Alvin Li Email & Phone Number

Bala Cynwyd, Pennsylvania, Us

Equity Options

Cambridge, Ma, Us

Published a paper to the AdvML Frontiers @ ICML2023Liquid time-constant networks (LTCs) have recently emerged as a promising method for processing complex continuous time-series data. To enhance the robustness of LTCs against adversarial attacks, we propose several novel adversarial training pipelines. These training pipelines introduce disruptive elements, such as data perturbations and temporal manipulation. We evaluate the performance of a standard LSTM network, a standard LTC network, and their adversarially trained counterparts on their performance across uniformly sampled, irregularly sampled, adversarially perturbed, and raw time series data. When evaluated on adversarially perturbed irregularly sampled time series data, our adversarial LTC model outperforms the prediction accuracy of state of the art standard LSTMs by over 35%, standard LTCs by over 20%, and adversarial LSTMs by 4%.Mentored by Alexander Amini, Mathias Lechner, Professor Daniela Rus (Director of MIT CSAIL)

Cambridge, Ma, Us

Completed the primary proof of concept for an enhanced ML image search querying process. We developed a novel deep feature variational autoencoder network with superior reconstructions compared to state of the art convolutional VAEs. Our model achieves this by incorporating a perceptual loss network (EfficientNet B0) that extracts differences in feature representations between ground truth and output images rather than evaluating them on a traditional pixel-by-pixel basis. We implemented a t-SNE algorithm to generate a visualization of the encoded data's latent space, then constructed a separate interface that displays a smooth linear interpolation of the data across each latent dimension. Our model allows users to construct a set of desired latent features, then reverse engineer the closest image in the dataset that matches their search query.Mentored by Alexander Amini, Igor Gilitschenski, Wilko Schwarting, Professor Daniela Rus (Director of MIT CSAIL)

Amsterdam, Nl

Seattle, Wa, Us

Launched a backend feature for Amazon Connect that automates agent rerouting, addressing a critical pain point and reducing downtime for over 1,000,000 user interactions

Hatch enables customers to seamlessly reserve a table and pre-order their food through a mobile app so that their dishes are served immediately upon arrival. Our team is working to revolutionize the restaurant industry through pioneering a novel dine in experience and optimizing operational efficiency for restaurants.

Mountain View, Ca, Us

HDR+, Portrait Mode, Nightsight for Pixel 4.