Graduate Researcher - Adaptive Resonance Theory & Machine Learning
Missouri S&T, Computer Science Department · Rolla, MO · Aug 2023 - Present
# Adaptive Representation Learning & Generative Modeling
- Developed ART-VQ, a discrete-continuous latent layer replacing VQ-VAE point codebooks with Fuzzy ART hyperboxes to preserve discrete region identity and continuous within-box position; introduced a projection readout that prevents representational collapse, reducing error 35x versus midpoint training.
- Achieved ~10x lower reconstruction MSE than VQ-VAE at matched token count on CIFAR-10 (1.01 vs. 10.55); rate-distortion controls attributed most of the improvement to increased representational rate. Under MNIST class shift, ART-VQ limited forgetting to +0.24 MSE vs. +19.3 for VQ-VAE.
- Built a generative pipeline combining a transformer prior over hyperbox identities with diffusion modeling of within-box positions; a pre-registered 12-seed CIFAR-10 study improved FID by 3.0 points over matched VQ (p = .033).
# Adversarial Robustness of Incremental Learners (First-Author Research)
- Developed WB-Softmax, a differentiable attack objective for non-differentiable prototype-based models, achieving 89-100% white-box attack success across USPS, MNIST, and Fashion-MNIST.
- Designed progressive two-stage selective adversarial training and separation-aware diagnostics using iCVIs; achieved best-in-class AURAC (28.2%, 64.5%, 41.3%) and improved high-perturbation USPS robustness 4x.
# DARPA Computer Vision Research
- Integrated ART classifiers into YOLO and built modular PyTorch infrastructure for classifier-head swapping, feature visualization, automated checkpointing, and SLURM multi-GPU training.