Projects

PROBLEM

Wanted a from-scratch understanding of how modern architectures are actually built, hand-deriving the backpropagation rather than leaning on framework abstractions, and to push that toolkit all the way to a working self-supervised model.

APPROACH

Built a modular neural network suite in low-level Python and CuPy, feed-forward, convolutional, activation, pooling, transformer and embedding layers as reusable, gradient-checked blocks, then used them to implement I-JEPA: a teacher/student/predictor pipeline with an EMA target and masked-patch latent prediction.

OUTCOME

The suite now trains a full I-JEPA end to end. Scaled from 8×8 digits up to 28×28 MNIST, the learned patch tokens reach ~0.95 linear-probe (concat) and 0.85 kNN accuracy by epoch 500, explored interactively below through their attention maps, nearest neighbours and t-SNE.

PROBLEM

Transformer-based temporal action localization models saturate performance on standard benchmarks, but degrade badly on fine-grained datasets with sub-second actions. It was unclear whether this was a localization failure or a feature-extraction failure.

APPROACH

Used a ballet dataset, AnnChor260, with an average action length of 0.66 seconds as a resolution stress test, comparing frame-level extractors (Frame-ResNet50), spatiotemporal baselines (I3D), and foundation models (InternVideo) at varying temporal strides.

OUTCOME

Found a performance inversion: simple frame-level features beat I3D by over 25% mAP on sub-second actions (31.24% vs 5.87%). InternVideo's near-saturated 89.83% mAP collapsed to 28.80% once temporal stride increased, showing token density, not model size, is the key factor for fine-grained localization.

PROBLEM

Autonomous vehicle perception models are statistical systems and therefore vulnerable to adversarial poisoning, but most attack research stops at digital perturbations without connecting them to real-world driving consequences.

APPROACH

Investigated the vulnerability of Tiny-VAD, a vectorized autonomous driving model, using the Image Joint Embedded Predictive Architecture (I-JEPA) as a novel baseline for generating semantically-aware adversarial perturbations.

OUTCOME

Ongoing coursework research at ETSU exploring whether surrogate-guided, semantically-rich adversarial features can induce measurable trajectory drift in otherwise robust perception models.

PROBLEM

AI-driven agriculture systems focus almost entirely on livestock, overlooking the wildlife, foxes, rats, pigeons, raccoons, and squirrels, that damage crops, spread disease, and prey on poultry.

APPROACH

Built two classification pipelines (VGG16 and ResNet50) trained on a compiled dataset combining existing wildlife datasets with real trail-camera footage, evaluating performance across clear and greyscale imagery.

OUTCOME

Found that models trained on clean images struggle to generalise to real trail-camera conditions (best accuracy 48.19%), establishing a baseline and surfacing the domain-gap challenge for real-world wildlife surveillance in agriculture.

PROBLEM

Every new research project meant re-writing the same boilerplate (data preparation, dataloaders, training loops, and evaluation) before any real experimentation could start.

APPROACH

Built a reusable personal PyTorch framework with standardised data preparation, dataloader, training, and evaluation components, designed to plug into a broad range of model architectures with minimal setup.

OUTCOME

A framework that now underpins new ML research projects, cutting the time between an idea and a first working experiment.

PROBLEM

State-of-the-art brain tumor segmentation models are often too computationally demanding for deployment in resource-constrained settings such as Sub-Saharan Africa, creating a tension between accuracy and efficiency.

APPROACH

Integrated efficient transformer variants (Reformer, Linformer) and a recurrent LSTM into a 3D V-Net architecture, benchmarking them against a standard Transformer on the BraTS Sub-Saharan Africa 2024 dataset.

OUTCOME

The Linformer achieved the highest volumetric overlap (Dice score 0.7916), while the Reformer gave the most precise tumor boundaries (95th percentile Hausdorff Distance 18.68), showing efficient transformers can match or beat costlier architectures on clinically relevant metrics.

PROBLEM

Most generative work treats a single modality at a time. The open question here was whether a latent space learned from audio could be decoded into coherent visual output, a genuinely multi-modal generation task.

APPROACH

Encoded audio spectrograms into a shared latent space with a variational autoencoder, then decoded that latent representation into corresponding images, working through the time-series nature of the audio input.

OUTCOME

A working audio-to-image pipeline and a much better feel for latent space exploration and the failure modes of multi-modal VAEs.

PROBLEM

Backpropagation dominates modern deep learning, but genetic algorithms offer a gradient-free alternative for evolving network topology and weights. The goal was to see how far neuroevolution could go on a well-understood benchmark.

APPROACH

Built GENEVO, a neuroevolutionary framework that evolves fixed-topology networks through genetic crossover and mutation instead of backpropagation, then tuned population size, mutation rate, tournament size, and batch size for classification and, separately, applied it to variational autoencoders.

OUTCOME

Achieved competitive validation accuracy of up to 89.16% on MNIST classification, confirming neuroevolution as a viable, computationally efficient alternative to backpropagation, though VAE reconstruction quality remained limited.

PROBLEM

Transformers are the default architecture for translation, but RWKV was proposed as an RNN-based alternative offering transformer-level performance with better inference efficiency. That claim needed independent verification.

APPROACH

Conducted a replication study of RWKV on a translation task, training and evaluating it under conditions comparable to standard transformer baselines.

OUTCOME

Concrete evidence on how RWKV's generalisation and parameter efficiency actually compare to transformers in practice, not just on paper.

PROBLEM

Weapon theft trends across South Africa are hard to see in raw tabular crime data, since the interesting structure is relational, not row-by-row.

APPROACH

Modelled the data as a graph and applied unsupervised clustering, using centrality metrics and modularity optimisation to surface pattern structure that tabular analysis would miss.

OUTCOME

A graph clustering model that identifies weapon theft trends and hotspots across South Africa, demonstrating the value of relational modelling for crime analytics.