Harish Karnick

Kernel-based online machine learning and support vector reduction

We apply kernel-based machine learning methods to online learning situations, and look at the related requirement of reducing the complexity of the learnt classifier. Online methods are particularly useful in situations which involve streaming data, such as medical or financial applications. We show that the concept of span of support vectors can be used to build a classifier that performs reasonably well while satisfying given space and time constraints, thus making it potentially suitable for such online situations. The span-based heuristic is observed to be effective under stringent memory limits (that is when the number of support vectors a machine can hold is very small).

Extracting semantic structure of web documents using content and visual information

This work aims to provide a page segmentation algorithm which uses both visual and content information to extract the semantic structure of a web page. The visual information is utilized using the VIPS algorithm and the content information using a pre-trained Naive Bayes classifier. The output of the algorithm is a semantic structure tree whose leaves represent segments having unique topic. However contents of the leaf segments may possibly be physically distributed in the web page. This structure can be useful in many web applications like information retrieval, information extraction and automatic web page adaptation. This algorithm is expected to outperform other existing page segmentation algorithms since it utilizes both content and visual information.

Personalized Web Search Using Probabilistic Query Expansion

The Web consists of huge amount of data and search engines provide an efficient way to help navigate the Web and get the relevant information. General search engines, however, return query results without considering users intention behind the query. Personalized Web search systems aim to provide relevant results to users by taking user interests into account. In this paper, we proposed a personalized Web search system implemented at proxy which adapts to user interests implicitly by constructing user profile with the help of collaborative filtering. A user profile essentially contains probabilistic correlations between query terms and document terms which is used for providing personalized search results. Experimental results show that our proposed personalized Web search system is both effective and efficient.

Scale independent raga identification using chromagram patterns and swara based features

In Indian classical music a raga describes the constituent structure of notes in a musical piece. In this work, we investigate the problem of scale independent automatic raga identification by achieving state-of-the-art results using GMM based Hidden Markov Models over a collection of features consisting of chromagram patterns, mel-cepstrum coefficients and timbre features. We also perform the above task using 1) discrete HMMs and 2) classification trees over swara based features created from chromagrams using the concept of vadi of a raga. On a dataset of 4 ragas- darbari, khamaj, malhar and sohini; we have achieved an average accuracy of ~ 97%. This is a certain improvement over previous works because they use the knowledge of scale used in the raga performance. We believe that with a more careful selection of features and by fusing results from multiple classifiers we should be able to improve results further.

Genre and Style Based Painting Classification

As the size of digitized painting collections increase, it becomes more difficult to organize and retrieve paintings from these collections. To manage search and other similar operations efficiently, it becomes necessary to organize the painting databases into classes and sub-classes. Manual tagging of these ever-increasing databases would become very costly and time consuming. The above challenging problem has motivated researchers to work in the area of painting analysis, genre and style classification, artist classification and automatic annotation of paintings with these tags. These problems are quite difficult as first the expected human performance for this task for non-expert but reasonably knowledgeable individuals is believed to be well below 100% percent. And second, there is a very big databases of paintings with relatively few painters painting in a single genre and style and many who paint in multiple genres and styles. In this paper, we explore the problem of feature extraction on the paintings and focus on classification of paintings into their genres and styles. We worked with 6 genres and 10 styles. We get an accuracy of 84.56% for genre classification. We achieved an accuracy of 62.37% for classifying the paintings into 10 styles. We include a comparison to existing feature extraction and classification methods as well as an analysis of our own approach across different feature vectors.

Cosine Distance Metric Learning for Speaker Verification Using Large Margin Nearest Neighbor Method

In this paper, a novel cosine similarity metric learning based on large margin nearest neighborhood LMNN is proposed for an i-vector based speaker verification system. Generally, in an i-vector based speaker verification system, the decision is based on the cosine distance between the test i-vector and target i-vector. Metric learning methods are employed to reduce the within class variation and maximize the between class variation. In this proposed method, cosine similarity large margin nearest neighborhood CSLMNN metric is learned from the development data. The test and target i-vectors are linearly transformed using the learned metric. The objective of learning the metric is to ensure that the k-nearest neighbors that belong to the same speaker are clustered together, while impostors are moved away by a large margin. Experiments conducted on the NIST-2008 and YOHO databases show improved performance compared to speaker verification system, where no learned metric is used.

Artificial ontogenesis of controllers for robotic behavior using VLG GA

In this paper we describe a method for the synthesis of robotic controllers using evolutionary techniques. A modified version of the recurrent neural network used for controlling the robots is evolved using genetic algorithm using the variable length genotype approach where the genotype encodes the network. It has been discovered that the separation of modalities in the network like vision and touch, for the first few layers helps in faster evolution as well as in the development of faster controller networks. The structure of the network that emerges from this kind of evolution is similar to brain-like networks. Performance of plastic versus non-plastic individuals has also been explored. Gene-blocking technique has been used for developing several behaviors in the same evolution cycle. The final controller developed at the end of the evolutionary process was tested on a Khepera.

Informed Weighted Random Projection for Dimension Reduction

Dimensionality reduction is a frequent pre-processing step in classification tasks. It helps to improve the accuracy of classification by better representing the dataset and also alleviates the curse of dimensionality by reducing the number of dimensions. Traditional dimensionality reduction techniques such as PCA or Kernel PCA are well known techniques that find a lower dimensional subspace which best represents the higher dimensional dataset. On the other hand, random projection can also be considered as a dimension reduction technique that tries to approximate the same topology of higher dimensional data in a lower dimensional space. Both approaches reduce dimensions but because of their different objectives they have not been successfully integrated. Here we show that in practice and more specifically in a supervised setting like classification, we can link the two methods to make random projection more informed in making the low dimensional representation competitive with the original data set with respect to classification accuracy. In this paper we propose a novel dimensionality reduction technique, namely informed weighted random projection, that combines Kernel PCA and random projection in an efficient way. The kernel PCA algorithm is applied initially to obtain a sub-space of reduced dimensions then the new lower dimensional bases derived by the kernel PCA are weighted in proportion to the measured robustness coefficient of each base. The proposed dimensionality reduction scheme has been applied on several benchmark datasets from the UCI repository and experimental results show that informed weighted random projection attains higher accuracy than the usual unweighted combination for all the datasets used in our experiments.

A resolution rule for well-formed formulae

A resolution proof procedure that operates on well-formed formulae with all quantifiers in place is presented. Extension of the unification algorithm to Q-unification (i.e. with quantifiers in place) is also discussed. The procedure involves a single inference rule called WFF-resolution which is proved to be sound and complete.

User bias removal in review score prediction

Review score prediction of text reviews has recently gained a lot of attention in recommendation systems. A major problem in models for review score prediction is the presence of noise due to user-bias in review scores. We propose two simple statistical methods to remove such noise and improve review score prediction. Compared to other methods that use multiple classifiers, one for each user, our model uses a single global classifier to predict review scores. We empirically evaluate our methods on two major categories (Electronics and Movies and TV) of the SNAP published Amazon e-Commerce Reviews data-set and Amazon Fine Food reviews data-set. We obtain improved review score prediction for three commonly used text feature representations.