1

A Machine Learning Framework for Image Collection Summarization Jorge Camargo, Fabio González {jecamargom, fagonzalezo}@unal.edu.co National University of Colombia

Abstract—In this paper, we propose a machine learning framework for summarizing and visualizing large collections of medical images. Because it is not possible to visualize all images of the collection it is necessary to visualize an overview that represents the complete collection. A kernel that involves domain knowledge is used in the visualization and summarization process. For building the summary we use a clustering method that selects an image subset that is visualized to user. We use dimensionality reduction techniques for visualizing the images in a 2D space. Our experiments show that it is possible to resume a large collection of medical images with the proposed framework. Index Terms—Image collection summarization, visualization, clustering, multidimensional scaling, Isomap

I. I NTRODUCTION The huge amount of visual and multimedia data is growing exponentially thanks to the development of Internet and to the easy of producing and publishing multimedia data. This generates two main problems: how to find efficiently and effectively the information needed, and how to extract knowledge from the data. These problems have been mainly addressed from the Information Retrieval (IR) perspective, approach that has been very useful dealing with textual data [1]. However, there are still a huge amount of work to do on other kind of non-textual data, such as images. Information visualization techniques [2] are an interesting alternative to address the problem in the case of large collection of images. Information visualization techniques offer ways to reveal hidden information (complex relationships) in a visual representation and allow users to seek information in a more efficient way [3]. Thanks to the human visual capacity for learning and identifying patterns, visualization is a good alternative to deal with this kind of problems. However, the visualization itself is a hard problem; one of the main challenges is how to find low-dimensional, simple and faithfully representations of the complete dataset and the relationships among data objects [4]. On the other hand, in the medical field, many digital images (x-ray, ultrasound, tomography, etc.) are produced for diagnosis and therapy. The Radiology Department of the University Hospital of Geneva generated more than 12,000 images per day in 2002, which requires Terabytes of storage per year [5]. Visualization tools are necessary in health centers to assist diagnosis tasks effectively and efficiently. For instance, a medical doctor may have a diagnostic image and wants to find similar images associated to other cases that helps him to assess the current case. Previously, the doctor would need to sequentially traverse the image database looking

for similar images, a process that could be unfeasible for moderately large databases. Nowadays, image visualization techniques provides a good alternative by generating compact representations of the collection, which are easier to navigate allowing the user to find quickly the information needed. The use of projection methods based only on low-level features is a common strategy in visualization of image collections, but it exists a huge semantic gap in the resulting visualization since domain knowledge is not taken into account. The present paper proposes a framework for summarizing large collections of medical images: a kernel that involves domain knowledge, an overview of the dataset built using a clustering method, and a visualization of the overview using a projection method. The reminder of this paper is organized as follows: In Section 2, related work is presented and briefly discussed; in Section 3, the kernel-based approach for improving the visualization is described; Section 4, shows the experimental evaluation of the strategy. Finally, Section 5 presents the conclusions and future work.

II. R ELATED W ORK In [6], authors propose an exploration system with visualization and summarization capabilities. They extract image features, summarize the collection and project the clusters in a low dimensional space. Li et al [7] propose an automatic method for summarizing collections of personal photos based on time partitions and content analysis. In [8], authors propose a system called PERSIVAL, which uses multimedia information from different sources as patient database, medical entities dictionary, evidence-based medicine rules, query knowledge base and library resources. They build summaries from echocardiogram videos and use some machine learning techniques. In this paper authors build a cluster hierarchy of images taken with digital cameras. The hierarchy is built based on keywords and pixel values, and representative images are selected for each cluster, that task is performed in a preprocessing step. Gomi et al [9] build a cluster hierarchy structure based on keywords and pixel values, and representative images are selected for each cluster. Authors propose a hierarchical data visualization technique to visualize the tree structure of images using nested rectangular region.

2

III. K ERNEL - BASED M EDICAL I MAGE C OLLECTION S UMMARIZATION We aim to generate an overview of the image collection that faithfully represents the complete collection. The main phases of our proposed framework are: first, to extract the image features; second, to build a kernel that reflects the similarity notion of expert pathologists; third, clustering the collection for building a summary; and fourth, to apply a projection method that reduces the dimensionality of the image summary for visualization purposes. The details of these four phases are presented in the following subsections. Figure 1 shows an overview of the framework proposed.

or statistical strategy may be used to discover relationships and patterns in that new space. Intuitively, kernel functions provide a similarity relationship between objects being processed, so they are widely used in similarity-based learning too. In this work, we use kernel functions with a twofold purpose: first, to model a more appropriate similarity measure between images using low-level visual features, and second, to learn a combination of features adapted to those particularities of the application domain. 1) Kernel functions : A histogram is a discrete and nonparametric representation of a probability distribution function. Although they may be seen as feature vectors, they have particular properties that may be exploited by a similarity function. There are different kernel functions specially tailored to histograms. In this work we use the histogram intersection kernel. Consider h as a histogram with n bins, associated to one of five different visual features. The Histogram Intersection Kernel is defined in Equation 1. k∩ (hi , hj ) =

n X

min (hi (k), hj (k))

(1)

k=1

Fig. 1: Framework for summarizing a large collection of medical images

A. Features extraction Histopathology images are a particular kind of medical images acquired under a microscope after special staining processes. The differential diagnosis of these images is based on the visual inspection of slides in which pathologists recognize distinctive patterns associated to diseases. We aim to model a similarity measure that approximates the similarity notion used by experts. The first step to calculate such a similarity function is the extraction of visual features. We used a set of low-level features considering four important visual characteristics: luminance, colors, textures and edges. The set includes the following global features: Gray Histogram (GH), RGB color histogram (RGB), Tamura Texture Histogram (TT), Sobel Histogram (SH) and Invariant Feature Histograms (IFH). All these five visual features are modeled as probability distribution functions. B. Kernel alignment (involving domain knowledge) Kernel functions have been successfully used in a wide range of problems in pattern analysis since they provide a general framework to decouple data representation and learning algorithms. A kernel function implicitly defines a new representation space for the input data in which any geometry

Intuitively, this kernel function is capturing the notion of common area between both histograms. This kernel is applied to two histograms of the same feature, i.e. the evaluation of similarity is made feature by feature in an independent fashion. Using k∩ and the five visual features we obtain five different kernel functions that will be used for learning and visualization. 2) Kernel function adaptation : A kernel function using just one low-level feature provides a similarity notion based on visual perception. For instance, the RGB histogram feature is able to indicate whether two images have similar color distributions. However, we aim to design a kernel function that provides a better notion of image similarity according to experts criteria. Histopathology patterns are a complex mix of different features, hence, we construct a new kernel function using a linear combination of kernel functions associated to individual features. The most simple combination is obtained by assigning equal weights to all base kernel functions, so the new kernel induces a representation space with all visual features. However, depending on the particular histopathology pattern, some features may have more or less importance. The present work uses the kernel alignment framework, initially proposed by Cristianini [10] in the context of supervised learning, to combine different visual features in an optimal way with respect to a domain knowledge target. 3) Kernel alignment: The empirical kernel alignment, is a similarity measure between two kernel functions, calculated over a data sample. If K1 and K2 are the kernel matrices associated to kernel functions k1 and k2 and a data sample S, the kernel alignment measure is defined as: AS (K1 , K2 ) = p

hK1 , K2 iF p hK1 , K1 iF hK2 , K2 iF

(2)

Where h·,P·i P is the Frobenius inner product defined as hA, BiF = i j Aij Bij .

3

We define K1 as the the linear combination of base kernels, that is the combination of all visual features. It is given by: X kα (x, y) = αf k∩ (hf (x), hf (y)) (3)

algorithm to images belong to each cluster represented by the medoid. It process it is repeated until the collection be totally divided. Figure 2 shows a representation of this hierarchy.

f

where x and y are images; hf (x) is the f -th feature histogram of image x; and α is the weighting vector. The definition of a target kernel function K2 , i.e. an ideal kernel with explicit domain knowledge, is done using labels associated to each image that are extracted from expert annotations. It is given by the explicit classification of images for a particular concept using yn as the labels vector associated to the n-th class, in which yn (x) = 1 if the image x is an example of the n-th concept and yn (x) = −1 otherwise. So, K2 = yy 0 is the kernel matrix associated to the target for a particular data sample. This configuration leads to an optimization problem, in which the objective is to find a weighting vector α that maximizes the alignment measure. It is modeled as the following quadratic programming problem with linear restrictions:

Fig. 2: Summarization of the collection using a hierarchy of clusters X X X max : αf yn0 Kf yn − αf1 αf2 hKf1 , Kf2 i − λ αf2 When the hierarchy is build, we apply a projection algorithm f f1 ,f2 f for reducing the original dimensionality to two dimensions that subject to : αf ≥ 0, (4) will be used to visualize each summary (in each level) in a In the present work, kernel-alignment is used to optimally 2D space. combine the individual feature kernels in one kernel that reflects semantic relatedness. This is accomplish by defining D. Visualization a target kernel function (ideal kernel) based on image annotaThere are different methods for reducing the dimensionality tions assigned by an expert. of a set of data points. Generally these methods select the dimensions that best preserve the original information. Methods C. Summarization like Multidimensional Scaling (MDS) [11], Principal ComDue to the huge amount of images, it is not possibe to ponent Analysis (PCA) [12], and Isometric Feature Mapping display all images to the user. Therefore, it is necessary to (Isomap) [13], have been useful for this projection task. provide a mechanism that summarize the entire collection. Classical MDS is a technique that focuses on finding the This summary represents an overview of the dataset and allows subspace that best preserves the inter-point distances and it to user begin the exploration process. In this framework, we uses linear algebra solution for the problem. The process use k-medoids clustering method from machine learning area involves the calculation of Eigenvalues and Eigenvectors of for building the overview. a scalar product matrix and proximity matrix. The input is a 1) Clustering method: The k-medoids algorithm is a clus- similarity matrix of images in a high-dimensional space and tering algorithm related to the k-means algorithm and the the result is a set of coordinates that represent the images medoidshift algorithm. Both k-means and k-medoids algo- in a low dimensional space [3]. ISOMAP uses graph-based rithms break the dataset up into groups and attempt to mini- distance computation in order to measure the distance along mize squared error, the distance between points labeled to be local structures. The technique builds the neighborhood graph in a cluster and a point designated as the center of that cluster. using k-nearest neighbors, then uses Dijkstra’s algorithm to In contrast to the k-means algorithm k-medoids chooses data find shortest paths between every pair of points in the graph, points as centers (medoids or images in our case). In our then the distance for each pair is assigned the length of this framework, we apply the clustering algorithm in the high shortest path and finally, when the distances are recomputed, dimensional space and then we project the medoids to a low MDS is applied to the new distance matrix [4]. dimensional space. The goal here is to obtain the k most Additionally to ISOMAP, which is a method that preserves representative images of the collection in order to show them the non-linear structure of the relationships, there exist other to user for beginning the exploration. methods like Locally Linear Embedding (LLE) [14], an un2) Hierarchy of clusters: For summarizing the collection supervised learning algorithm that computes low-dimensional we apply successively the clustering method in order to break neighborhood preserving embeddings of high dimensional the collection in a hierarchy of clusters. The first overview data. SNE [15] is a method based on the computation of probis obtained applying k-medoids for selecting the k most abilities of neighborhood assuming a Gaussian distribution, in representatives images. Then, it is applied again the clustering both the high dimensional and the 2D space. The method then

4

tries to match the two probability distributions. Nguyen in [4], proposes a combination of non-linear methods to build new methods. On the other hand, all projection methods described above and the used in this paper need a distance matrix as input. We have adapted kernel functions with different visual features and domain knowledge. Since a kernel function gives the dot product in an embedded space, we can calculate the point distances in that embedded space using the following transformation: d(xi , xj )2 = k(xi , xi ) − 2k(xi , xj ) + k(xj , xj )

(5)

where k(xi , xj ) is the similarity (kernel1 ) between xi and xj . IV. E XPERIMENTAL E VALUATION The main goal of the experimentation phase is to show how our framework allows to build an overview of different image collections. In this experimentation we apply the proposed framework with kernel alignment strategy to Carcinoma data set because we had pathologist annotations. In the experiments with Corel and ImageClef data sets, we used a kernel based only on low-level features. The following subsections describe the experimental setup as well as the experimental results and their discussion. A. Prototype We developed a system prototype using Matlab for the preprocesing steps and JEE web application for visualizing and exploring purposes. Figure 3 shows a screen shot of the web user interface.

Fig. 3: Screen shot of the software prototype

B. Data sets We used 3 data sets in the experimentation, which are described in Table I. 1 The similarity measures used in this work are kernel functions, which corresponds to the dot product in a particular Hilbert space, this makes it possible to define a distance function based on them [10].

TABLE I: Data sets used in the experimentation Dataset Corel ImageClef Carcinoma

Kind of images general medical images histopathology images

Collection size (images) 2500 200 1502

Corel data set is a collection of photographic stock images and clip art, it is the most widely used standard data set for testing content based image retrieval systems CBIR. From this collection we selected a subset consists of 2500 images. ImageCLEF dataset [16] is a subset collection of 195 images in JPEG format used in the Cross-Language Retrieval in Image Collections (ImageCLEF 2007), which has medical radiography’s taken randomly from medical routine at the RWTH Aachen University Hospital. Carcinoma data set has been used to diagnose a kind of skin cancer known as basal-cell carcinoma. The histopathology collection is composed of 5,995 images from which a subset of 1,502 images was studied and annotated by a pathologist to describe its contents. The annotation process and the complete description of the dataset is detailed in [17]. The pathologist has determined that in this collection there are examples of 18 histopathology concepts. C. Experimental results A summary for each data set was generated using the framework proposed. Figure 4, shows the visualization of the entire Corel data set, highlighting the medoids (pivots) of each cluster (Figure 4a) and the visualization only of the medoids (Figure 4b). This visualization represents a summary of the collection that may be shown to user at the beginning of the exploration. In this experiment we are visualizing only 50 images that represents 2500, which is more suitable due to the limit of the screen display. Note that each pivot is representing a group of similar images, in this case, the similarity is based on color (RGB) feature. Figure 5, shows a visualization of the summary of the ImageClef data set. We used also RGB feature for building this visualization. Pivots in this case are representing histological images, radiographic images, and other kind of images. In Figure 6, it is visualized the Carcinoma collection with the pivots obtained. We used a kernel function calculated using a summation of kernels correspond to visual low level features (color, edges, textures and gray). Each kernel function was obtained using the process described in Section 3.2. Histological images are more uniforms, so it is more difficult to say whether pivots obtained are a good summary of the collection or are not. However, an intuitive analysis says that these prototype images are covering the entire space. The last visualization in Figure 7, corresponds to the Carcinoma data set highlighting the pivot images. The visualization was generated involving domain knowledge in the calculation of the kernel function. The visualization shows that there are images near each other although they are visually dissimilar. It is an interesting result that shows how the domain knowledge involved in the similarity calculation process, allow us to build a more suitable kernel function that reflects similarities among

5

(a) Visualization of the Corel summary with 50 pivots highlighted

(b) Visualization of 50 Corel summary pivots

Fig. 4: Corel data set summary using RGB feature and MDS. (a) Visualization of the entire collection highlighting 25 pivots. (b) Visualization of the 25 pivots that summarize the entire collection.

(a) Visualization of the ImageClef summary with 15 pivots highlighted

(b) Visualization of 50 ImageClef summary pivots

Fig. 5: ImageClef data set summary using RGB feature and Isomap projection method. (a) Visualization of the entire collection highlighting 15 pivots. (b) Visualization of the 15 pivots that summarize the entire collection.

images different to low level visual features in the visualization and summarization process. D. Discussion The results obtained show that for data sets like Corel it is easy to say whether a summary is intuitively good. In fact, summaries for medical image data sets like ImageClef allow us to use the same analysis. Medical image data sets like histopathology collections make it necessary to obtain physicians feedback to establish the quality of the summary. However, these preliminary results allow us to see the power of the clustering methods as mechanism for building image collection summaries. C ONCLUSIONS AND F UTURE W ORK We have presented a machine learning framework for summarizing and visualizing of large collections of images. The

preliminary experiments allow us to see the power of the clustering methods as mechanism for building image collection summaries. The proposed framework is based on machine learning techniques: dimensionality reduction techniques for projecting the high dimensional space of the original images, kernel alignment for involving domain knowledge in the calculation of the similarity measure, and clustering for summarizing the image collection. Kernel alignment allows us to tune up the image similarity measure used by the projection method, using expert domain knowledge represented as image labels. It is important to notice that this labels are only used in a training phase. This phase produced a optimized similarity function that can be applied to new unlabeled images. In future work, we will evaluate the proposed framework with physicians in order to validate how good is the summary obtained. Experiments that involve searching objective images, navigation time and the user experience, will be addressed in

6

(a) Visualization of the Carcinoma summary with 18 pivots highlighted

(b) Visualization of 18 Carcinoma summary pivots

Fig. 6: Carcinoma data set summary using all features kernel and MDS. (a) Visualization of the entire collection highlighting 18 pivots. (b) Visualization of the 18 pivots that summarize the entire collection.

(a) Visualization of the Carcinoma summary with 18 pivots highlighted

(b) Visualization of 18 Carcinoma summary pivots

Fig. 7: Carcinoma data set summary using knowledge-based kernel and MDS. (a) Visualization of the entire collection highlighting 18 pivots. (b) Visualization of the 18 pivots that summarize the entire collection.

future work. Image over-fitting and occlusion will be addressed in future work for optimizing the screen layout.

Medical image collection visualization is an unexplored area that offers interesting and challenging problems. First of all, a huge amount of medical images is produced routinely in health centers that demand effective and efficient techniques for searching, exploration and retrieval. Second, these images have a good amount of semantic, domain-specific content that has to be modeled in order to build effective medical decision support systems. The work presented in this paper is an initial exploration, which suggests that information visualization methods coupled with machine learning techniques may provide meaningful representation of medical image collections.

ACKNOWLEDGMENTS This work was partially funded by Sistema para la Recuperación por Contenido en un Banco de Imágenes Médicas number 1101393199 of Ministerio de Educación Nacional de Colombia through Red Nacional Académica de Tecnología Avanzada RENATA in the Convocatoria 393 de 2006: Apoyo a Proyectos de investigación, desarrollo tecnológico e innovación. R EFERENCES [1] A. Del Bimbo, “A perspective view on visual information retrieval systems,” Content-Based Access of Image and Video Libraries, 1998. Proceedings. IEEE Workshop on, pp. 108–109, Jun 1998. [2] J. D. Stuart K. Card and B. Shneiderman, Readings in Information Visualization: Using Vision to Think. Morgan Kaufmann Publishers, 1999. [3] J. Zhang, Visualization for Information Retrieval. Springer, 2008.

7

[4] G. P. Nguyen and M. Worring, “Interactive access to large image collections using similarity-based visualization,” Journal of Visual Languages & Computing, vol. 19, no. 2, pp. 203–224, April 2008. [Online]. Available: http://dx.doi.org/10.1016/j.jvlc.2006.09.002 [5] H. Muller, N. Michoux, D. Bandon, and A. Geissbuhler, “A review of content-based image retrieval systems in medical applications– clinical benefits and future directions,” International Journal of Medical Informatics, vol. 73, no. 1, pp. 1–23, February 2004. [Online]. Available: http://dx.doi.org/10.1016/j.ijmedinf.2003.11.024 [6] D. Stan and I. K. Sethi, “eid: a system for exploration of image databases,” Inf. Process. Manage., vol. 39, no. 3, pp. 335– 361, May 2003. [Online]. Available: http://dx.doi.org/10.1016/S03064573(02)00131-0 [7] J. Li, J. H. Lim, and Q. Tian, “Automatic summarization for personal digital photos,” in Information, Communications and Signal Processing, 2003 and the Fourth Pacific Rim Conference on Multimedia. Proceedings of the 2003 Joint Conference of the Fourth International Conference on, vol. 3, 2003, pp. 1536–1540 vol.3. [Online]. Available: http://dx.doi.org/10.1109/ICICS.2003.1292724 [8] K. R. Mckeown, S.-F. Chang, J. Cimino, S. Feiner, C. Friedman, L. Gravano, V. Hatzivassiloglou, S. Johnson, D. A. Jordan, J. L. Klavans, A. Kushniruk, V. Patel, and S. Teufel, “Persival, a system for personalized search and summarization over multimedia healthcare information,” in JCDL ’01: Proceedings of the 1st ACM/IEEE-CS joint conference on Digital libraries. New York, NY, USA: ACM Press, 2001, pp. 331–340. [Online]. Available: http://dx.doi.org/10.1145/379437.379722 [9] A. Gomi, R. Miyazaki, T. Itoh, and J. Li, “Cat: A hierarchical image browser using a rectangle packing technique,” in Information Visualisation, 2008. IV ’08. 12th International Conference, 2008, pp. 82–87. [Online]. Available: http://dx.doi.org/10.1109/IV.2008.8 [10] J. Shawe Taylor and N. Cristianini, Kernel Methods for Pattern Analysis. Cambridge University Press, 2004. [11] M. Torgerson, “Multidimensional scaling: I. theory and method,” Psychometrika, vol. 17(4), pp. 401–419, 1958. [12] I. Jolliffe, “Principal component analysis,” Springer-Verlag, 1989. [13] V. Tenenbaum, J. B. de Silva and J. C. Langford, “A global geometric framework for nonlinear dimensionality reduction,” Science, vol. 260, pp. 2319–2323, 2000. [14] L. S. S. Roweis, “Nonlinear dimensionality reduction by locally linear embedding,” Tech. Rep., 2000. [15] G. Hinton and S. Roweis, “Stochastic neighbor embedding,” in Advances in Neural Information Processing Systems 15. MIT Press, 2003. [16] E. K. J. K. T. M. D. W. H. Henning MÃijller, Thomas Deselaers, “Overview of the imageclef 2007 medical retrieval and annotation tasks,” ImageCLEF, 2007. [17] J. Caicedo, “A prototype system to archive and retrieve histopathology images by content,” Master’s thesis, National University of Colombia, 2008.

A Machine Learning Framework for Image Collection ...

exponentially thanks to the development of Internet and to the easy of producing ..... preprocesing steps and JEE web application for visualizing and exploring ...

2MB Sizes 2 Downloads 267 Views

Recommend Documents

A Machine Learning Framework for Image Collection ...
of our proposed framework are: first, to extract the image features; second, to build .... preprocesing steps and JEE web application for visualizing and exploring ...

A Machine Learning Framework for Image Collection ...
Abstract—In this paper, we propose a machine learning frame- work for ..... [14] Jock D.Mackinlay Stuart K. Card and Ben Shneiderman. Readings in. Information ...

Image Collection Summarization A Machine Learning ...
image features are extracted, domain knowledge is involved in the computation of the similarity ... Hierarchical clustering of www image search results using visual, textual and link information. ... images are not available to query the system.

Image Collection Summarization A Machine Learning ...
We propose a machine learning framework for summarizing and visualizing ... Hierarchical clustering of www image search results using visual, textual and link ...

A kernel-Based Framework for Image Collection ...
for summarizing and visualizing large collections of ... to improve the visualization and summarization pro- cess. ... Information visualization techniques [16].

A Machine Learning Framework
ASD has attracted intensive attention in the last decade. [see Tanaka & Sung, 2013, for a review]. Overall ... emotion, are analyzed with computer vision and speech techniques based on machine learning [Bartlett, ..... responses to own name, imitatio

FVD: a High-Performance Virtual Machine Image Format for Cloud
on-write image formats, which unnecessarily mixes the function of storage space allocation with the function of dirty-block tracking. The implementation of FVD is ...

Medical image registration using machine learning ...
Medical image registration using machine learning-based interest ... experimental results shows an improvement in 3D image registration quality of 18.92% ...

A robust incremental learning framework for accurate ...
Human vision system is insensitive to these skin color variations due to the .... it guides the region growing flow to fill up the interstices. 3.1. Generic skin model.

A conceptual framework for the integration of learning ...
Test LT in situ. • Students using the LT. Monitor and adapt the integration. • Continuous “integrative evaluation”. • Adapt the LT and the REST of the course “system”. Evaluation of implementation ..... operates, but whether it does so

A Learning-Based Framework for Velocity Control in ...
Abstract—We present a framework for autonomous driving which can learn from .... so far is the behavior of a learning-based framework in driving situations ..... c. (10a) vk+1|t = vk|t + ak|t · ∆tc,. (10b) where the variable xk|t denotes the pre

A Potential-based Framework for Online Learning with ...
This framework immediately yields natural generalizations of existing algorithms. (e.g. Binomial Weight [CFHW96] or Weighted Majority [LW94, Vov95]) onto online learning with abstentions. 1 Introduction. In many applications of machine learning, misc

A Potential-based Framework for Online Learning with ...
Show xt ∈ /. Predict yt ∈ 1-1, +1, +l. Reveal yt ∈ 1-1, +1l. Reliable predictions on non-abstention examples. Performance Metrics: ▷ Mistakes: ∑t I(yt = -yt) ...

A General Kernelization Framework for Learning ...
Oct 1, 2009 - In summary, after defining a between-class scatter matrix Sb and a within-class matrix Sw ..... Kaufmann, San Francisco, CA, 1998, pp. 515–521 ...

Generation 2.0 and e-Learning: A Framework for ...
There are three generations in the development of AT. ... an organising framework this paper now conducts a ... ways in which web 2.0 technologies can be used.

A Feature Learning and Object Recognition Framework for ... - arXiv
K. Williams is with the Alaska Fisheries Science Center, National Oceanic ... investigated in image processing and computer vision .... associate set of classes with belief functions. ...... of the images are noisy and with substantial degree of.

A Potential-based Framework for Online Learning with ...
A Potential-based Framework for Online Learning with Mistakes and Abstentions. Chicheng Zhang joint work with Kamalika Chaudhuri. UC San Diego. NIPS Workshop on Reliable Machine Learning in the Wild ...

A Feature Learning and Object Recognition Framework for ... - arXiv
systematic methods proposed to determine the criteria of decision making. Since objects can be naturally categorized into higher groupings of classes based on ...