Agenda

Extracting key-fields from a variety of document types remains a challenging problem. Services such as AWS, Google Cloud and open-source alternatives provide text extraction to "digitize" images or pdfs to return phrases, words and characters. Processing these outputs is unscalable and error-prone as varied documents require different heuristics, rules or models and new types are uploaded daily. In addition, a performance ceiling exists as downstream models depend on good yet imperfect OCR algorithms.

We propose an end-to-end solution utilizing computer-vision based deep learning to automatically extract important text-fields from documents of various templates and sources. These produce state-of-the-art classification accuracy and generalizability through training on millions of images. We compare our in-house model accuracy, processing time and cost with 3rd party services and found favorable results to automatically extract important fields from documents.
Bill.com is working to build a paperless future. We process millions of documents a year ranging from invoices, contracts, receipts and a variety of others. Understanding those documents is critical to building intelligent products for our users.

At One Concern, we develop models for digital twins and their resilience by employing machine learning and advanced statistical methods to build a platform where organizations, communities, and private and public sectors understand, forecast, and mitigate climate risk. This session will cover how One Concern develops digital twins and resilience models by applying machine learning algorithms.

Asset prices are primarily affected by fundamental, technical, and market sentiment factors. The last factor refers to the public perception of the market and individual instruments traded within. While this perception is not solely based on information published in the media, online news sources and social media websites have a high impact on market sentiment and consequently, asset prices. A journalist with an investigation disclosing publicly traded company's unethical or illegal business practices will more-than-likely cause a change in market sentiment, which in-turn would result in a stock price change. Due to the impact of online media on the financial markets, sentiment-driven trading strategies have been and still remain an integral part of hedge fund portfolios. In this session, Vladyslav Ivanov, a Quantitative Researcher for Outremont Technologies digital assets manager, will cover the current state-of-the-art transformer-based deep learning models for sentiment analysis. He will discuss use of these models in the context of forecasting technology sector asset prices and will perform empirical assessment of models performance. The subjects of models application in latency-sensitive environments, as well as integration into the signal generation processes will also be addressed.

Long short term memory neural networks (LSTMs) originally rose to prominence in natural language processing, but have also shown value in time series. One drawback of LSTMs, and neural networks more generally, is that it can be difficult to derive effective confidence intervals. Two primary methods for calculating confidence intervals in LSTMs have been proposed: dropout-based simulation and estimating probability distributions.

In this case study we look at dropout-based simulation, which we found to be effective and more flexible than distributional estimation. However, added flexibility came at the cost of elevated computational burden.

In this session, Google Research Scientist Hamza Farooq will cover a data-driven, NLP-based framework for understanding Google customer feedback across any text-based data sources. He will cover a framework named Stardust that uses an ensemble of machine learning / deep learning models to parse and score topics to measure sentiment and context pertaining to trust and customer experience

Historically, edge machines have been viewed as dumb devices with the simple expectation that they perform tasks on command. With the recent advancements in deep learning, can we change this norm? Yes — with adaptive intelligence but the definition of this intelligence varies by form factor and associated experiences.

This session gives an inside view of inferencing on edge devices at Qualcomm. We discuss two classes of use cases: high performance and ultra low power. Our high performance use cases span multiple verticals that leverage computer vision applications (streaming, gaming and auto). Our ultra low power use cases are Always on Vision, Always on Audio and sensing in the Mobile and IoT markets.

Building an ML solution leveraging deep learning is not just about building a core model. A solution may entail building and maintaining multiple models. In addition, several enabling technologies and tools are also required. Can an organization build each of these technologies in-house? How does this impact overall costs and time to market? With several open source and commercial off-the-shelf offerings becoming available, when should an organization build these technologies and when should it buy them? In this session, we will discuss how to make these build vs buy decisions, including benefits, decision criteria and critical success factors.

In this informal but informative presentation, Dr. James McCaffrey from Microsoft Research will explain advanced neural architectures including variational autoencoders, generative adversarial networks, Siamese networks, and Transformer architecture encoders. You’ll see running demos of each type of network. No math equations with Greek letters. Well, maybe one or two. You’ll leave this presentation with a solid understanding of what type of problem each network can (and cannot) solve, and have the all the knowledge you need to communicate with subject matter experts. No experience with neural networks is necessary to understand and benefit from this presentation.

Pull requests (PRs) have become  de facto standard for code review/merge process by development teams using SCM tools like Github, Bitbucket. PR is a rich source of information about developers & reviewers. PRs can give us quite a lot of insights about the coding styles, logical skills of the developers  as every single line of code is being reviewed and bad smells are getting highlighted by the reviewer. The comments/suggestions that reviewer gives helps in understanding the proficiency of the reviewer. We have developed a set of PR Analytics by applying Transformers based NLP, Decision Trees & Statistical Analysis on PR data.

PR Analytics can  be used to perform skill assessment in order to find out the areas of improvement for the development team in a quantitative manner. PR Analytics can also help the Scrum masters & the project managers to better plan their deliverables since now they know strengths & weaknesses of the development team and can allocate the right developers for the right type of tasks
In this talk I would present some of the analytics that we have developed using data from bitbucket and how we are using them for improving the efficiency of our development teams.

Offering well-designed rewards is essential for increasing the engagement of the users. Boosting user activity in the desired products helps economies of scale and increases the efficiency of the platform. However, rewards incur significant cost. Therefore, designing an efficient learning algorithm to optimize marketing campaigns can significantly increase the profitability of these platforms.

In this talk, we consider the following problem: given a set of different promotion structures, assign each user to a promotion structure from the set, or nothing at all. The goal is to maximize user retention while respecting the budget/keeping the cost low. I propose a novel methodology to maximize the treatment effect in a budget constraint setting. Furthermore, we use Boltzmann-exploration to balance exploration and exploitation. This enables us to efficiently collect data and update the model regularly. Finally, I show that our approach outperforms the other alternatives including R-linear and generalized random forest.