Sparse coding is an influential unsupervised learning approach proposed as a theoretical model of the encoding process in the primary visual cortex (V1). While sparse coding has been successful in explaining classical receptive field properties of simple cells, it was unclear whether it can account for more complex response properties in a variety of cell types. In this dissertation, we demonstrate that sparse coding and its variants are consistent with key aspects of neural response in V1, including many contextual and nonlinear effects, a number of inhibitory interneuron properties, as well as the variance and correlation distributions in the population response. The results suggest that important response properties in V1 can be interpreted as emergent effects of a neural population efficiently representing the statistical structures of natural scenes under resource constraints. Based on the models, we make predictions of the circuit structure and response properties in V1 that can be verified by future experiments.

In this thesis we modify the Made-To-Measure (M2M) algorithm to be dynamically selfconsistent and apply it to the problem of generating equilibrium collisionless systems with non-spherical halos. Our M2M algorithm systematically adjusts the masses of particles in a system slowly, keeping the system in equilibrium. The adjustments are performed according to some given constraints and proceed until pseudo-observations of the system match the constraints. We use this algorithm to generate isolated triaxial dark matter halos and disk-halo systems with prolate halos. The isolated triaxial dark-matter halo simulations provide a test for the algorithm. These tests show that our algorithm can generate equilibrium collisionless systems with non-spherical halos, but we also find that our algorithm requires a large amount of computational time to converge to the final target system. The disk-halo simulations show that prolate halos modify the morphology and velocity profile of dark matter dominated disks that cause errors in the measurement of the inclination and understanding the rotation curve. As a result of these errors, a mass estimate from the observed rotation curve of a disk in a prolate halo will depend on the observers position relative to the disk. The mass estimates from the same disk observed at different positions may vary by up to a factor of three. / Thesis (Master, Physics, Engineering Physics and Astronomy) -- Queen's University, 2010-04-13 18:32:29.294

Machine translation has advanced considerably in recent years, primarily due to the availability of larger datasets. However, one cannot rely on the availability of copious, high-quality bilingual training data. In this work, we improve upon the state-of-the-art in machine translation with an instance-based model that scores each instance of translation in the corpus. A translation instance reflects a source and target correspondence at one specific location in the corpus. The significance of this approach is that our model is able to capture that some instances of translation are more relevant than others. We have implemented this approach in Cunei, a new platform for machine translation that permits the scoring of instance-specific features. Leveraging per-instance alignment features, we demonstrate that Cunei can outperform Moses, a widely-used machine translation system. We then expand on this baseline system in three principal directions, each of which shows further gains. First, we score the source context of a translation instance in order to favor those that are most similar to the input sentence. Second, we apply similar techniques to score the target context of a translation instance and favor those that are most similar to the target hypothesis. Third, we provide a mechanism to mark-up the corpus with annotations (e.g. statistical word clustering, part-of-speech labels, and parse trees) and then exploit this information to create additional perinstance similarity features. Each of these techniques explicitly takes advantage of the fact that our approach scores each instance of translation on demand after the input sentence is provided and while the target hypothesis is being generated; similar extensions would be impossible or quite difficult in existing machine translation systems. Ultimately, this approach provides a more exible framework for integration of novel features that adapts better to new data. In our experiments with German-English and Czech-English translation, the addition of instance-specific features consistently shows improvement.

The amount of multimedia content available via the Internet, and the number of formats in which it is encoded, stored and delivered continues to grow rapidly. So too the number and diversity of the devices and software applications which produce, process and consume such content. This constantly changing landscape presents an increasing challenge to interoperability, since more and more software and hardware must be upgraded as new formats are developed. However, many of the operations performed on multimedia content are similar across coding formats. In recognising this, this thesis proposes several approaches to format-independent media processing, with an emphasis on content delivery. This considerably simplifies interoperability, since support for a new content format may be provided by disseminating a data file, rather than requiring application and device providers to extend and modify their software and hardware. A fundamental requirement for format-independence is the ability to describe the structure of any given format in a way that exposes how it may be fragmented for delivery or processing, and how other data important to the processing (for instance temporal or scalability parameters) can be extracted from the binary data. Several meta-syntax languages are evaluated that (to greater or lesser degree) perform this function. Of these, the most suitable for general use in format-independent processors is found to be MPEG-21’s Bitstream Syntax Description Language (BSDL). Its general suitability notwithstanding, BSDL exhibits several critical flaws when used to describe and process modern content formats. In response, this thesis proposes several new features for the language which significantly reduce processing complexity, and provide extensibility for complex data types. These features are implemented and validated using bitstreams of real-world length, which enable a linear response of approximately 10 times the speed of playback (on the particular test machine used), for videos up to one hour in duration. Digital media increasingly encompasses a wide range of metadata, as well as collections of related content (a DVD and it’s “special features”, for instance). Several recent standards address generic virtual containers for such rich content. While these standards—which include MPEG-21 and TVAnytime—provide numerous tools for interacting with rich media objects, they do not provide a framework for streaming or delivery of such data. This thesis presents the Bitstream Binding Language (BBL), a format-independent tool that describes how multimedia content and metadata may be bound into delivery formats. Using a BBL description, a generic processor can map rich content (an MPEG-21 Digital Item, for example) into a streaming or static delivery format. BBL provides a universal syntax for fragmentation and packetisation of both XML and binary data, and allows new content and metadata formats to be delivered without requiring the addition of new software to the delivery infrastructure. The BBL framework is validated and tested against a number of application scenarios including a format-independent streaming server, generic metadata syntax translation, virtual container assembly, and a format-independent hinter.Finally, it is observed that much of the semantic metadata that is generated to describe multimedia content could also be used to improve the decisions that must be made in order to transmit it effectively. Indeed, methods have been proposed for using specific semantic concepts in the delivery process. However, until now, no high-level system has been proposed that is able to take arbitrary semantic metadata, and utilise it in the multimedia delivery decision-making process. This thesis proposes such a system. It combines the aforementioned semantic concepts with other existing work in Rate-Distortion Optimisation for multimedia delivery, scalable content formats, and syntax description, and then develops a generalised framework to permit an arbitrary range of semantic metadata and optimisation techniques to be utilised. This objective is accomplished by utilising schema languages to describe the details of any given content or metadata, so that declarative mapping rules can be specified for translating from format-specific data points to format-independent concepts that are directly used by the framework. This translation can then be performed using software or hardware that knows nothing about the specific format it is processing.This thesis describes a particular embodiment of the semantic-aware multimedia delivery system which was implemented in order to verify its key assertions. It presents the results of subjective testing that was performed on several short news clips encoded using H.264/SVC scalable video coding, and Scalable-To-Lossless (SLS) au dio coding. Each clip was adapted to four target bitrates, using both of two methods: (a) using the semantic-aware system to devote a greater proportion of the available bandwidth to that part of the content (audio or video) that was conveying more of the semantics at any given time; and (b) at a constant bit-rate with the same average rate as clip (a). Test participants were shown each pair of clips (a and b) in a random order and were asked to evaluate which was more successful at conveying the meaning of the story. The result of this subjective testing was a 72% preference for those clips which had been adapted so as to devote more bandwidth to the semantically-important parts of the content.

The amount of multimedia content available via the Internet, and the number of formats in which it is encoded, stored and delivered continues to grow rapidly. So too the number and diversity of the devices and software applications which produce, process and consume such content. This constantly changing landscape presents an increasing challenge to interoperability, since more and more software and hardware must be upgraded as new formats are developed. However, many of the operations performed on multimedia content are similar across coding formats. In recognising this, this thesis proposes several approaches to format-independent media processing, with an emphasis on content delivery. This considerably simplifies interoperability, since support for a new content format may be provided by disseminating a data file, rather than requiring application and device providers to extend and modify their software and hardware. A fundamental requirement for format-independence is the ability to describe the structure of any given format in a way that exposes how it may be fragmented for delivery or processing, and how other data important to the processing (for instance temporal or scalability parameters) can be extracted from the binary data. Several meta-syntax languages are evaluated that (to greater or lesser degree) perform this function. Of these, the most suitable for general use in format-independent processors is found to be MPEG-21’s Bitstream Syntax Description Language (BSDL). Its general suitability notwithstanding, BSDL exhibits several critical flaws when used to describe and process modern content formats. In response, this thesis proposes several new features for the language which significantly reduce processing complexity, and provide extensibility for complex data types. These features are implemented and validated using bitstreams of real-world length, which enable a linear response of approximately 10 times the speed of playback (on the particular test machine used), for videos up to one hour in duration. Digital media increasingly encompasses a wide range of metadata, as well as collections of related content (a DVD and it’s “special features”, for instance). Several recent standards address generic virtual containers for such rich content. While these standards—which include MPEG-21 and TVAnytime—provide numerous tools for interacting with rich media objects, they do not provide a framework for streaming or delivery of such data. This thesis presents the Bitstream Binding Language (BBL), a format-independent tool that describes how multimedia content and metadata may be bound into delivery formats. Using a BBL description, a generic processor can map rich content (an MPEG-21 Digital Item, for example) into a streaming or static delivery format. BBL provides a universal syntax for fragmentation and packetisation of both XML and binary data, and allows new content and metadata formats to be delivered without requiring the addition of new software to the delivery infrastructure. The BBL framework is validated and tested against a number of application scenarios including a format-independent streaming server, generic metadata syntax translation, virtual container assembly, and a format-independent hinter.Finally, it is observed that much of the semantic metadata that is generated to describe multimedia content could also be used to improve the decisions that must be made in order to transmit it effectively. Indeed, methods have been proposed for using specific semantic concepts in the delivery process. However, until now, no high-level system has been proposed that is able to take arbitrary semantic metadata, and utilise it in the multimedia delivery decision-making process. This thesis proposes such a system. It combines the aforementioned semantic concepts with other existing work in Rate-Distortion Optimisation for multimedia delivery, scalable content formats, and syntax description, and then develops a generalised framework to permit an arbitrary range of semantic metadata and optimisation techniques to be utilised. This objective is accomplished by utilising schema languages to describe the details of any given content or metadata, so that declarative mapping rules can be specified for translating from format-specific data points to format-independent concepts that are directly used by the framework. This translation can then be performed using software or hardware that knows nothing about the specific format it is processing.This thesis describes a particular embodiment of the semantic-aware multimedia delivery system which was implemented in order to verify its key assertions. It presents the results of subjective testing that was performed on several short news clips encoded using H.264/SVC scalable video coding, and Scalable-To-Lossless (SLS) au dio coding. Each clip was adapted to four target bitrates, using both of two methods: (a) using the semantic-aware system to devote a greater proportion of the available bandwidth to that part of the content (audio or video) that was conveying more of the semantics at any given time; and (b) at a constant bit-rate with the same average rate as clip (a). Test participants were shown each pair of clips (a and b) in a random order and were asked to evaluate which was more successful at conveying the meaning of the story. The result of this subjective testing was a 72% preference for those clips which had been adapted so as to devote more bandwidth to the semantically-important parts of the content.

Thesis (M.Sc.). / Written for the School of Computer Science. Title from title page of PDF (viewed 2008/12/09). Includes bibliographical references.

Thesis (Ph. D.)--University of Washington, 2004. / Vita. Includes bibliographical references (p. 206-213).

Diss. (sammanfattning) Uppsala : Uppsala universitet, 2009. / Härtill 5 uppsatser.

Thesis--Wisconsin. / Vita. Includes bibliographical references (leaves 118-122).

Thesis (Ph. D.)--University of Chicago, Dept. of Computer Science, August 1997. / Includes bibliographical references. Also available on the Internet.