MIT EECS Event

MASTERWORKS99

A PRESENTATION OF MASTER'S THESIS RESEARCH
IN
ELECTRICAL ENGINEERING AND COMPUTER SCIENCE

Monday, April 26, 1999
from
3:00 to 5:00 pm
in
Building 34
(3rd floor)

Masterworks is an annual presentation of thesis research by Master's students in the Department of Electrical Engineering and Computer Science. It is open to the public; students, undergraduate and graduate, are particularly welcome.

Students who are nearing the end of their thesis research and who wish to make a presentation submit an abstract to Masterworks and those who are selected present a 15-minute talk on their research. Prizes are awarded for the best presentations.

The talks are scheduled in 8 1-hour sessions of 3 or 4 talks each; 4 sessions occur simultaneously.

Abstracts

SESSION A

Optimal Path Planning and High Level Control of an Autonomous Gliding Underwater Vehicle
Anna M. Galea

Glider vehicles present a special problem to path planning in shallow water and high currents due to their slow speed and the increase in required propulsion energy in shallow water operations. The energy-limited and time-efficient nature of optimal paths for these vehicles was studied in a simulated environment without vehicle dynamics considerations. Dynamic programming, analytic methods, and numerical programming methods were applied to the problem of obtaining an optimal path with time and energy required to travel between two waypoints as the minimization criteria. Numerical programming generated the most general results and the greatest insight into this problem. In order to run missions with the strategies thus learned, a layered control architecture was implemented for these vehicles. This architecture allows for mission flexibility and modularity of vehicle behaviors. The specific control strategy can be transported to any glider vehicle, regardless of the hardware used to affect dynamic changes.

Queue Occupancy in Single Server, Deterministic Service Time Tree Networks
Michael J. Neely

Tree networks of single server, deterministic service time queues are often used as models for packet flow in data communication systems with Asynchronous Transfer Mode (ATM) traffic. In this talk, we present a simple method of analyzing packet occupancy in these systems without making any assumptions about the nature of the underlying input processes. We demonstrate how analysis of multi-stage tree systems can be reduced to the analysis of a much simpler 2-stage equivalent model. We also develop an expression for first moments of queue occupancy in terms of first moments of a simple 1-stage equivalent model. From this, we observe an interesting phenomenon for general types of "distributable inputs": Expected occupancy at any queue within a multi-stage tree network will be a concave function of the loadings produced by the multiple exogenous inputs.

Approximation Algorithms for Model-Based Object Recognition
Louay M. J. Bazzi

We consider the problem of model-based object recognition in the presence of noise, spurious features, and occlusion. We address the case where the model is allowed to be transformed by elements in a given space of allowable transformations. Since known algorithms for the problem have unacceptable running times, we turn our attention to approximation algorithms. We present an algorithm that can achieve any desired approximation factor in a relatively low order worst case asymptotic running time. The time constant of the algorithm is polynomial in the inverse of the desired approximation factor. The solution we provide is general enough to handle different cases of allowable transformations, such as planar affine transformations, and scaled rigid motions in arbitrary dimensions.

SESSION B

A Fundamental Study of Feature Evolution During High Density Plasma Etching
Jennifer M. Lane

A comprehensive study of feature evolution during high density plasma etching has been conducted. Polycrystalline silicon nested trench, isolated trench, nested line, isolated line, and window profiles were examined. Source powers of 250, 500, and 750 W, bias powers of 150, 250, and 350 W, pressures of 2, 10, 25, and 50 mTorr, and Cl2, HBr, and HCl feedgases were utilized. A strong correlation was found between sidewall contour and microtrench shape. SiO2-masked and maskless crystalline silicon profiles were examined to elucidate the role of mask charging on microtrench formation. No charging effects were observed in the resultant profiles. Sensitivity of HBr etching to oxygen addition in the plasma as well as faceting of the silicon lines was also examined.

Electrical Degradation of InAlAs/InGaAs Metamorphic High-Electron Mobility Transistors
Samuel D. Mertens

InAlAs/InGaAs High Electron Mobility Transistors (HEMT) hold promise for power-millimeter wave applications. A major reliability concern in some of these devices is the degradation of the drain resistance that is observed when the device is electrically stressed for a long time at bias conditions necessary for power applications. The goal of this thesis was to find the physical origin of this reliability problem and to suggest solutions to it. State-of-the-art InAlAs/InGaAs metamorphic HEMTS, provided by our sponsor, Hewlett Packard, were stressed under different bias schemes. It was found that most figures of merit of the device degrade under severe bias stress. Experiments on mHEMTs have shown that the degradation of RD is strongly correlated to the amount of impact-ionization during stress. We have studied the degradation of simpler Transmission Line Method (TLM) structures. These are devices that have exactly the same material structure as the transistors, but no gate has been fabricated. We have found that these TLMs show different degradation modes. At least one of them appears to be related to impact-ionization. We have identified several possible mechanisms for this degradation.

Wafer Bonding for Monolithic Integration of Si CMOS VLSI Electronics with III-V Optoelectronic Devices
Joanna M. London

GaAs-on-silicon epitaxy techniques as well as wafer bonding GaAs to Si, have been developed to overcome lattice mismatch in order to integrate optoelectronic and Si devices. However, the thermal expansion differences between these materials continues to be a limitation in using either of these approaches. After recognizing that Si devices, such as MOSFETs, are intrinsically thin and relatively strain tolerant, while optoelectronic devices, such as LEDs and lasers, are thick and very strain sensitive, this research was based on developing a better approach which involved bonding thin Si layers to thick GaAs substrates with various dielectric layers as the interface, to produce silicon-on-gallium arsenide (SonG) wafers. Such wafers are suitable for the fabrication of Si SOI CMOS electronics and the subsequent monolithic integration of high performance optoelectronic devices. Future goals for this work include bonding fully processed SOI CMOS wafers to the GaAs, rather than silicon wafers containing no electronics. With the successful development of SonG techniques for monolithic integration, it will be possible to use full-wafer and batch processing techniques for the production of sophisticated economically viable optoelectronic integrated circuits.

There has been considerable recent interest in the use of organic materials as thin film semiconductors. Several advantages including light weight, ease of application, flexibility, and contamination tolerance make these materials potentially attractive alternatives to other technologies such as amorphous silicon. The thesis investigates a high performing organic semiconductor--pentacene--and studies its performance as a function of electrode placement, thickness, electric field used, and temperature. The trap density is also briefly studied. The thesis concludes with the results from an experiment which uses some of the knowledge gained to increase the performance of the pentacene devices.

SESSION C

In recent years, the recognition of handwritten mathematical expressions has received a small, but consistent amount of attention in pattern recognition research. The diversity of approaches to the problem, however, and the lack of a commercially viable system, indicate that there is still much research to be done in this area. In this talk I will describe an on-line approach for converting a handwritten mathematical expression into an equivalent expression in a typesetting command language such as TeX or MathML. The three primary components of this system are a method for classifying isolated glyphs, an algorithm for partitioning a set of strokes into glyphs, and an algorithm for converting a two-dimensional arrangements of glyphs into a typeset expression. For glyph classification, a Gaussian classifier is used to rank order the interpretations of a set of strokes as a single glyph. To partition a set of strokes, the weights generated by the glyph classifier are used to perform a constrained search of possible partitions for the one with the minimum summed cost. Finally, the glyphs are parsed using simple geometric relationships between their bounding boxes as cues for their intended arrangement.

A Cluster-Based Statistical Model for Object Detection in Visual Scenes
Thomas D. Rikert

This thesis presents an approach to object detection which is based on recent work in statistical models for texture synthesis and recognition (Heeger and Bergen 95, DeBonet and Viola 97, Zhu et.al. 98, Simoncelli and Portilla 98). Our method follows the texture recognition work of De Bonet and Viola. We use feature vectors which capture the joint occurrence of local features at multiple resolutions. The distribution of feature vectors for a set of training images of an object class is estimated by clustering the data and then forming a mixture of Gaussian model. The mixture model is further refined by determining which clusters are the most discriminative for the class and retaining only those clusters. After the model is learned, test images are classified by computing the likelihood of their feature vectors with respect to the model. We present promising results in applying our technique to face detection and car detection.

Discrete models of physics are an attractive alternative to continuous models such as partial differential equations. In discrete models, such as cellular automata, space is treated as having finitely many locations per unit volume. Physical processes are modeled by rules that depend on a small number of nearby locations. Such models depend critically on a regular (crystalline) lattice, as well as the global synchronization of all sites. Is it possible to do without these and still model physics? Or are they somehow fundamental? We will answer this question by presenting a class of models that are "extremely local": the update rule does not depend on synchronization with other sites, or on detailed knowledge of the lattice geometry. All interactions involve only a single pair of sites. The models have the further advantage that they exactly conserve quantities such as momentum and energy which are conserved in physics. We will show that the models agree with continuous models. We will draw parallels between kinds of physical models and kinds of computing architectures, and show that the class of models presented corresponds to a new parallel computing architecture known as an amorphous computer.

SESSION D

Automatic Target Recognition Based on Collection of Evidence
Matthew D. Blum

The problem of automatically recognizing an object in an image scene is very difficult. This thesis develops an image-based object recognition algorithm in which information from different features is combined using Dempster-Shafer reasoning. Specific attention is paid to cases in which only partial information is available because of occlusion or sensor limitations. The structure of the recognition system developed herein is as follows. First, some image processing techniques are used to filter out noise, detect edges, and find features in the raw image. Further preprocessing is performed to isolate objects of interest. Finally, Dempster-Shafer reasoning is used to combine from the edge features into a working model of the objects seen in the raw image. The preceding object recognition system was tested on simulated data, dealing with sets of geometrical shapes. Two experiments were performed, one with unoccluded objects, and one with up to four occluded objects in each raw image. Its performance was compared to the Bayesian approach and human classification, where it was found that although it did not outperform human reasoning, Dempster-Shafer reasoning performed considerably better than the Bayesian approach.

The MASC project is adopting a modular, network approach to the design of the next-generation information appliance. Our design separates the computation resources from the appliance using computer cards and embedded backplanes. This allows the appliance, the computer, and the software to be upgraded independently. We have built a prototype computer card that uses an Intel StrongArm-1100 as CPU, has 32 megabytes of DRAM and 4 megabytes of ROM. Software packages that have been ported to the prototype card include the Angel debugger and bootloader, Linux kernel 2.2, and the Linux PCMCIA module. By rewriting the kernel I/O communication functions, existing Cardbus devices and device drivers can be used on our prototype architecture. This hardware and software infrastructure allows a developer to rapidly deploy information appliances simply by interconnecting computer cards and PCMCIA peripherals, and writing intelligent software applications.

Demonstration system for an ultra-low-power video coder and decoder
Rex K. Min

A digital system showcasing two custom video coding chips was designed and implemented. These chips, designed by Thucydides Xanthopoulos at MTL, perform forward and inverse discrete cosine transforms (DCT) with only milliwatts of power. This demonstration captures and buffers full-motion video data, routes the data through the DCT and inverse DCT devices in real-time, and displays the resulting video stream. The system is implemented in a printed circuit board with off-the-shelf components, including an NTSC video decoder, RAM and ROM, programmable logic devices, and an LCD display. Control logic generated from VHDL handles the flow of real-time video data through the system, coefficient quantization, synchronization signals for the LCD, and an I2C serial bus interface. This work not only demonstrates the functionality and low power consumption of the DCT chipset with arbitrary real-time video data, but it also demonstrates hardware-based applications of the most popular design concepts from undergraduate computer science in a successful, real-world application.

SESSION E

The tectorial membrane (TM) is a fragile gelatinous structure that overlies the mechanically sensitive hair bundles of sensory cells in the inner ear. Based on its location, the TM presumably plays an important role in hearing. However, little is known about the physical properties of the TM. Studies of its composition and structure suggest that the TM is a polyelectrolytic gel whose mechanical and electrical properties are strongly linked and depend on the ion content of the solutions that surround it. To test this notion, we have made measurements of the Donnan equilibrium potential between the TM and bathing solutions. Interpreting these results in the context of an isotropic polyelectrolytic gel model yields a fixed charge concentration of nearly 200 mmol/L, which is 6 times larger than any previously published estimate. Despite the large value of this estimate, the stability and repeatability of the measurements gives us confidence in our method. This result has important implications for hearing. For instance it is known that the ion content of the solutions that surround the TM change in response to acoustical and mechanical trauma. Given its fixed charge concentration, the TM may play an important role during trauma.

High-Speed Permanent Magnet Motor Generator for Flywheel Energy Storage
Tracey C. Ho

This thesis is part of a joint project between MIT and SatCon to develop a high-speed motor-generator for a flywheel energy storage system. Such systems offer environmental and performance advantages over chemical batteries, with potential applications in hybrid electric vehicles and uninterruptible power supplies. The development of high-energy Neodymium Iron Boron magnets, as well as advances in composites, electric drives and magnetic bearings, has helped make flywheel systems more commercially viable. A 30 kW high-speed permanent magnet synchronous motor-generator was designed, built and tested. The basic electric design is due to Professor James Kirtley, while much of the mechanical design is from SatCon. This thesis covers the design of the cooling system, construction details, and the development of theoretical models for thermal performance and loss mechanisms. The modeling of rotor magnet eddy currents is of particular interest, since the high-energy magnets have appreciable electrical conductivity. At present testing is still in progress but it should be completed soon. Experimental data for all the quantities predicted by our electrical, mechanical and thermal models will be collected and analyzed, in order to evaluate these models.

Improved Performance of a Virtually Imaged Phased Array for Optical Demultiplexing
Afsana N. Akhter

In fiber optic communications, the transmission of multiple optical channels over the same fiber provides unprecedented capacity. Optical demultiplexing is a critical function in high density wavelength division multiplexing for channel spacing of 1 nanometer or less. A Virtually Imaged Phased Array (VIPA) produces a large angular dispersion that can be applied to optical demultiplexing. A numerical simulation of the VIPA demultiplexer has been constructed and used to develop VIPA designs with improved performance. Results show that a glass plate with linear transmissivity increases VIPA's efficiency. If the plate in the VIPA has non-uniform thickness, considerable beam distortion occurs, which increases channel crosstalk. It is demonstrated that the distortion due to a linearly varying plate thickness can be corrected using a proper lens system.

An Analysis of Cross Gain Saturation Effects in Systems of Cascaded Erbium-Doped Fiber Amplifiers
Dedric A. Carter

For centuries people have sought methods of communicating from a distance. These methods have evolved from wireless communications in their most basic and natural sense to wireline communications. The invention of the laser in the 1960's moved wireline communications into a new realm of possibility, with lightwaves carrying information for communications through small distances. The inherent attenuation effects of silica fiber, the chosen medium for light propagation, made long-haul communications very expensive and not very practical. The invention of the Erbium-Doped Fiber Amplifier (EDFA) in the late 1980's introduced a whole new era of lightwave communications. As data rates increase, the EDFA introduces problems into all optical systems in the form of Cross Gain Saturation (CGS) Effects . As the input power to optical systems changes, gain fluctuations occur on the remaining channels causing modulation and phase shifts. This is the first work to examine real world communication systems under CGS conditions. A characterization of CGS effects in a real 4 link 400km 10Gb/s transmission system is presented, and a solution to the problem of CGS is proposed in the form of an all-optical technique called overpumping. The contributions of this work are in the area of noticing non-linear effects in amplifier cascades constructed using actual fiber, realizing the impact of the undershoot transient on the systems bit error rate, demonstrating that the receiver pre-amplifier dramatically manifests the problem of CGS, and demonstrating the effectiveness of a single and double stage overpumping solution to the CGS problem.

SESSION F

CSFS: A New Encrypted File System with Group Sharing and Integrity Protection
Kevin E. Fu

Traditional cryptographic storage uses encryption to ensure confidentiality of file data. However, encryption can prevent efficient random access to file data. Moreover, no cryptographic storage file system allows file sharing with similar semantics to UNIX group sharing. The Cryptographic Storage File System (CSFS) provides confidentiality and integrity of data while enabling efficient random access and file sharing using mechanisms similar to UNIX groups. CSFS uses a delayed-write-encryption policy for caching, delayed re-encryption for distributed re-encryption, and a hash tree structure beneath the inode for integrity. While maintaining confidentiality and integrity, the cost of reading a block is O(1) amortized over a sequential read of the entire file of n blocks. Writes execute in worst-case O(lg n) time. CSFS also implements user authentication on the file server to overcome a serious loophole in the NFS security architecture.

We present a protocol for authenticating an individual's membership in a group without revealing that individual's identity and without restricting how the membership of the group may be changed. Existing protocols that authenticate membership by identifying individuals do not provide anonymity. Those in which members share a common key require a new key to be distributed whenever an individual leaves the group. To overcome these limitations we introduce the \emph{verifiably common secret encoding} to construct anonymous authentication protocols. These protocols both authenticate membership without identifying the member and enable a trusted third party to add and remove members of the group instantly, in a single message to the authenticator. Applications in electronic commerce and communication can now provide anonymous authentication while accommodating frequent changes in membership. Because a verifiably common secret encoding grows linearly with the size of the group, we describe techniques for partitioning groups to improve performance.

We present improved algorithms for testing monotonicity of functions. Namely, given the ability to query an unknown function f : Sigma -> Xi, where Sigma and Xi are finite ordered sets, the test always accepts a monotone f, and rejects f with high probability if it is epsilon-far from being monotone (i.e., every monotone function differs from f on more than an epsilon fraction of the domain). For any positive epsilon, the query complexity of the test is O((n / epsilon) * log |Sigma| * log |Xi|). The previous best known bound was O((n^2 / epsilon) * |Sigma|^2 * |Xi|). We also present an alternative test for the boolean range Xi = {0,1} whose query complexity O(n^2 / (epsilon)^2) is independent of alphabet size |Sigma|.

Query-Efficient Checking of Proofs and Improved PCP Characterizations of NP
Venkatesan Guruswami

Recent research in the theory of computing has shown the following intriguing result (known as the PCP Theorem). "One can write proofs of mathematical assertions in such a manner that there exists a probabilistic verifier that looks at the proof in only a constant number of bit positions that is independent of the length of the proof, and satisfies: (Completeness) For every valid theorem there exists a proof that is always accepted; and (Soundness) For invalid assertions every purported proof is rejected with 50% probability." This result characterizes the fundamental complexity class NP in terms of languages that have such "probabilistically checkable proofs" (PCP), and also has significant applications to approximately solving important optimization problems. The appearance of this result spurred a lot of research devoted to quantitative strengthenings of it, and these culminated in the striking recent result that the constant number of queries above can be as small as three, provided we allow for a small but non-zero probability that even a correct proof is rejected. We give a "tight" PCP construction that makes only 3 queries to the proof and always accepts "correct" proofs, thus proving that the constant above can really be 3 (it is known that it cannot be smaller than 3). We also show how to get improved soundness for a slightly higher number of queries.

SESSION G

If It Works, It's Not AI: A Commercial Look at Artificial Intelligence Startups
Eve M. Phillips

The "AI Bubble" of the 1980s has much to teach us about the difficulty of bringing new technology products to market and creating lasting businesses. This period was marked by a large number of startup firms, such as the LISP machine makers, AI languages firms, and expert systems companies; most of these companies, however, failed to come close to their original projections. From their successes, we can see examples of positive methods of introducing new technologies into the marketplace. From their stumbles, we can see many warnings for high tech firms: management inexperience and academic bias; business models which confuse products and consulting; misunderstanding the target market; and failing to manage customer and press expectations. I hope to provide some warnings and suggestions for any company trying to build a business around a new technology.

I will examine the issues and the value-added, from both a technical and economic perspective, of solving the problem of integrating information in the retail banking industry. I will outline a systematic approach for approaching the problem of integration and report on an early implementation of a prototype for the Universal Banking Application using technologies available today. I will also describe some of the issues our research group discovered and suggest improvements for future work.

Given the recent explosion of the Internet, the trade of information goods and services may become the dominant form of commerce in the future global economy. Software agents, or semi-autonomous computer programs that run with little or no user input, may play a significant role in this information economy, perhaps even as direct participants in economic transactions. Such software agents might trade different types of information commodities with other software agents in return for electronic currency, ultimately to benefit their human user. This thesis addresses several problems that must be resolved before such an economy of software agents could become a reality, and offers some potential solutions to these problems. We consider a number of different electronic payment systems in the context of an agent economy, and then describe the design and implementation of an electronic payment system in Magenta, a fully open and extensible multi-agent platform developed at IBM Research. We discuss a number of different mechanisms that could protect consumers from dishonest merchants, such as reputation, escrow, and certification mechanisms; escrow and reputation mechanisms are also implemented in Magenta.

Implementation of a Congestion Manager and its Application to Concurrent Transmission Control Protocol (TCP) Connections
Hariharan S. Rahul

The success of the Internet to date has been in large due to the sound congestion control principles on which its dominant protocol, Transmission Control Protocol (TCP), is based. Since most traffic in the Internet has been dominated by long-running TCP flows, the network has shown relatively stable behavior and has not undergone large-scale collapse in the past decade. However, trends in traffic patterns today threaten the long-term stability and efficiency of the Internet. Web applications use multiple, independent and short concurrent TCP flows. The short durations of transfers provide TCP insufficient information to adapt to network congestion, while independent concurrent connections compete with each other. There is also an increasing use of applications like real-time streaming audio over UDP which do not adapt appropriately to congestion. These trends make it likely that large portions of the network might suffer congestion-triggered collapse due to unresponsiveness in the face of congestion or aggressive mechanisms to probe for spare bandwidth. This thesis presents the implementation (in Linux) of a novel framework, called the Congestion Manager (CM) for managing congestion from an end-to-end perspective, and allowing applications to adapt to network congestion. The CM maintains congestion parameters and exposes a powerful API to enable applications to learn about network characteristics, pass information to the CM, and schedule data transmissions. The thesis also describes how TCP can be implemented using the CM. Our performance results show that an ensemble of concurrent TCP connections can effectively share bandwidth and obtain consistent performance, without adversely affecting other network flows. We conclude that the CM provides a useful and practical framework for building adaptive Internet applications with predictable and consistent performance.

SESSION H

Embedded systems must exhibit reactive behaviors that can rapidly respond to stimuli, yet keep a long-term goal in sight. The subsumption architecture attempts to solve this issue by presenting a system in terms of a network of horizontally layered behaviors implemented using Augmented Finite State Machines (AFSMs). Subsumption works very well but has not found its way into widespread use. One possible reason is the mindshift required to express behaviors in terms of a state machine. Another barrier is the traditional implementation of subsumption architectures in hardware or specialized programming languages. This thesis describes an attempt to address both issues using a subsumption architecture implemented in Java and an encapsulation system designed to allow fragments of robot behaviors coded using traditional methods to be embedded in AFSMs. The approach allows a bottom-up incremental approach to system implementation while encouraging an eventual reimplementation from the top down, and was introduced to contestants in the 1999 6.270 Autonomous Robot Design Competition with a fair degree of success.

Code safety has become an increasingly important concern as the frequency of program distribution has increased, and there is a need for a safety system that can adapt to new threats and specific needs without sacrificing performance or usability. I discuss issues in applying a new approach to code safety by adapting the Naccio architecture to enforce safety policies on Microsoft Windows executables. Naccio is a platform-independent architecture for defining safety policies and enforcing them by transforming programs. Naccio/Win32 transforms Windows executables by inserting safety checking code and modifying program code so that safety checking cannot be circumvented. The transformed application can be run normally without interfering with other applications or requiring any specialized operating system support. I describe a prototype implementation of Naccio/Win32 and report its effects on the behavior and performance of sample applications under a variety of policies.

Existing wireless networks are designed with static physical layers that limit performance and lead to inefficient use of the spectrum. Since the physical layer of a wireless network link is usually designed to meet the worst case channel and traffic conditions, this leads to poor use of the available spectrum resources and limits performance when the conditions are better than the worst case. In addition, applications have varying throughput requirements and error tolerances, but the processing in the physical layer cannot be tailored to optimize link performance for specific applications. Software radio technology provides a mechanism for dynamically modifying the physical layer, allowing functions such as coding, modulation, and filtering to be changed "on the fly". By exploiting software radio technology, the physical layer of any network link can be modified with minimal interruption in service. The ability to modify the physical layer has two primary benefits: dynamic optimization for changing environmental conditions and better utilization of the spectrum.

An Automatic Annotation System for Audio Data Containing Music
Janet Marques

This presentation will describe an automatic annotation system for audio files. Given a sound file as input, the application outputs time-aligned text labels that describe the file's audio content. Audio annotation systems such as this one will allow users to more effectively search audio files on the Internet for content. In this annotation system, the labels include eight musical instrument names and the label 'other'. The musical instruments are bagpipes, clarinet, flute, harpsichord, organ, piano, trombone, and violin. The annotation tool uses two sound classifiers. These classifiers were built after experimenting with different parameters such as feature type and classification algorithm. The first classifier distinguishes between instrument set and non instrument set sounds. It uses Gaussian Mixture Models and the mel cepstral feature set. The classifier can correctly classify an audio segment, 0.2 seconds in length, with 75% accuracy. The second classifier determines which instrument played the sound. It uses Support Vector Machines and also uses the mel cepstral feature set. It can correctly classify an audio segment, 0.1 seconds in length, with 67% accuracy.

MIT Department of Electrical Engineering & Computer Science