A next generation insights engine

Machine Learning Engineer, applying computer vision techniques (specifically CNNs) to identify cancerous regions in mammograms

Motivo DataMachine Learning expert, design and implement algorithms to obtain insights from largedatasets.• Implemented KDB-tree, means of multidimensional indexing (essentially a multidi-mensional B+ tree), currently being used to efficiently search dimensional data fromsemiconductor space. Also implemented various graph-search algorithms, again cur-rently being used to work with data in semiconductor space, as well as several toolsfor visualizations.• Helped implement company core technology (extension of suffix tree) to provideefficient search in combinatorial spaces, and currently applying the technology tonatural language space.• Obtained research experience in Recursive Neural Networks (using Python Theanoimplementations), applied to sentiment analysis and automatic text generation.

Served as teaching assistant under Prof. Alex Biliris for course in Database Systems. Heldweekly office hours, explained fundamentals of relational databases, covered concepts in-cluding relational algebra, sql, indexing (B+ tree and hash tables), query optimization,supervised teams of students in developing real-world database application. Graded home-works/exams, held recitation sessions to reinforce concepts, had a great experience teachingstudents!

Phoneme Classification Project, Natural Language ProcessingProject completed under the supervision of PhD Avner May (who was under the supervi-sion of Prof, Michael Collins) in conjunction with IBM Watson.• Purpose was using some interesting kernel-tricks, we could replicate the best resultsin phoneme classification obtained from deep learning (multi-layer neural nets) usingSVMs ... my job was to try and get the best results using neural nets, for whichother team members would try to replicate with other methods.• Using large multi-layer neural networks, attempted to classify phonemes in Can-tonese and Bengali. Implemented architecture in matlab (and code was GPU com-patible). Altered hyper-parameters to minimize perplexity on the validation set, in-cluding activation function (used sigmoid, tanh, relu, etc.), number of layers/degreeper layer, tried recursive pretraining, etc.• Obtained results comparable to prior results in literature, i.e. for cantonese, a per-plexity of 6.7 on the heldout (test) set. Received an A+ for class credit for job welldone!

R¨aestch Lab, Memorial Sloan Kettering Cancer CenterPredicted plan sections in the initial consultation notes of prospective cancer patientsbased on patient background information.• Applied techniques of SVM, Logistic Regression and k-NN to predict word space ofthe plan section. Nearest neighbor was easily the most successful in indicating wordpresence, and was a noticeable improvement over predicted no plan (SVM and LRwere also improvements, but far less). Accuracy was determined by comparing theword presence of the ML predicted plan to the actual physician’s plan.• Made improvements to the NLP aspect of this problem by extending code to parseExcel files.

Produced business to business recommendation systems for a large telecom company.• With a set of 180,000 success (current customers) and 70,000 failures, predicted howlikely future businesses were to purchase from the telecomm company (likelihoodmodel), and returned specific product recommendations for each business (productrecommendation model)• Implemented a nearest neighbor model to predict likelihood a business will purchasetelecomm from our client; predicted 84% of successes and 76% of failures based ona small testing set (20% of the training set).• Constructed a product recommendation system using the current customer databaseto predict what telecomm products (out of 23 total products) businesses would pur-chase from our client by using nearest neighbor, clustering, and regression algorithms.• Both models, written in Java, accommodate very large data sets (minimal overheadwith structures, parallelized, etc.); models will dynamically update when more datais inputted; currently, due to data limitations the likelihood model is biased towardsuccess data (lacks failures), but in time when more data is recorded, predictionaccuracy will increase and likelihood model will become more generalizable. Also,product recommendations will become more specific as the clientele size increases.

Caltech Preceptorship ProgramI attended a program offered to Caltech students by Dr. William Caton, the chairman ofneurosurgery at the Huntington Memorial Hospital. I spent over 200 hours at the hospitaland was exposed to numerous medical fields, including neurology, emergency medicine,pediatrics, pathology, cardiology and radiology. I was able to witness firsthand approx-imately 50 surgeries as well as physicians making their rounds, and learned quite a lotabout medicine along with some of the limitations (and advancements) in medical tech-nologies and software. The most glaring software deficiency was in the Emergency Room,where doing something very simple (i.e. prescribing an IV) would require inconvenientlygoing through multiple windows. Perhaps in the future I will have an opportunity tomake improvements to this outdated software. Also interesting is the potential of usingcomputer science to make more accurate diagnosis. Applying machine learning to pastpatient data could help physicians verify/reconsider difficult medical decisions.

Performed research with Professor Tracey Ho regarding anonymous communication.• Made practical improvements to the Bus Protocol for anonymous communicationproposed by Beimel and Dolev in prior research.• Provided robustness against the problem of probabilistic overwriting in the originalmodel.• Quantified a tradeoff between entropy and anonymity so a user will be able to cus-tomize parameters in our protocol based on how much they are willing to relax theperfect anonymity requirement.• Resulted in publication (see publications) (Actually worked in this project over several years as a student researcher)