Learning Spectral Filters for Single- and Multi-Label Classification of Musical Instruments (PhD Defense)
Date/Time: Tuesday, July 28, 2015 from 9:00 a.m. - 10:00 a.m.
Location: EPS 126
Presenter: Patrick Donnelly, Montana State University
Abstract: Musical instrument recognition is an important research task in the area of music
information retrieval. While many studies have explored the recognition of individual
musical instruments in isolation, the field has only recently begun to explore the
more difficult multi-label classification problem of identifying the musical instruments
present in polyphonic mixtures. This dissertation presents a novel method for feature
extraction in multi-label instrument classification and makes important contributions
to the domain of instrument classification and to the general research area of multilabel
In this work, we consider the largest collection of instrument samples to date in the musical instrument classification literature. We examine 13 musical instruments common to four datasets, including the first use of a dataset in this research domain. We consider multiple performers, multiple dynamic levels, and all possible musical pitches within the range of the instruments.
To the area of multi-label classification, we introduce a binary-relevance feature extraction scheme to couple with the common binary-relevance classification paradigm. This approach allows consideration of a unique feature space for each binary classifier, allowing selection of features unique to each class label. We present a data-driven approach to learning areas of spectral prominence for each instrument and use these locations to guide our binary-relevance feature extraction. We use this approach to estimate source separation of our polyphonic mixtures.
We contribute the largest study of single- and multi-label classification in musical instrument literature and demonstrate that our results track with or improve upon the results of comparable approaches in the literature. In our solo instrument classification experiments, we provide the seminal use of Bayesian classifiers in the domain, introduce the grid-augmented topology for na ive Bayes, and demonstrate the utility of conditional dependencies between frequency- and time-based features for the instrument classification problem. For multi-label instrument classification, we explore the question of dataset bias for polyphonic test sets derived from the monophonic training sets in cross-validation study controlled for dataset independence. Additionally, we present the most comprehensive cross-dataset study in the instrument classification literature and demonstrate the generalizability of our approach.
Furthermore, we consider the difficulty of the multi-label problem with regards to label density and cardinality and present experiments with a reduced label set, comparable to many studies in the literature, and demonstrate the efficacy of our system on this easier problem. We provide a comprehensive set of multi-label evaluation measures with the goal of aligning the instrument classification literature with the standard evaluation practices of the general multi-label community.
Bounding Rationality by Computational Complexity
Date/Time: Monday, May 4, 2015 from 4:10 p.m. - 5:00 p.m.
Location: Byker Auditorium, Chemistry and Biochemistry Building
Presenter: Lance Fortnow, Georgia Institute of Technology
Traditional microeconomic theory treats individuals and institutions of completely
understanding the consequences of their decisions given the information they have
available. These assumptions may not be valid as we might have to solve hard computational
problems to optimize our choices. What happens if we restrict the computational power
of economic agents?
There has been some work in economics treating computation as a fixed cost or simply considering the size of a program. This talk will explore a new direction bringing the rich tools of computational complexity into economic models, a tricky prospect where even basic concepts like "input size" are not well defined.
We show how to incorporate computational complexity into a number of economic models including game theory, prediction markets, forecast testing, preference revelation and awareness.
This talk will not assume any background in either economics or computational complexity.
End of Year Celebration and Awards Ceremony
Date/Time: Monday, April 20th from 4:10 p.m. - 5:00 p.m.
Location: EPS 108
We will reflect on some of our accomplishments from the 2014-2015 academic year. Awards will be given to recognize achievement of graduate students (e.g. Outstanding Ph.D. Researcher, Outstanding GTA) and faculty (e.g. Researcher of the Year). Light snacks and refreshments will be served.
Reactive Game Engine Programming for STEM Outreach
Date/Time: Monday, February 23, 2015 from 4:10 p.m. - 5:00 p.m.
Location: EPS 108
Presenter: Alan Cleary, Montana State University
Science, Technology, Engineering, and Mathematics (STEM) are pervasive in our society. For this reason it is important that we incorporate STEM topics in our education system. The crux of the problem is how to make these topics accessible to younger students in an engaging manner. We present our experiences using a novel programming style, reactive programming, to deliver a summer camp for students in grades 8 through 12. This software uses a declarative programming approach to allow students without a background in computing to explore a wide variety of subject material within a 3D virtual environment, including computer science, mathematics, physics, and art. This work is based on PyFRP, a reactive programming library written in Python. We describe our camp experience and provide examples of how this style of programming supports a wide variety of educational activities.
Predicting Metamorphic Relations for Testing Scientific Software: A Machine Learning Approach Using Graph Kernels
Date/Time: Monday, February 2, 2015 from 4:10 p.m. - 5:00 p.m.
Location: EPS 108
Presenter: Upulee Kanewala, Colorado State University
Comprehensive, automated software testing requires an oracle to check whether the output produced by a test case matches the expected behavior of the program. But the challenges in creating suitable oracles limit the ability to perform automated testing in some programs including scientific software. Metamorphic testing is a method for automating the testing process for programs without test oracles. This technique operates by checking whether the program behaves according to a certain set of properties called metamorphic relations. A metamorphic relation is a relationship between multiple input and output pairs of the program. Unfortunately, finding the appropriate metamorphic relations required for use in metamorphic testing remains a labor intensive task, which is generally performed by a domain expert or a programmer. This talk describes MRpred: an automated technique for predicting metamorphic relations for a given program. MRpred applies a machine learning based approach that uses graph kernels to create predictive models. MRpred achieves a high prediction accuracy, and the predicted metamorphic relations are highly effective in identifying faults in scientific programs.
Topological Data Analysis and Road Network Comparison
Date/Time: Friday, January 30, 2015 from 4:10 p.m. - 5:00 p.m.
Location: Roberts 301
Presenter: Brittany Fasy, Tulane University
Vast amount of data are routinely collected, and analyzing them effectively has become a central challenge we face across science and engineering. Topological data analysis (TDA) is a field that has recently emerged in order to tackle this challenge. This talk will focus on the problem of comparing two road networks (for example, to detect where and by how much a road network has changed over the course of a year). Surprisingly, only recently have distance measures between embedded graphs (representing road networks) been studied. We will see how one of the tools from TDA, namely, persistent homology, can be used to define a local distance measure between two graphs. Persistent homology describes the homology (in particular, the number of connected components and loops) of a data set, at different scales. An example to keep in mind is impressionistic paintings: at one scale, all that is seen are brush strokes; at a larger scale, the brush strokes blur together to form the subject of the painting. The (local) persistent homology distance measure is one of the first theoretically justified approaches to road network comparison. This talk should be accessible to both students and faculty.
How to Use CS to Become a Nuclear Physicist: Applying Computational Geometry to Reactor Physics
Date/Time: Friday, January 30, 2015 from 1:30 p.m. - 2:30 p.m.
Location: CS Conference Room
Presenter: David Millman, Lead Cloud Developer at ProductionPro
Simulating a nuclear reactor is challenging. Often, it involves many computers, working together to solve a very complex differential equation. While many methods exist for solving the equation, the most accurate are Monte Carlo (MC) methods. MC methods use Constructive Solid Geometry (CSG) to model a complex domain with high fidelity. Recent efforts to include feedback effects (e.g., depletion, thermal, xenon, etc.) have forced MC methods to calculate volumes and tight bounding boxes of spatial regions quickly and accurately.
In this talk, I describe a framework for approximating (to a user specified tolerance) volumes and bounding boxes of regions given their equivalent CSG definition. The framework relies on domain decomposition to recursively subdivide regions until they can be computed analytically or approximated. While the framework is general enough to handle any valid region, it was optimized for models used in criticality safety and reactor analysis. For bounding boxes, this is the first algorithm that has strong enough accuracy guarantees yet is fast enough for use within a production level nuclear reactor code. For volume calculations, numerical experiments show that the framework is over 500x faster and two orders of magnitude more accurate than the standard stochastic volume calculation algorithms currently in use.
Virtual Reality and Cybersickness
Date/Time: Monday, January 26, 2015 from 4:10 p.m. - 5:00 p.m.
Location: EPS 108
Presenter: Lisa Rebenitsch, Michigan State University
Often, familiarity with virtual reality is due to movies and shows such as Star Trek, The Matrix, Sword Art Online, and Iron Man's holograms. Real world virtual reality exists beyond flight simulators, and Oculus Rift has increased interest in the field. However, virtual reality implementations differ from television versions. Most rely strictly on sight with some including sounds and, more rarely, touch. Two common paradigms in virtual reality are projection screen systems such as CAVEs and visor display systems such as Oculus Rift. Applications for virtual reality include medical training, military training, museums, collaboration, design, and entertainment.
One safety issue inhibiting use of virtual reality is cybersickness, or the feeling of motion sickness-like symptoms in virtual environments. For example, three-dimensional and shaky camera movies have reports of "movie theater sickness." The likelihood and severity of these symptoms increase in virtual environments. The source of the issue is unclear with research in the field posing over forty potential factors. Factors lie in three categories: individual, hardware, and software. Prior attempts predicting cybersickness are the Cybersickness Dose Value (CSDV) and Kolasinski's linear model. The CSDV correlates well with cybersickness, but only includes software factors. Kolasinski's model explains 34% of the variance and excludes individual factors. Cybersickness is highly individual with a resistant population upwards of 50%. New models using individual characteristics and include the effect of resistance are needed. Statistical and modeling methods called zero-inflated models are examined for better comparison of factors and prediction of cybersickness.
Privacy in Social Computing and Mobile Networking
Date/Time: Friday, January 23, 2015 from 4:10 p.m. - 5:00 p.m.
Location: Roberts 301
Presenter: Na Li, Northwest Missouri State University
With the development of web and wireless technologies and mobile devices, more and more people are conducting their daily activities online. These activities generate large amounts of data including the sensitive information that people are not willing to share with others. Therefore, the disclosure of users’ privacy becomes an intensive concern. This talk will focus on preserving users’ privacy in social media and mobile networks. Specifically, two projects will be introduced. One is to design a privacy-aware friend search engine in Online Social Networks (OSN), and the other one is to preserve users’ relationship privacy in OSN operators sharing data with the third parties. Additionally, this talk will briefly discuss the problems of privacy disclosure in mobile networks.
Classification of Musical Instruments
Date/Time: Wednesday, January 21, 2015 from 4:10 p.m. - 5:00 p.m.
Location: EPS 108
Presenter: Patrick Donnelly, Montana State University
Musical instrument classification is an important task in the area of Music Information Retrieval. While there have been many approaches to recognize individual instruments, the majority of these are not extensible to the more complex case of identifying the musical instruments present in polyphonic mixtures. We present a data-driven clustering technique for learning regions of spectral prominence in an instrument's timbre, exploiting these regions as spectral filters in the feature extraction stage of a binary relevance classification task. We demonstrate the approach over several large datasets consisting of multiple articulations, dynamics, and performers, validating the approach across datasets and with several classifiers. Lastly, we discuss ongoing work in the extension of these spectral filters for source separation estimation in identification of instruments present in polyphonic mixtures.
Seminars from 2014.