R. Fergus, P. Perona, and A. Zisserman. Object class recognition by unsupervised scale-invariant learning, In Proc. CVPR, Jun 2003.

In this paper, author propose a image recognize system for determinate whether an object is in the query image. The system can be break into three parts: representation, learning, and recognition. Representation means break image into features, and used these features to represent an object. An object is modeled as a constellation parts, where appearance, shape, relative scale and occlusion are all considered by the model. EM algorithm is used as an unsupervised learning method in this paper. And recognition is done by compare the probability ratio of query image having object present to a threshold. The result of this system is significantly better than the baseline. But the scale of the dataset they used is maybe not large enough, one of the dataset only contain couple hundred images. Also they aligned the images make them facing the same way before learning.
This paper provide a complete probability model to model object in a image. It is kind of hard to digest at the first glance. The model consider several parts including scales, but not rotation. Therefore they need to preprocessing the image before learning. The author suggest to improve the system by adding rotation invariance in the future works.

"Non-linear dimensionality reduction techniques for classification and visualization," Vlachos, et al., KDD 2002.

In this paper, they use non-linear dimensionality reduction techniques to address two problem. One is visualization of high dimensional data, another is curse-of dimensionality problem of classification. In problem one, it propose that LLE and Isomap can not preform well when data are sample from different clusters. They improve the solution by keep distance to both nearest and farthest object when mapping. They solve the second problem by introduce two techniques, weightedIso and Iso+Ada which modifying the transformation performed by the Isomap and produce homogeneous neighborhoods in reduced space. The experiment shows that it perform as well as other method but in a much lower dimensional.

"Nonlinear dimensionality reduction by locally linear embedding," Roweis & Saul, Science, 2000.

In this article, author propose a simple unsupervised algorithm that compute low dimensional embeddings of high dimensional inputs called locally linear embedding(LLE). The algorithm can be separate into three parts:

1. Assign neighbors to each data points Xi.
2. Compute the weights Wij, that best linear reconstruct Xi from its neighbors.
3. Compute low dimensional embedding vector Yi, using the wights Wij that minimize the embedding cost function.

This method find a low-dimensional representation of the data in a straight forward way and preserved the neighborhood property. There are not much math used in this article, all we need is some background in linear algebra(solving least square problem), which make this paper not hard to understand. Even though, this is still quite a innovate method and can be widely used in many applications.

"Eigenfaces for recognition," M Turk, A Pentland - Journal of Cognitive Neuroscience, 1991.

This paper propose a method call eigenface to recognize face in images. This method is fast, simple, and accuracy in constrained environment like household or office. The method decompose the face image into a small set of image call "eigenface" which is basically principle components of set of training face images. And then recognize is preform by project the query image into subspace spanned by enigenfaces.
The method is quite simple and fast, but it has a lot of constrain in image such as lightening and view of the face, which cause limit use in application. Despite of these constrain, it still has great learning ability and performance that let this paper become a classic one.

"Latent Dirichlet allocation," D. Blei, A. Ng, and M. Jordan. . Journal of Machine Learning Research, 3:993–1022, January 2003

Latent Dirichlet allocation(LDA) is a generative model that model hidden topics in documents, it is similar to pLSA with an assumption that the topic distribution have a Dirichlet prior. This is basically all I know from the paper and wikipedia. But I don't really know what "Dirichlet prior" is due to my lack of knowledge in statistic. The math in the paper is quite complicate, and I need more time to digest, hopefully I can gain more insight of this paper in the class.
--
class notes:
@generate doc:
find N in poisson, observe doc, use dirichlet estimate parameter theta, use parameter to model multinomial to find topic for each word, find a word given topic z and beta.
@jensen's inequality:
log(xa + (1-x)b) >= xlog(a) + (1-x)log(b)

"Probabilistic latent semantic indexing," T. Hofmann, SIGIR, 1999.

Probabilistic latent semantic indexing(pLSA) is evolve from latent semantic indexing(LSA) with a statistical foundation which use to automatic indexing by analyze the co-occurrence data. The core of pLSA is a statistical model called aspect model. Consider an word, document pair (w,d),
probability p(w,d) can be express as

P(w,d) =∑P(c)P(d ! c)P(w ! c) = P(d)∑P(c ! d)P(w ! c)

where c is latent class. Then it use a generalization of maximum likelihood for mixture models called tempered EM to fit the model. The experiment show the benefit from pLSA which has significant gain on precision over LSA.

[paper reading]"Efficient visual search of videos cast as text retrieval," J. Sivic, and A. Zisserman, IEEE TPAMI, 2009.

This paper introduce a method to search object in video using text retrieval approach. The algorithm contain two parts, first part is pre-processing. It finds affine covariant regions in each keyframes and represent it using SIFT descriptor, and then discard regions that are unstable in temporal domain. It then randomly select frames to form the visual vocabulary using the stable regions clusters. It also apply a stop-word remove method which discard visual word that appear in nearly all frames to improve the result. After remove stop-word, it calculate tf-idf weighted document frequency, and then build inverted file indexing structure. The last three steps are borrow the idea from text retrieval. The second part is run-time retrieval, when a user select a query region(in a image), it first find the visual word in the query region, and then retrieval keyframes using visual word frequency, finally it re-rank the result using spatial consistency of the visual word to further improve the ranking result. This paper doesn't use much complicate method or math, but it has a great result just using idea borrow from other field. It shows that we don't necessary have to come up with complicate idea to be success, sometimes the idea could be simple but useful. The other good things about this paper is that it has an extensive expirement on a lot of different aspect. These are really worth us to learn.