James Webb Telescope Issue - 31

Graduate Research SP
William Oswald
Theta Lambda
University of South Alabama, Ph.D. Student in Systems Engineering
RESEARCH TOPIC
Memory Optimizations In Machine Learning Systems
It was only 2005 when Intel released its first commercial CPU to offer dual-core
processing, whereas in 2020 Nvidia released a single GPU with 6912-cores
of computational power, specifically to meet demands for deep learning
applications. With demand ever increasing for computational performance, the
entire computational system needs to evolve to meet demand. As these deep
learning models increase in computational complexity, so does the demand on
hardware. This hardware demand in deep learning systems remains a problem for
Internet of Things (IoT) devices, spacecraft, and even large data centers; therefore,
reducing the hardware demands of deep learning will affect computer systems
at many different levels. Specifically, deep learning applications require a massive
memory footprint which drives up power demands. These memory circuits usually
consume several orders of magnitude more energy than the computational
circuits. Subsequently, achieving greater energy efficiency in memory is one
of the key design considerations for deep learning. Designers have developed
various different low-power memory techniques, however, existing power-efficient
memories usually come with a significant overhead. The deep learning systems
I aim to optimize include Video Streams, and hyper-spectral imaging samples
for cancer detection. I chose these systems to investigate because of the high
memory demands in these two research topics.
LEARN MORE
https://scholar.google.com/citations?user=5EuCKBgAAAAJ&hl=en&oi=sra
Kurt Butler
Theta Mu
Stony Brook University, Ph.D. Student in Electrical Engineering
RESEARCH TOPIC
Causal Inference and the Neuroscience of Consciousness
While the origin of consciousness has been hotly debated since the inception
of humanity, modern neuroscience has revealed that the networked activity of
multiple brain regions may be essential to supporting conscious experience. To
develop mathematical models that describe how the network functions, it is
important to distinguish causation, based on which one brain region activates
another directly, from the case where both regions might both be driven by a
common cause. Even with multichannel brain signal recording technologies
such as electroencephalography (EEG) and local field potential (LFP), it is
impossible to observe the entirety of the brain at once, and so it is important
to explicitly consider hidden states when modeling brain activity. By discovering the mechanisms that underlie consciousness,
researchers hope to improve the clinical treatment of patients with traumatic brain injuries or in comas.
Kurt's work has focused on using machine learning, and in particular, the use of Gaussian processes (GPs) to model the complex
signals that arise from the nonlinear and stochastic behavior of neural populations. His recent work has focused on detecting and
quantifying the strength of causal interactions between brain regions. He also uses GPs to describe how the complicated brain
signals observed by EEG and LFP can be produced by unobserved low-dimensional state spaces.
The low-dimensional representation of a brain state is
estimated (left) and the model learned using GPs reproduces
similar structures in its predictions (right).
LEARN MORE
https://sites.google.com/view/kurt-butler/home
CONTACT
https://selcrec.ece.stonybrook.edu/
For 3D Hyper-Spectral Images, the data can be
compressed in meaningful ways before reaching
the Neural Network.This shows the logical
flow of data through a proposed algorithm,
starting from the raw 3D data collected from
the laser, to flattening, Principal Component
Analysis (PCA) compression, and then Region of
Interest (ROI) extraction. This is a very abstract
oversimplification of the algorithm, but serves as
a useful visual aid to the logical flow.
TLIGHT
Video Stream Decoding Process,
Deemed Content-Adaptable
ROI-aware low-power video
memory. The direction
of an arrow indicates the
development from one previous
technique to a newer one.
CONTACT
linkedin.com/in/william-liam-oswald-915871171
A Bottom-Up example of a Neural Network
Search system, including novel additions to
hardware and algorithm memory optimizations.
This image contains all research topics and
goals of my dissertation moving forward.
HKN.ORG
31

https://scholar.google.com/citations?user=5EuCKBgAAAAJ&hl=en&oi=sra https://www.linkedin.com/in/william-liam-oswald-915871171/ https://sites.google.com/view/kurt-butler/home https://selcrec.ece.stonybrook.edu/ https://hkn.ieee.org/

James Webb Telescope Issue

Table of Contents for the Digital Edition of James Webb Telescope Issue