Theses and Dissertations
This collection includes both ASU Theses and Dissertations, submitted by graduate students, and the Barrett, Honors College theses submitted by undergraduate students.
Displaying 1 - 3 of 3
Filtering by
- All Subjects: Computer Engineering
- All Subjects: Colloquial language
Description
Deep neural networks (DNN) have shown tremendous success in various cognitive tasks, such as image classification, speech recognition, etc. However, their usage on resource-constrained edge devices has been limited due to high computation and large memory requirement.
To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model.
This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs.
To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model.
This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs.
ContributorsSrivastava, Gaurav (Author) / Seo, Jae-Sun (Thesis advisor) / Chakrabarti, Chaitali (Committee member) / Berisha, Visar (Committee member) / Arizona State University (Publisher)
Created2018
Description
Glottal fry is a vocal register characterized by low frequency and increased signal perturbation, and is perceptually identified by its popping, creaky quality. Recently, the use of the glottal fry vocal register has received growing awareness and attention in popular culture and media in the United States. The creaky quality that was originally associated with vocal pathologies is indeed becoming “trendy,” particularly among young women across the United States. But while existing studies have defined, quantified, and attempted to explain the use of glottal fry in conversational speech, there is currently no explanation for the increasing prevalence of the use of glottal fry amongst American women. This thesis, however, proposes that conversational entrainment—a communication phenomenon which describes the propensity to modify one’s behavior to align more closely with one’s communication partner—may provide a theoretical framework to explain the growing trend in the use of glottal fry amongst college-aged women in the United States. Female participants (n = 30) between the ages of 18 and 29 years (M = 20.6, SD = 2.95) had conversations with two conversation partners, one who used quantifiably more glottal fry than the other. The study utilized perceptual and quantifiable acoustic information to address the following key question: Does the amount of habitual glottal fry in a conversational partner influence one’s use of glottal fry in their own speech? Results yielded the following two findings: (1) according to perceptual annotations, the participants used a greater amount of glottal fry when speaking with the Fry conversation partner than with the Non Fry partner, (2) statistically significant differences were found in the acoustics of the participants’ vocal qualities based on conversation partner. While the current study demonstrates that young women are indeed speaking in glottal fry in everyday conversations, and that its use can be attributed in part to conversational entrainment, we still lack a clear explanation of the deeper motivations for women to speak in a lower vocal register. The current study opens avenues for continued analysis of the sociolinguistic functions of the glottal fry register.
ContributorsDelfino, Christine R (Author) / Liss, Julie M (Thesis advisor) / Borrie, Stephanie A (Thesis advisor) / Azuma, Tamiko (Committee member) / Berisha, Visar (Committee member) / Arizona State University (Publisher)
Created2015
Description
A distributed wireless sensor network (WSN) is a network of a large number of lowcost,multi-functional sensors with power, bandwidth, and memory constraints, operating
in remote environments with sensing and communication capabilities. WSNs
are a source for a large amount of data and due to the inherent communication and
resource constraints, developing a distributed algorithms to perform statistical parameter
estimation and data analysis is necessary. In this work, consensus based
distributed algorithms are developed for distributed estimation and processing over
WSNs. Firstly, a distributed spectral clustering algorithm to group the sensors based
on the location attributes is developed. Next, a distributed max consensus algorithm
robust to additive noise in the network is designed. Furthermore, distributed spectral
radius estimation algorithms for analog, as well as, digital communication models
are developed. The proposed algorithms work for any connected graph topologies.
Theoretical bounds are derived and simulation results supporting the theory are also
presented.
ContributorsMuniraju, Gowtham (Author) / Tepedelenlioğlu, Cihan (Thesis advisor) / Spanias, Andreas (Thesis advisor) / Berisha, Visar (Committee member) / Jayasuriya, Suren (Committee member) / Arizona State University (Publisher)
Created2021