Matching Items (753)
Filtering by
- Creators: Computer Science and Engineering Program
Description
The following addresses the challenge of effectively connecting mentors with student-run ventures at Arizona State University (ASU). Based on observations and interviews conducted with project sponsor Dr. Byrne and other mentors, the existing informal, referral based approach is inefficient, meaning that potential mentorship opportunities were lost.
To streamline the process, a web platform was developed. This site enables ventures to create structured profiles highlighting concise value propositions and other key indicators, empowering mentors to proactively identify suitable ventures. The technical implementation utilized Next.js for the frontend, Firebase and Firestore for the authentication and storage, and TailwindUI for styling. The result is a user friendly and scalable minimal viable product.
ContributorsMulderink, Matthew (Author) / Osburn, Steven (Thesis director) / Byrne, Jared (Committee member) / Barrett, The Honors College (Contributor) / Computer Science and Engineering Program (Contributor) / Tech Entrepreneurship & Mgmt (Contributor) / Dean, W.P. Carey School of Business (Contributor)
Created2025-05
Description
Machine learning is defined as: the use and development of computer systems
that are able to learn and adapt without following explicit instructions, by using
algorithms and statistical models to analyze and draw inferences from patterns in
data. Neural Networks are one such system that uses a series of connected nodes
(called neurons) where each connection is adjustable according to certain parameters
called ”weights”. Additionally each neuron has an adjustable bias value which adds
a fixed amount to the sum of the neurons connections. Deep learning is an algorithm
for tuning the parameters (i.e. weights and biases) of a network in order to best fit a
given problem.
For this project the problem I have selected is that of symbol recognition. I am
using the MNIST Handwritten Digit dataset which contains 70,000 images of digits
(0-9). Each image is a 28x28 grid of pixels with values from 0-255. The goal of my
system is to take an image and produce the matching 7 segment display representation
of the number in the image.
The goal of this project was to investigate methods for reducing the cost to complete
this identification task. These methods are separated into three main sections:
1. Topology
2. Knowledge Distillation
3. Network Pruning
The Topology section investigated the impacts of changing the layer sizes of a
network. In this section I found that it is better to have more connections to the
output layer than to any other layer in the network. This makes sense as the output
layer is what we expect to have the results we are looking for and so giving it more
data allows for better differentiation.
The Knowledge Distillation section focused on a training method of the same
name. This method involves the use of a larger, well trained teacher model. This
model is used as an example for a student model to try and mimic. I found that
this setup can work very well, with the student often outperforming the teacher after
the same amount of training. However, the target of the training must be chosen
carefully to avoid interfering with the student’s learning process.
The final section focused on network pruning. Pruning is a process that happens in
biology to remove weak connections to make a neural network more efficient. I found
that automatically removing connections throughout the training process worked exceptionally
well with results of our pruned network matching the control. However, I
did find that more investigation is needed to identify which connections are the most
important before removing them at the start.
ContributorsFrink, Ethan (Author) / Osburn, Steven (Thesis director) / Bazzi, Rida (Committee member) / Barrett, The Honors College (Contributor) / Computer Science and Engineering Program (Contributor)
Created2025-05
Description
Conventional four-point-probe (4PP) stations achieve high-accuracy sheet-resistance measurements but often lack the ability to perform mapping across an area of a sample. Commercial tools also feature a large (80-100 mil) probe spacing, which limits the spatial resolution of the sheet-resistance measurement.We retrofitted an R-θ-Z wafer stage with a 3-D-printed probe head and spring-loaded, 410 µm-diameter gold pins, controlled through a new Python automation stack, to build a wafer-mapper. While the gold probe pins work well on metal films, sheet-resistance measurements on silicon samples require the probe tips to mechanically pierce the native SiO₂ that spontaneously grows on silicon wafers. For device-grade specimens this mechanical scratching is undesirable,because it could lead to the introduction of mechanical defects. Initial experiments were conducted that applied high-voltage (≤ 105V), low-current (≤ 1 mA) pulses to break down the oxide electrically. However, tip deformation increased the effective contact area, raising the breakdown voltage beyond practical limits and preventing reliable contact formation, causing large variations in the mapping data. We therefore explored a contact-less eddy-current approach using a single-loop RF coil. The RF excitation signal was swept from 100 kHz to 6 GHz while its complex reflection coefficient ( S₁₁ ) was captured. The resulting resonance-splitting or “fan-out” of S₁₁ spectra correlates monotonically with the sheet resistivity of test wafers (1-140 Ω □⁻¹). LTSpice models of the coil-wafer system reproduced the measured trends, lending confidence that calibrated peak-tracking can yield quantitative resistivity maps. This work demonstrates the feasibility of a hybrid probe station that performs non-contact characterization of bulk silicon samples. In future iterations this characterization technique can also be applied to thin-film measurements. Key design lessons and an outline for refining the probe head and extraction algorithms are presented.
ContributorsStringer, Evan (Author, Co-author) / Goryll, Michael (Thesis director) / Celano, Umberto (Committee member) / Barrett, The Honors College (Contributor) / Electrical Engineering Program (Contributor) / Computer Science and Engineering Program (Contributor)
Created2025-05