RE•WORK Deep Learning Summit

Summary, in this talk, Dr. Edgar Kalns who is the lead researcher on the SRI - Squirrel AI Joint Lab will discuss two exciting AI and Machine Learning applications for adaptive education. He will illustrate the theoretical foundation, technical approach, and research findings from the MILBA and VPA projects.

In this talk we explore real world use case applications for automated “Fake News” evaluation using contemporary deep learning article vectorization and tagging. We begin with the use case and an evaluation of the appropriate context applications for various deep learning applications in fake news evaluation. Technical material will review several methodologies for article vectorization with classification pipelines, ranging from traditional to advanced deep architecture techniques. We close with a discussion on troubleshooting and performance optimization when consolidating and evaluating these various techniques on active data sets.

Searching for homes is the primary mechanism guests use to find the place they want to book at Airbnb. The goal of search ranking is to find guests the best possible options while rewarding the most deserving hosts. Ranking at Airbnb is a quest to understand the needs of the guests and the quality of the hosts to strike the best match possible. Applying machine learning to this challenge is one of the biggest success stories at Airbnb. Much of the initial gains were driven by a gradient boosted decision tree model. The gains, however, plateaued over time. This talk discusses the work done in applying neural networks in an attempt to break out of that plateau. The talk focuses on the elements we found useful in applying neural networks to a real life product. To other teams embarking on similar journeys, we hope this account of our struggles and triumphs will provide some useful pointers. Bon voyage!

Goal-oriented conversational systems are typically multi-turn, relying on the entire conversation thus far to generate a response to user input. Many of these systems use dialog state tracking or belief tracking, to either rank candidate responses from a pool of templates or generate responses directly while others are end-to-end. End-to-end models for goal-oriented conversational systems have become an increasingly active area of research.
In this talk, I will present our recent efforts to build end-to-end conversational models for customer service application. We use historical chat transcripts and customer profile data to build models, and test with live customers using a human-in-the-loop research platform. We experiment with sequence-to-sequence model that generates responses word by word, and multi-encoder based ranking model to score template responses. I will compare these approaches as they apply to customer service domain.

The combination of breakthroughs in AI and Machine Learning and increasing amounts of digitized medical data have generated significant excitement about the potential to automate medical decision making processes. Among these, there are significant opportunities in designing solutions for settings where AI and ML systems can work seamlessly with human experts to provide more efficient and accurate patient care. In this talk, I outline one such problem, that of medical expert disagreement. We study the application of machine learning to predict patient cases which are likely to give rise to maximal expert disagreement. We show one can develop and train AI models to predict an uncertainty score for a patient, identifying cases where large disagreements ensue, and flagging that patient for a medical second opinion. Methodologically, we formalize the importance of doing direct prediction of these uncertainty scores, instead of a two step process of diagnosis and postprocessing, evaluating on a  gold-standard adjudicated dataset.

Over 5 million Americans are affected by Alzheimer's disease and have cost over 200 billion dollars in direct and indirect costs . Early and accurate diagnosis of Alzheimer's disease is important because it opens the possibility of therapeutic intervention to slow or halt the disease progression. Unfortunately, Alzheimer disease (AD) remains a diagnosis based on clinical grounds and most diagnosis gets established at a late stage when too many neurons have been lost. Much research effort has been made on biochemical and imaging tests
to improve our early diagnostic capability but most have met with mild to moderate accuracy. 18F-FDG PET scans of the brain utilize radioactive glucose to image the energy uptake pattern in various parts of the brain, which has recently been implicated to change in subtle ways in Alzheimer's disease. In this study, we develop and validate a deep learning algorithm that predicts the final diagnosis of Alzheimer disease (AD), mild cognitive impairment, or neither, approximately 6 years before the final diagnosis from these PET scans.

Prospective 18F-FDG PET brain images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) (2109 imaging studies from 2005 to 2017, 1002 patients) and retrospective independent test set (40 imaging studies from 2006 to 2016, 40 patients) were collected. Final clinical diagnosis at follow-up was recorded. Convolutional neural network of Inception architecture was trained on 90% of ADNI data set and tested on the remaining 10%, as well as the independent test set, with performance compared to radiologic readers. Model was analyzed with sensitivity, specificity, receiver operating characteristic (ROC), saliency map, and t-distributed stochastic neighbor embedding. The algorithm achieved area under the ROC curve of 0.98 (95% confidence interval: 0.94, 1.00) when evaluated on predicting the final clinical diagnosis of AD in the independent test set (82% specificity at 100% sensitivity), an average of 75.8 months prior to the final diagnosis, which in ROC space outperformed reader performance (57% [four of seven] sensitivity, 91% [30 of 33] specificity; P < .05). Saliency map demonstrated attention to known areas of interest but with focus on the entire brain.

By using fluorine 18 fluorodeoxyglucose PET of the brain, a deep learning algorithm developed for early prediction of Alzheimer disease achieved 82% specificity at 100% sensitivity, an average of 75.8 months prior to the final diagnosis.