Explored the robustness and efficacy of auto-encoders, VAEs, and GANs as pretrained feature extractors within the context of the Differential Private Model Inversion framework, particularly when confronted with challenging and unstable training conditions.

Investigated the practical advantages of integrating a stereoscopic depth sensor into the Point-NeRF framework for 3D scene reconstruction. Designed and conducted real-world experiments to demonstrate qualitative improvements without sacrificing efficiency.

Proposed a novel ensemble approach employing a gaussian pyramid of resized inputs and an adversarially-trained denoiser to robustly defend against white-box adversarial attacks on CNNs.

Leveraged pretrained ResNets and GANs for the efficient training of super-resolution models, enabling the synthesis of high-fidelity outputs from low-quality Pokémon images.

Created interactive simulations to describe and communicate the mathematics behind importance and rejection sampling.

Implemented the Asynchronous Advantage Actor Critic paper in PyTorch, leveraging parallel processing to efficiently train an agent for playing Space Invaders.