=Dw w ∈ R n : the vector of the reconstruction coefficients; x ∈ R d : feature vector of the image to be reconstructed; D ∈ R d*n (d < n) : a matrix formed by the feature vectors of the other images in the dataset.

11.  Handle the noise on certain elements of x:  Reformulate x = Dw+ξ, where ξ ∈ R d is the noise term.  Then :  Set the edge weight of the sparse graph:

12.  Result:

13.  The problem with :  M uu is typically very large for image annotation  It is often computationally prohibitive to calculate its inverse directly  Iterative solution with non-negative constraints: may not be reasonable since some samples may have negative contributions to the other samples  Solution:  Reformulate: The generalized minimum residual method (usually abbreviated as GMRES) can be used to iteratively solve this large-scale sparse system of linear equations effectively and efficiently.

14. √: correct; ?: ambiguous; m: missing

15.  We cannot assume that the training tags are fixed during the inference process.  The noisy training tags should be refined during the label inference.  Solution: adding two regularization terms into the inferring framework to handle the noise:

16.  Solution:  Set the original label vector as the initial estimation of ideal label vector, that is, set, and then solve and we can obtain a refined f l.  Fix f l and solve  Use the obtained to replace the y in the previous graph-based method, and we can solve the sparse system of linear equations to infer the labels of the unlabeled samples.

17.  It is well-known that inferring concepts based on low-level visual features cannot work very well due to the semantic gap.  To bridge this semantic gap  Construct a concept space and then infer the semantic concepts in this space.  The semantic relations among different concepts are inherently embedded in this space to help the concept inference.

18.  Low-semantic-gap: Concepts in the constructed space should have small semantic gaps;  Informative: These concepts can cover the semantic space spanned by all useful concepts (tags), that is, the concept space should be informative;  Compact: The set including all the concepts forming the space should be compact (i.e., the dimension of the concept space is small).

19.  Basic terms:  Ω : the set of all concepts;  Θ : the constructed concept set.  Three measures:  Semantic Modelability: SM(Θ)  Coverage of Semantic Concept Space: CE(Θ, Ω)  Compactness: CP(Θ)=1/#(Θ)  Objective:

20.  Simplification: fix the size of the concept space.  Then we can transform this maximization to a standard quadratic programming problem.  See the paper for more details.

21.  Image mapping: x i D(i)  Query concept mapping: c x Q(c x )  Ranking the given images:

23.  Dataset  NUS-WIDE Lite Version (55,615 images)  Low-level Features  Color Histogram (CH) and Edge Direction Histogram (EDH), combine directly.  Evaluation  81 concepts  AP and MAP

24. Ex1: Comparisons among Different Learning Methods

26.  Ex2: Concept Inference with and without Concept Space

27. Ex3: Inference with Tags vs. Inference with Ground-truth We can achieve an MAP of 0.1598 by inference from tags in the concept space, which is comparable to the MAP obtained by inference from ground-truth of training labels.

28.  Exploited the problem of inferring semantic concepts from community-contributed images and their associated noisy tags.  Three points:  Sparse graph based label propagation  Noisy tag handling  Inference in a low-semantic-gap concept space

29.  Training set construction from the web resource

30. Thanks! Questions?