Hi, my name is Ben Wortman
I'm a Data Scientist!

Adversarial patches have been of interest to researchers in recent years due to their easy implementation in real world attacks. In this paper I expand upon previous research by demonstrating a new hidden patch attack on optical flow. By altering the transparency during training I generate patches that are invariant to their background meaning they can be inconspicuously applied using a transparent film to any number of objects.

An end-to-end model for the simultaneous detection and inpainting of images obstructed by fences. This was achieved by overlaying raw images with existing fence masks and training to minimize SSIM, L2, and Adversarial loss. Despite training exclusively on synthetic data, this method was able to generalize to unseen, real world data and effectively inpaint ~90% of the test images.

During my summer as an intern at Carnegie Mellon SEI, I prototyped an interface for autonomous drone controllers. This included training a DL bayesian object detector, modifying a DL depth estimator with MC dropout to give uncertainty estimates on distance, and finally object localization using the camera's intrinsic projection matrix and GPS coordinates to give users a birds eye view of detected objects in the field.

Through the use of NLP word embeddings from Facebook’s FastText model, I was able to produce an updated emotion model which provides a 5% increase in performance using only 46% of the labels. When it comes to annotation, this increases the amount of information each label has while nearly cutting in half the costs for an equivalent dataset.

Since emotion labels are not completely independent, I am using semantic embeddings as targets to encode additional information into the model's training. In addition to this I have trained a complementary model using transfer learning from a pretrained Resnet to make predictions on LMA components which have been shown to be correlated with emotion. I have also crowdsourced an additional dataset annotated for human interaction to be processed in a separate channel.

While working as an intern at Impact Radius, I matching walled garden Ad impressions to customer conversion data in order to give customers insight into how their ad campaigns were performing. To do this, I created features from browser metadata and census demographic data local to the conversion IP before using a fuzzed decision tree algorithm for classifcation. In addition to this, I also developed a preprocessing step that identified 30% fraudulent TV promo code linked conversions. When filtered out using an SVM, these dramatically improved predictive performance and the explainability of the model (shapely values).

A fun personal project. I built a streamlit chat bot powered by gpt-3.5 and that provides custom tarot card readings.

To get some practice in interfaces I programmed a simple collage builder application for my younger brother's etsy page.

For fun I built a webcrawler to gather ~10k images from online and then trained a GAN to switch the seasons in the images.

One of my first projects. I explored interface design using Conway's game of life. I include tunable parameters so users can explore how changing paramters affects the simulation.

During my free time I love to make music. I play guitar, piano, and drums. On weekends you can usually catch me playing out with my band.