David Bick Email & Phone Number

Silicon Valley, California, United States

Pittsburgh, Pennsylvania, United States

- Published in InterSpeech 2022 and 2 papers in ICASSP 2023, on improving speech enhancement through fine-grained speech characteristics - Prepared presentation summarizing the three papers to present to Meta Reality Labs Research Audio- Applied CNN to learn phonetically significant acoustic features- Used the pre-trained acoustic model to apply differentiable loss that enforces de-noised speech to retain fine-grained acoustic features from clean speech

San Francisco Bay Area

- Implemented prompt-tuning for T5 to guide frozen pre-trained language models to perform new tasks using a small amount of learnable parameters prepended to input - Trained T5 to convergence on SQuAD with model-tuning and prompt-tuning - Accounted for idiosyncrasies of compilation stack lowering to Cerebras hardware when evaluating code- Debugged reference T5 implementation by comparing outputs to HuggingFace- Generalized SQuAD data-processing scripts to support more models and tokenizers

Washington, District Of Columbia, United States

- Fine-tuned a BERT pre-trained masked language model on documents specific to Fannie Mae to gain understanding of financial and company terminology - Investigated multiple clustering pipelines, including PCA + K-Means and UMAP + HDBSCAN, and then used class-based TF-IDF to identify idiosyncratic words within clusters - Analyzed error cases and conducted further literature review, identifying relevant paper to address the issues (from EMNLP 2020)- Incorporated method from paper of normalizing flows to induce a smooth and isotropic embedding space that improved clustering results, qualitatively confirmed by manager - Tested sentiment analysis pipeline on top of the language model to provide scalable analysis of relationship between FNMA and client banks - Performed regular maintenance work and feature addition to team's primary report, a Flask app deployed on AWS EBS- Transitioned R-Shiny production process to AWS from on-prem Netezza

Greater Pittsburgh Area

- Mentored by Tom Mitchell, investigated ability to predict a stimulus word shown to a subject based on the MEG brain scan images of subject after they were shown the word- Implemented CNN's to reach accuracy of almost 80% on the stimulus word, out of a set of 60 words - Visualized and interpreted activations to identify which areas of the MEG scan contributed most to the prediction of a word- Utilized Gaussian Naive Bayes and Logistic Regression to perform experiments of feature extraction prior to inputting data to NN's

Washington D.C. Metro Area

- Developed a model converting natural language to SQL, incorporating pre-trained BERT for word embeddings, LSTM feature extractor, and three prediction heads to predict select column, aggregator operation, and where clauses- Acheived 86% accuracy on an out of sample set of Fannie Mae specific questions, while working with a very limited training dataset - Presented on end product and concepts behind it four separate times, including to VP of Analytics, VP of Credit Risk Analytics, and 5/8 team leaders within Analytics division

Greater Pittsburgh Area

- Explained assignments with topics of MLP for phoneme classification, CNN for phoneme classification, LSTM with CTC Loss for phoneme classification, and LSTM encoder-decoder with attention for ASR - Advised final projects involving GAN's for image generation, text-to-speech, and image segmentation- Answered over 1,200 student questions on Piazza (Q&A forum), most of all TAs, helping with theoretical and applied questions and debugging- Gave recitations to over over 200 Masters/PhD students on using AWS EC2 instances for training neural nets, recurrent neural net implementation in PyTorch, and theory behind attentional seq2seq models- Debugged student's neural network homework's for 2-4 hours every week during office hours

Washington D.C. Metro Area

- Analyzed epidemiological datasets of interest to the Army, mostly in relation to vectors of disease in areas that affect troops- Learned and applied tidyverse packages for visualizing and analyzing data in R, and created a small NN in PyTorch to predict wind speeds - Utilized circular statistics as part of analysis of wind direction in Sahara and effect on mosquito travel

Masters, Language Technologies

Bachelor Of Science - Bs, Statistics And Machine Learning

Activities and Societies: Jazz Ensemble, Chair of Funding Committee Student Dormitory Council, Orientation Counselor, Music Lessons

Education record

Activities and Societies: Class Treasurer, Jazz Band, Jazz Combo