Nuwan Ishanta Senaratna

Projects

Decision Dashboard

The Decision Dashboard (DD) is a research and development project undertaken by the DD research team at Stanford University's Center for Integrated Facility Engineering (CIFE). It is based upon the research of Dr. Calvin Kam (Consulting Assistant Professor at Stanford) who is also head of the DD team.

The Decision Dashboard uses a symbolic model in a graphical window to represent different decisions and their corresponding options. If decision makers mix and match different domain-specific or cross-disciplinary models, preferences, or criteria, the ripple consequences will be dynamically reflected in the symbolic model.

Currently, as a Research Assistant on the DD team, I am the sole Developer/Designer of a Java based application implementing the DD.

Automatic Multiple Document Summarization

A system that would automatically produce summaries of multiple document sets has many applications. The exponential increase in the amount of textual information produced today, particularly over the internet, has, in the last couple of years, accelerated research in this direction. Although there have been several advances in the state of the art, current methods are not without their drawbacks.

In this project, Stanford Grad student Anish Johnson and I, designed and implemented a novel and highly effective method for Multiple Document Summarization aided by the use of FrameNet frames. We use FrameNet, the Stanford POS Tagger and WordNet to enhance a conventional Multiple Document Summarization (MDS) pipeline. Our tests indicated that our method performs better than several current approaches.

Enhancing still image 3-D scene structure learning using human input meta-information

Inferring the 3-D structure underlying a single still 2-D image, given no other information, is an extremely challenging problem. The challenge is compounded by the inherent ambiguities between various 2-D features and there corresponding 3-D features. The work of Saxena et al in solving this problem has proved particularly promising. They use a Markov Random Field (MRF) to infer a set of plane parameters that capture both the 3-D location and 3-D orientation of a segmentation of homogeneous image patches. This models both image depth information and the relationships between different parts of the image. Although results obtained for a range of images have proved both accurate and visually pleasing, the model does not perform too well certain classes of images.

Source: Make3d

This project consisted of demonstrating that these problems can be mitigated by utilizing a small amount of human input meta-information to enhance the model. The meta-data can be input in a very short time, typically 5 to 10 seconds per image. I proposed two such enhancement methods for incorporating human meta-information: Utilizing Horizon Information and Utilizing Scribbles.

My implementations have resulted in significant improvements in the accuracy of the reconstruction and the visual quality of the reconstructed 3-D structure.

Results of this project have been incorporated in Make3d, a Stanford project led by Ashutosh Saxena and Prof. Andrew Ng.

Stanford Little Dog Robot Project

Stanford Little Dog Robot Project confronts the task of making a robot dog walk across hilly terrain towards a goal, a surprisingly difficult problem. A simple A-Star search is doomed to fail from the beginning simply because of the astronomically large number of steps that the dog can take while at each location on the terrain.

I worked Stanford Little Dog with Stanford Grad Students Todd Sullivan and Lawrence McAfee as a course project for Machine Learning. The result of our effort was a capable dog that can efficiently navigate most difficult terrains. During the project we,

• Abstracted the problem to a generic setting
• Created several classifiers using a multitude of features to predict when steps will fail in the simulator
• Created an automatic system for collected data for the classifiers and for algorithm analysis
• Developed a high level search to find an adequate path along the terrain
• Made the dog try to turn towards goals instead of sidestepping or walking backwards
• Created many visualization and analysis tools to help us analyze our algorithms.

The implementation was in in C. We also used Matlab to test various machine learning techniques before hardcoding it in C.

Extending WordNet using Generalized Automated Relationship Induction

WordNet-like semantic lexicons are vital for research and application development in natural language processing and related fields. However, attempts at extending such lexicons using traditional methods such as hand-crafted patterns or human insertion of new words into the taxonomy to cover more general vocabularies, have proved to be very expensive and tedious. Recently, there has been some work into using machine learning-based methods to automatically learn word relationships from large collections of text.

The goal of our project is to use machine learning techniques to build a framework that is capable of identifying a range of word relationships. As will be elaborated on later, the large amount of flexibility that our package has to offer comes from the fact that, given a user-inputted set of examples for the type of relationship that is to be learned (i.e., such as input derived from WordNet), our algorithm will be able to learn lexical patterns from these examples and use these patterns to identify new word pairs that have the given relationship and then use this knowledge to learn further word pairs using a feedback loop.

Secondly, our package is also very powerful because of its ability to integrate both the Stanford Parser and WordNet into a single system. The Stanford Parser is used to parse sentences into typed dependency parse trees, from which we can extract patterns out of the sentences that are used in our feature vectors.

Previously, an algorithm was developed by Snow et al. to automatically learn new word pairs from text that exhibit the hypernym relationship. We build upon by creating a generic word induction algorithm that can operate on any word relationship where the word pairs exist in the same sentence. Additionally, we invoke the induction process described in repeatedly, training on the previous training set and the new word pairs generated from the previous iteration. A good technique for the hypernym relationship (as is also for our algorithm) to capture structured relational knowledge is to use dependency trees.

Dependency trees represent dependencies between individual words in a sentence.

Our project improves on the general layout of this system by 1) increasing accuracy through the use of a better parser (the Stanford Parser), 2) broadening the use for such an algorithm by incorporating a wider range of word relationships, and 3) by being able to train our algorithm on a much larger corpora than was previously available. Our package will make use of the Stanford Parser (for creating typed dependencies from text), WordNet (for supplying word relationship example data), and various classifier packages to compare and determine what works best for word-pair induction, including SVMs, Naive Bayes, and logistic regression.

Automatic Music Composition using a Tree of Interacting Emergent Systems

Automatic music composition (or AMC) is a "hot field" in computer science. In addition to being one of the fastest growing areas in Artificial Intelligence (AI), it has enormous commercial potential. Entertainment companies such as Sony, Disney and Phillips have invested billions into developing AMC applications. Naturally, a wide range of approaches and techniques have been explored in attempts to build AMC systems. Of these, emergent-computing techniques have proved very promising. However, most attempts have tended to use a single, centralized emergent technique for the entire AMC process. This has many drawbacks.

In this project, I proposed a solution to these drawbacks in the form of a tree-like structure made up of several interacting emergent systems. The author calls this "A Tree of Interacting Emergent Systems" or TIES. TIES is a context-independent framework, applicable to many applications, including AMC. The author demonstrates the latter by using it to implement the product, AMCTIES (An application capable of Automatic Music Composition using a Tree of Interacting Emergent Systems).

AMCTIES defines a TIES framework for musical composition. Musical entities are represented as emergent entities. These are generated by emergent systems known as generators. Low-level emergent entities are used to generate high-level emergent entities, and hence form a tree of emergent systems. For example, Motifs and Rhythm are used to generate Phrases. TIES allows for complex interactions between its emergent systems and implements several additional design components in order to facilitate this. The TIES framework is independent of the implementation of the actual generators. The idea is to design a scalable, flexible and robust framework that will maximize practical results.

To demonstrate how TIES can be used for AMC, several generator emergent systems have been implemented. These use Fractals, Cellular Automata and Genetic Algorithms. The particular implementations have been optimized specifically for AMC.

AMCTIES is delivered as a set of software libraries. These have been developed on the Microsoft .NET framework. The outputs produced by AMCTIES are in the form of MIDI files. The libraries can be used to design a host of applications.

A set of music files generated by AMCTIES was subjected to the scrutiny of several experts in the music field. This scrutinization process was designed to validate the generation of each of the component musical entities. Expert responses indicate that AMCTIES can indeed generate music with a high level of creativity and novelty. Tests also show that it is difficult to significantly differentiate the output of AMCTIES from compositions by experienced human composers.

AMCTIES has huge commercial potential and would be of much interest to the music and entertainment industries.

Timetable and Resource Scheduling System