A Brief Insight Into SAP’s First ML Research Retreat

SAP ML Research Team

Last week, our machine learning (ML) research team held the first SAP ML Research Retreat to bring together a diverse group of about 40 ML experts ranging from the realm of academia to business. The list of attendees included members from different SAP ML units, as well as distinguished researchers coming from top-tier partner universities who presented their current ML projects. The main purpose of this one-day workshop at Innovation Center Postdam was to create an informal environment of knowledge exchange between SAP’s research partners and ML teams. The retreat also laid the foundation for the transfer of the latest ML technology into the plethora of enterprise machine learning use cases.

The keynote of the event was presented by Max Welling, a distinguished machine learning professor from the University of Amsterdam, who is well-known for his seminal work on variational auto-encoders. In his opening session he introduced new deep learning approaches developed in his lab, notably a novel framework for embedding raw graph data into the deep learning setting. Their approach allows many real-world problems to be cast into graph problems, which naturally would not fit into generic deep learning frameworks. In his words “the beauty of this new technology is the unifying view of a lot of data pulled together”. Specifically, the approach’s intrinsic incorporation of semantic information provides high application potential for several classification tasks or recommendation problems e.g. in the field of e-commerce.

Max Welling gave way to his PhD student Thomas Kipf to introduce their recent work on Graph Convolutional Matrix Completion. His talk touched upon the application of graph convolutional networks to link prediction and entity classification, as well as matrix completion for recommender systems.

Showcasing the various use cases of ML, Paul Squires of New York University (NYU), introduced a new ML approach for quantitative human capital management. The joint research project with NYU colleague Harold Kaufman explores career paths of knowledge workers in organizations, such as programmers, accountants or product developers, along with the role of machine learning in career path prediction. Arguing that human capital management poses challenges in terms of ambiguity and invalidity of job titles and descriptions, their research project questions the exact meaning of a career path. Finally, the findings of their research will be of value for the development of tools and methods that allow for improved career management, for both knowledge workers and organizational leaders.

With a research background in the medical domain, several scholars showed that their ML research projects offer multi-domain applicability and great value prospects for various other industries. The research partner from University of Pittsburgh/Carnegie Mellon University, Kayhan Batmanghelich, who has a strong ML background in the imaging context, notably imaging genetics, initiated the discussion of ML in the area of healthcare. He based the outline of his collaborative research project on the concept of multi-domain learning for image data and unstructured text. In this regard, he presented probabilistic models of medical images and genetic variation. He concluded with an extension of his work, outlining the integration of deep learning for tackling general multi-modal problems, such as text and image data while preserving interpretability.

Another ML approach with roots in the medical domain was introduced by Narges Ahmidi from Johns Hopkins University’s Malone Center for Engineering in Healthcare. Her presentation outlined a technology to analyze and optimize pathways. In the context of hospitals, these pathways can be considered as the sequence of interactions between receiving a patient until discharge. However, such a concept has applicability far beyond healthcare e.g. in the industry domain, where a pathway may reflect all manufacturing or delivery steps within a supply chain. Ahmidi argued that each person or machine in an organization is an influencing point in a complex system and underlined the need for a solution to identify current deficiencies that negatively influence the system’s functioning. Her research project specifically aims to develop generic automated techniques to understand the system’s cross-linked components and interactions. Moreover, pursuing the goal to uncover systemic bottlenecks, her team applies machine learning techniques, such as causal inference and mediation analysis, which eventually leads to optimized pathways.

Benjamin Gutierrez Becker and Abhijit Guha Roy, two research scientists of the Al-med Lab of Ludwig-Maximilians-University of Munich, began a lively discussion regarding their recent developments in medical imaging. Both have worked on creating alternative solutions for the implementation of ML for medical applications, where annotated data is scarce and complex. Gutierrez Becker outlined the application of a reinforcement training strategy based on the multi-armed bandit problem to improve the selection of medical training data. He was followed by his colleague Guha Roy who gave insight on into a fully convolutional neural networks for fast brain image segmentation.

Michael Rotman from Tel Aviv University,, a PhD student of Lior Wolf, introduced a more general problem with regards to training very deep fully connected neural networks and presented a novel training method. With the aim to speed-up training, he proposes an approach with dynamic parameter-specific learning rate.

The last academic presentation of the day was by Serhii Gavrylov, from Ivan Titov’s research lab at the University of Edinburgh. He shared his latest work on the emergence of language with multi-agent games, which just recently has been accepted at NIPS 2017. Gavrylov began his presentation explaining a setting where two agents engage in playing a referential game and, from scratch, develop a communication protocol necessary to succeed in said game. Within this setting, machines can be forced to learn a new artificial language (e.g. sequences of discrete symbols) that is efficient. He showcased his work in the context of learning a language used to communicate the contents of images between a sender and a receiver.

After a full day of presentations from our academic partners and vivid discussion sessions, Zbigniew Jerzak, head of the SAP Deep Learning Center of Excellence and Machine Learning Research, resumed the stage to conclude with a synopsis of SAP’s aims and use cases in the context of machine learning. He hereby emphasized on the importance of building and integrating strong ML components to enhance existing enterprise solutions. With this aim in mind, the retreat offered all participants the opportunity to expand their ML research network and learn about other distinct use cases machine learning has to offer. A big thanks to all the participants for making this possible!