Skip to main content

Neural Network and Machine Learning Verification

Primary Investigators:
Taylor Johnson
 
Brief Description of Project:
In this project, students will help develop benchmarking processes for recent machine learning and neural network verification algorithms and tools, such as our nnv tool (https://github.com/verivital/nnv). These approaches allow, for example, to detect or prove the absence of perturbations that can cause various computer vision and machine perception tasks to misbehave, known colloquially as adversarial perturbations, but the source of which could be due to environmental uncertainty, noise, attackers, etc. Anticipated contributions include developing scripts for performing benchmarking of our methods and other research groups' recent approaches, to primarily be evaluated on convolutional neural networks (CNNs) on standard data sets, such as MNIST, CIFAR, and ImageNet.

Desired Qualifications:
Students at all levels (freshman through senior) are welcome and will be able to help refine our prototype systems and approach. Programming experience in Matlab, Java, and Python would all be desirable, as would prior experience with machine learning frameworks, such as Keras, TensorFlow, etc. All code will be version controlled using Git/Mercurial, which experience with is desired, but not required.
 
Nature of Supervision:
The adviser will hold weekly group meetings with the undergraduates, current PhD students, and postdocs, as well as approximately weekly individual meetings with undergraduate students. The current group members are available here: http://www.taylortjohnson.com/?m=people
 
A Brief Research Plan (period is for 10 weeks):
In the first 2-3 weeks, students will learn about machine learning and our existing prototype framework called nnv (https://github.com/verivital/nnv). In weeks 4-9, students will develop and test extensions to our framework, as well as developing benchmarking processes for our tools as well as other research groups' tools, evaluating on standard convolutional neural networks and image data sets, such as MNIST, CIFAR, and ImageNet. In the final week, students will prepare and submit a written report describing their prototype enhancements, accuracy evaluation, and their experience with the research program. Students will present an oral presentation on their summer research in the final week.
 
Number of Open Slots: 1
 
Contact Information:
Name: Taylor T. Johnson
Title: Assistant Professor
Department: Electrical Engineering and Computer Science
Campus Address: ISIS 401D
Mailing Address: 1025 16th Avenue South Room 401D
Nashville, TN 37212
United States
Email: taylor.johnson@vanderbilt.edu
Phone: (979) 251-6215