Date :12th May, 2017
Presenter: Uday Shankar Shanthamallu, PhD Student, SenSIP Center, Arizona State University
Introduction to Basic Paradigms of Deep Learning in the field of Visual Computing
Abstract
This is a short informal tutorial and introduction to basic paradigms for Deep Learning with focus on Visual Computing. There will be approximately 3-4 presentations on various topics. The tutorial begins with basics of visual representation and introduction to artificial neural networks, training of neural networks using back propagation algorithm. Different optimization techniques such as stochastic gradient descent and its variants are discussed. It will also talk about various activation functions used in neural networks and their advantages. Loss function such as softmax loss are discussed. Several ways to optimize the neural networks are presented. An introduction of Convolutional neural networks (CNN) is given along with the discussion on various architectures in literature. Some of the latest paradigms in deep learning such as Autoencoder, de-noising autoencoders and Generative Adversarial Networks (GAN) are presented.
Date: 21 st April, 2017 Time: 2:00PM
Presenter: Yishan Jiao, PhD Student, Arizona State University
Minimally supervised and interpretable automated speech analysis for clinical applications
Abstract
Neurological disorders can affect a person’s ability to communicate. To improve the efficiency and effectiveness of communication, some patients visit a speech-language pathologist (SLP) for evaluation and treatment. To formulate a treatment plan, the SLP will first listen to the patient’s speech and provide a subjective assessment based on his/her perception, and then develop an intervention strategy based on the assessment. However, subjective assessment can be unreliable and biased due to the familiarity of a clinician with his/her patients. Our work focuses on developing automated speech analysis tools to facilitate clinical work and complement subjective evaluation. In this seminar, we present a series of recently developed automated acoustic measures, which are interpretable to both clinicians and patients. These include online speaking rate estimation, a measure of articulatory precision, a set of interpretable phonological features, among others. Since the amount of the annotated pathological data is limited, the measures are either minimally supervised or completely unsupervised. We will provide the details of the algorithms and the evaluation on real data from patients with various neurological diseases. We will conclude with a discussion of future work based on feedback from clinicians at the Mayo Clinic.
Date: 7th April, 2017 Time: 2:00PM
Presenter: Raksha Ramakrishna, PhD Student, Arizona State University
A Compressive Sensing framework for the analysis of Solar Photo-Voltaic Power
Abstract
Given the rise in solar power generation it has become imperative to help with its integration into the electric grid operations. The challenge here is that solar energy is not entirely deterministic due to weather and atmospheric effects which introduces stochasticity and variability in the system. In this presentation, a new stochastic model for solar Photo-Voltaic (PV) power that explicitly models cloud cover as a random mask and the effect of this cloud coverage as an attenuation of the deterministic solar irradiation pattern is elucidated. Relying on compressive sensing methods we are able to fit a set of solar PV power data from California with the components of this stochastic model and extract the parameters of such a process thus effectively capturing the variability of solar power production. We then leverage the rich information coming from this parametric model for providing solar power forecasts over a short horizon.
Date: 17th March, 2017 Time: 3:00PM
Presenter: Sai Zhang, PhD Student, Arizona State University
Parameter and Structure Estimation in Wireless Sensor Networks
Abstract
Distributed wireless sensor networks (WSNs) are important and widely used in various applications, such as military surveillance, health care and distributed estimation. There are many advantages of using distributed WSNs, including robustness to link failures and scalability. In this presentation, we have addressed the problem estimating the structure of distributed WSNs. Firstly, a distributed consensus algorithm for estimating the number of nodes in a WSN in the presence of communication noise is introduced. Then, a distributed network degree distribution and degree matrix estimation algorithm is proposed. Finally, a distributed algorithm for estimating the center and the area of a WSN is described. The network area estimation problem is formulated as a convex optimization problem and distributed optimization methods are used. The performance of the algorithms are analyzed, and simulations corroborating the theory are also provided.
