We propose a simple yet effective alternative to pre-training and finetuning called Don't Pre-train, Teach (DPT). When compared to when they were pre-trained on ImageNet and finetuned, our small models achieve superior performance or lag behind by at most 1.5%, while only taking 5% of the training time.

We meta-optimize Deep Operator Nets to jointly learn the operator and an optimal set of sensor points for the function family. For the nested gradient, we leverage the Implicit Function Theorem and use a Neumann approximation of the inverse Hessian. Compared to the fixed sensor points, we achieved a 20% decrease in test error for Diffusion PDEs and a 72.42% decrease for Wave PDEs.

Wrote custom CUDA kernels to speed up the forward pass in convolutional neural networks. Leveraged tiling, shared memory, matrix unrolling, loop unrolling, reductions, and streams to achieve a 58.82% decrease in processing time.

We found that, for a user's first post, any kind of feedback from the community, even negative, led to users more likely posting again compared to if they received complete silence.

We predict what note someone is listening to from EEG signals. We first converted the time series signals into to a 2D image via an Azimuthal Equidistant Projection and Bezier curve interpolation. Then, we fed that into various logistic regression and neural network models, achieving an accuracy of 58.63%.

By leveraging phoneme-level information in the TIMIT dataset, we relax Convolutional Neural Networks' reliance on the Connectionst Temporal Classification (CTC) function, enabling the model to generate word-boundary timings with an average error of 67 ms, while also introducing interpretability via class activation maps.

A Karaoke web app that pulls lyrics and audio from the internet, obtains the word-timing information, and highlights the lyrics in real time for the singer to follow.

Implemented the original Style Transfer algorithm of Gatys et. al, to transfer artistic styles to images. I modified their approach to be more faithful to the desired content.

Implemented Seam Carving to decrease/increase the size of photos while preserving content. Additionally, modified method for content-aware object deletion.

Recreated Mario Kart with Rumbas. Features included smooth driving via Nintendo Switch controllers and pick-up powerups (nearest-neighbor attack, speedup, stop) via Decawave DWM1001C Tags.

Programmed a robot to find a tennis ball on a pendulum, predict the trajectory, and catch it, all from a single image.