WikiNLS.com: Wikipedia natural language search web application• Used Pytorch pretrained BERT architecture as backbone. Fine-tuned on SQuADquestion-answering dataset. Deployed app with Flask/nginx backend at WikiNLS.com.Published multiple articles in Towards Data Science on state-of-the-art techniques for deep learning and natural language processing (https://medium.com/@dwbressler):• Speed up your deep learning language model up to 1000% with the adaptive softmax, Part 1. https://towardsdatascience.com/speed-up-your-deep-learning-language-model-up-to-1000-with-the-adaptive-softmax-part-1-e7cc1f89fcc9• Speed up your deep learning language model up to 1000% with the adaptive softmax, Part 2: Pytorch implementation. https://towardsdatascience.com/speed-up-your-deep-learning-language-model-up-to-1000-with-the-adaptive-softmax-part-2-pytorch-d47fe9a56152• Building a convolutional neural network for natural language processing. https://towardsdatascience.com/how-to-build-a-gated-convolutional-neural-network-gcnn-for-natural-language-processing-nlp-5ba3ee730bfbPublished code on state-of-the-art techniques for deep learning and natural language processing:• Adaptive Softmax: https://github.com/DavidWBressler/adaptivesoftmax• Gated Convolutional Neural Networks (GCNN): https://github.com/DavidWBressler/GCNN

Built a classifier classifying skin lesions into >50 categories. Top-5 average sensitivity >75%, specificity >98%. Used Fastai library with Pytorch. Model based on pretrained 34-layer ResNet fine-tuned on user-submitted images. Model uses 1-cycle superconvergence for quicker training times, as well as discriminative layer training to assure optimal transfer learning. Model uses late fusion to incorporate metadata and textual inputs. Implemented label hierarchy with multiple layers of meaningful coarser labels, to assure each final label contains a sufficient number of examples. Implemented both oversampling and thresholding to address class imbalance. Final accuracy improved by a custom test-time augmentation (TTA) procedure and ensembling.

• Received US patent (20160005320) for developing Brainful, a novel internet-based pain-relief brain-training program, used by >5K patients to reduce pain without drugs.• Led clinical trial with >1K online subjects to show efficacy of Brainful platform for pain-relief. Led trial design, conducted the experiment, analyzed results.• Drafted Matlab code to automatically query MongoDB database, pull select clinical data in JSON format, analyze data, and produce summary graphs and results dashboard.• Designed and conducted extensive user-testing with >50 users in person and >5K users online to improve Brainful platform.• Created ‘Neurotype’ web application. User answers 50 web-based questions that are stored in SQL table, table is queried through PHP/mysql, and data is crunched to create and return html report.

• Led experiments using EEG/ERP to study visual spatial attention.• Helped develop new technique for EEG source localization.

• Led neuroscience projects resulting in five publications. Research focused on brain patterns of visual spatial attention, and effects of attention on perception. Research cited >450 times in scientific literature.• Created novel fMRI time-series analysis, enabling discovery of a new brain mechanism.Was first to describe attention-modulated suppression of infraslow noise in visual cortex.• Created new retinotopic mapping procedure using reverse correlation andnonparametric permutation significance testing, enabling mapping across visual cortexwhile simultaneously estimating eccentricity-dependent attention neural modulations.• Generated MATLAB scripts to control visual display of experimental stimuli, collecthuman responses and fMRI data, and analyze resulting 4-dimensional brain datasets.• Presented research through talks and posters at eight international scientific conferences.

• Led projects that resulted in seven publications (two first-author)• Managed three undergraduate students (who are co-authors on publications).• Increased brain analysis code efficiency by 2000%. Analysis ran in hours, not days.