Matthew T. Radice

Limitations of Q-Learning for partially observable reinforcement learning domains and compensatory mechanisms via working memory.

David W. Ludwig

Deep learning models for dynamic task switching and permutation invariant representations.

Laurel Koenig

Quantum machine learning methods for improved generalization and industrial applications.

Tingting Sun

Deep learning models for image generation and GC/MS embeddings.

Divyashree Koti

Deep learning models for 3D point cloud data.

Arthur S. Williams

Hierarchical and indirection-based working memory models for task generalization in partially observable reinforcement learning domains.

Jackson L. Goble

Episodic memory models for frequency-based task generalization in reinforcement learning domains.

Karuna D. Gujar

Recurrent and transformer neural networks for predicting biophysical traits from fMRI.

Scott P. Morton

Currently researching the ElectroStatic Surface Charge Pipeline for HIV antibody binding characteristics. Additionally converting all Bash scripts into a unified python language format with a json based configuration file. Further research goals include a rewrite of frodaN for ease of use and parallel processing potentials involving CUDA and/or MPI, Massively Parallel Processing (MPP) of model protein sequences for predicting binding characteristics of HIV and HIV antibodies, potential electronic circuitry characteristics of protein sequences as a means of binding energy.

Ivan Syzonenko

Ivan's research has focused on understanding and overcoming the limitations of machine learning algorithms, particularly clustering, in biomolecular simulation analysis. More recently, he is developing and testing methods for accelerated unbiased targeted MD simulations using classical pathfinding approaches with a focus on protein folding.

Heena Khan

Comparison of various machine learning and deep learning models for the identification of Islamophobic content on social media.

Terryn J. Seaton

OpenLDAP deployment for a Singularity/SLURM HPC cluster in the cloud via Kubernetes.

Chaning B. Mullinax

Nested LSTM/Indirection models of working memory for human-like extrapolatory generalization.

Will H. Haase

Minimizing catatrophic interference via unitization and generative recurrent neural network models.

Nibraas A. Khan

Combined working memory and N-task models for partially-, non-observable reinforcement learning problems.

Lucas Remedios

Keras/TensorFlow tools for N-task learning.

Huizhi Wang

Dimensional attention mechanisms for holographic reduced representational encodings and category learning.

Ngozi C. Omatu

Dimensional attention for working memory: accelerated early learning with asymptotically optimal performance.

Mike Jovanovich

Neurobiologically plausible models of working memory, task switching, and task generalization.

Joshua M. Arnold

Coupled action-working memory learning for partially observable reinforcement learning problems.

Cody Newbold

Journalism today is dealing with so much data that better methods are needed to process it. Latent Dirichlet Allocation (LDA) is often used to sort text into topics. The Afghan War Diary (AWD) was processed with LDA and model trees to ascertain fatality numbers. The AWD was used in a separate study that analyzed the documents with point process modeling (PPM) to predict where conflicts would occur in space and time. We have combined the two approaches in this study, hopeful that the results will allow us to predict where and when conflicts occur and if the fatality numbers can also be obtained in reference to that. We anticipate that our results will show that PPM combined with LDA and model trees will give more useful results than using either of the methods separately.

Jonathan Howton

Jonathan's research focuses on using high-throughput structural analysis to understand how environmental factors affect protein binding particularly with regard to viral transmission.

Grayson Dubois

Grayson's research looks into a new method of representing concepts in artificial neural networks (ANNs) that mimic human working memory systems. This new method uses something called Holographic Reduced Representations (HRRs), which are powerful tools capable of representing compositional structure in distributed representations. Grayson has developed an engine for encoding and decoding HRRs and is working on integrating it into the Working Memory toolkit, a software library written in ANSI C++ designed to allow researchers to write simulations of learning tasks using ANNs modeled after working memory. The current toolkit requires the programmer to explicitly provide methods to convert concepts used by working memory from symbolic encodings (SE) to distributed encodings (DE). Grayson's HRR Engine will automate the process of SE/DE conversion, thus taking the burden off of the programmer and opening the door to future possibilities not possible with previous DE methods. These include but are not limited to the chunking of similar concepts in memory, transferability of learned behaviors between tasks, and long term memory.

Gary Hammock

Gary's research focuses on the development and testing of novel simulation-inspired methods for cryptography.

Robert Myers

Robert's research focused on the development and testing of distributed pathfinding algorithms.

Michael Murphy

Fractal dimension is a number that describes the self-similarity, or "complexity", of a geometry. In image processing, fractal dimension is often used as a novel method for contrasting and comparing image content. The Box-Counting Algorithm is one of the most popular methods of computing an estimate for the fractal dimension of an image, but the algorithm is influenced by many factors such as filtering and noise. Our research found a relationship between dimensional estimations and the variability in those estimations when using the Box-Counting Algorithm in the presence of increasing levels of uniform noise. This relationship provides a way to strengthen relative dimensional rankings between noisy images.

Stephen Kinser

Stephen's research focused on converting a semi-autotmated protein electrostatics pipeline to a fully-automated version which requires significantly less user intervention.