论文读书笔记-on the difficulty of nearest neighbor search[and so on]

Paper 1：on the difficulty of nearest neighbor search

这篇论文主要介绍的是在最近邻搜索中困难度的相关内容，并且提出了一个具体化的估计困难度的方法（relative contrast），下面是本文摘抄

1、  the term difficulty here has two different but related meanings: in the context of NN search problem, difficulty represents meaningfulness. In the context of approximate NN search methods like tree or hashing based indexing methods, difficulty represents
complexity.

2、  Suppose we are given a data set X containing n d-dim points,X={xi,i=1..n}, and a query q where xi, q are samples from an unknown distribution p(x), let D(,) be the distance function for the d-dimensional data.



Lp distance:





Cr captures the notion of difficulty of NN search in X. Smaller the Cr, more difficult the search. If Cr is close to 1, NN search in database X is not very meaningful.
3、





σ' is usually very small for high dimensional data. It is clear that smaller  leads to smaller relative contrast, more difficult nearest neighbour search.
4、



 
5、dataproperties and relative contrast
Data dimensionality:the larger d will lead to smaller relative contrast, making NN search lessmeaningful.
Data sparsity: thesmaller s will lead to larger relative contrast.
Data size: relativecontrast increases monotonically with n.
Data metric norm:relative contrast with respect to p is not so straightforward.
 

Paper 2. Efficient active algorithm for hierarchical clustering

这篇论文提出的是对于层次聚类的一种算法，更确切的说是一种计算框架。作者通过研究前人的成果，结合常见的聚类算法提出自己的改进建议。下面是摘抄
1、A-any flat clustering algorithm,which takes as parameters a dataset and a natural number k, indicating the number of clusters to produce. k –denote the number of clusters at any split, it is known and fixed across the hierarchy. n –the number
of objects in a datasets.s –a parameter, influencing the number of measurements used by the algorithm, smaller s implies fewer measurements,increasing s increase the robustness of our method.K-possibly noisy similarity function, model both cases where similarities
are measured directly and where they are computed via some kernel function based on observed object features.

 




 

Restrictions:







2、  active spectral clustering

 

Paper 3 choosing linguistics over vision to describe images

这篇论文介绍的是如何对一幅图片进行语言描述，显然这种描述不是人工而是机器自动实现的。文中提出了相比前人而言更好的方法，得到的描述不再是孤立的单词，而是一个句子。摘抄如下：
1、作者首先给出了一个例子，指出对每一幅输入的图片是如何处理的。
(1)givenan unseen image
(2)findK images most similar to it from the training images, and using the phrasesextracted from their descriptions
(3)generatea ranked list of triples which is then used to compose description for the newimage



在上面的第二步中，对每幅图片的描述是人为添加的，但是这些人为添加的东西一开始都是一些句子，需要将其分解为((attribute1,object1),verb),(verb,prep,(attribute2,object2)),(object1,prep,object2)这些形式，具体例子如下：



 
2、  model for predicting phrase relevance







3、  phrase integration