Skip to main content

2019 Fall

Semester Short
20199

Optimized Assembly Tests from Machine Learning Correlation Algorithms

We explore ways of developing more efficient x86 assembly tests which check for functional coverage. The current testing suite used by our sponsor contains redundancies caused by the pseudo randomly generated test files. These redundancies lead to wasted computational time and resources. In our solution, we analyze pseudo randomly generated tests with known functional coverage checks, or "bin hits," using machine learning. For the training phase, our algorithm looks for common sequences of assembly instructions within tests that hit a given bin.

Building Energy System Modeling using Machine Learning

Predicting the future state and energy consumption of office buildings with an interpretable model can be highly useful for both study by designers of buildings and control systems alike. We present a comparison of different modeling approaches, from recurrent neural networks to a hybrid modeling approach which combines an accurate deep state predictor with an easily interpretable linear energy predictor.

Team Members: 

Sameer Bibikar

Rohan Koripalli

Srinjoy Majumdar

Christopher Mao

Vivian Nguyen

Mock Data Generator

An application to generate mock-data for Teradata to test their algorithms. Teradata would like to test their algorithms without violating the privacy of their customers. Therefore, our application generates realistic data to help the testing of those algorithms.

Team Members: 

Mahmood Alam

Waseem Mehany

David Terral

Asset Issuance Application for Blockchain System and BlockReduce

Bitcoin is arguably the most prolific cryptocurrency available and was the first to employ blockchain technologies. Our senior design project consists of adding asset issuance to Bitcoin and providing a platform for the asset owner to sell and trade their assets. Adding assets will expand Bitcoin’s overall functionality, thus making it appealing to a more expansive audience.

Team Members: 

Donald Maze-England

Jennifer Sin

Humza Tariq

Jonathan Wu

Lauren Wolf

Acceleration of Deep Learning on a FPGA

Modern machine learning methods based on deep neural networks commonly rely on hardware acceleration of critical computational kernels in order to improve performance and reduce energy consumption. Our project explores this new approach by mapping a convolutional neural network onto an FPGA to classify images faster than a traditional CPU-only implementation.

Team Members: 

Nafis Akbar

Christian Han

Nick Morin

Ghoshank Patel

Machine Learning for Prediction and Analysis of Public Financial Data

The U.S. Securities and Exchange Commission requires publicly traded companies to file reports of their financial statements every quarter. Our project utilized this data along with daily stock price information for nine different companies to create an interactive web application. Users can input financial data on the front-end, and our machine learning algorithms will cluster companies with similar stock trends and make predictions on the future value of the queried stock.

Team Members: 

Rajat Ahuja

Casey Hu

Seungjun Lee

Utileye: A Tool to Monitor Room Utilization

In cities and crowded areas around the world, space is becoming an increasingly valuable commodity. Office building and university administrators, particularly, have an increased need to optimize room usage. With Utileye, we aim to directly address this problem by creating a battery operated device that can detect room utilization over a sustained period of time. The development of Utileye involved a significant embedded systems component, a custom PCB, and a user-facing website for monitoring and configuring devices.

Wireless Power Transfer System for Electric Vehicles

Our project is a wireless power transfer system that provides a platform for future electric vehicle charging systems. The system uses inductive coupling to transfer power across a transmitting and receiving coil that is separated by an air gap. The transmitting coil is powered by a DC/AC converter with a user adjustable frequency that determines the system's output power. The receiving coil produces a proportional AC current which is then rectified and can be applied to a load or charging system.

User simplified modular cloud-based machine learning using hardware accelerators

A software module that allows a user to be able to input an image dataset to determine what the cost and train time will be on Google TPU's and GPUs running on TACC. It also allows for a simplified training process for users who lack knowledge in the areas of machine learning.

Team Members: 

Qinyu Diao

Elvin Galarza

Andy Hoang

Shiv Lalapet

Attila Szabo

Subscribe to 2019 Fall