The Connected Data Podcast Connected Data World

Graph-based technologies became first-class citizens in various industries and many practical applications. Still, building performant and reliable machine learning pipelines over graph data, e.g., graph machine learning applications and products, remains a non-trivial task.


This panel discussion brings together academic and industrial experts from fields where Graph ML yields significant gains and greatly improves traditional processes. In addition to highlighting successful business cases, the panel concentrates on questions often dismissed or hidden behind the curtains of modern Graph ML applications.


In particular, we will talk about the origins of graph data, its modeling, organization, and processing aspects; best communication interfaces; bridging a gap between products and ML algorithms as well as measuring their practical impact.


On a higher level, the panel will discuss upcoming trends in industrial Graph ML and prospective disruptive applications.


Key Topics


Graph Machine LearningDeep LearningGraph Data ManagementKnowledge GraphsGraph ML in Production

Target Audience


Machine Learning PractitionersData ScientistsData ModelersCxOsInvestors

Goals


Explore the interplay between machine learning and knowledge based technologiesHow to get the “actionable” knowledge from the graph data?How can those approaches complement one another, and what would that unlock?What is the current state of the art, how and where is it used in the wild?NLPBiomedIndustry (say, Ad Tech)What are the next milestones / roadblocks?Where are the opportunities for investment?We know drug discovery is on its highs, NLP is being democratized rapidly, what else?

Session outline:


IntroductionMeet and GreetSetting the stageKnowledge Graphs, meet Graph Machine LearningIt’s all about the data:How do you create, maintain, and process graphs?Databases or tabular sources?Do you consider data modeling aspects?Best communication interface: natural language or structured query languages?How can machine learning help create and populate graphs (including KGs)?Cover some of the current state of the artAn edge - does it appear naturally or derived from node similarities?What kind of problems can we solve by using it?NLP, Biomed, Industry (say, adtech)How academic datasets align with real-world tasksWhere is this used in production?Success stories and business casesWhat are the major roadblocks / goals, how could we address them, and what would that enable?How to bridge the gap between business goals and Graph ML models?How do we measure the impact of applying ML models in real-world tasks? Metrics, A/B testing, generally about setting things in productionWho are some key players to keep an eye on?Both from industry and research

Panelists:


Mikhail Galkin. Researcher, Mila | McGill University


Dr. Tiffany Callahan. Researcher, University of Colorado, Anschutz Medical Campus


Andreea Deac. Researcher, Mila | Université de Montréal


Dr. Charles Hoyt. Researcher, Harvard Medical School, Laboratory of Systems Pharmacology


Sergei Ivanov. Research Scientist, Criteo AI Lab


---


Subscribe to our YouTube channel for more gems from the vault:


https://www.youtube.com/@ConnectedDataWorld

After the amazing breakthroughs of machine learning (deep learning or otherwise) in the past decade, the shortcomings of machine learning are also becoming increasingly clear: unexplainable results, data hunger and limited generalisability are all becoming bottlenecks.


In this talk we will look at how the combination with symbolic AI (in the form of very large knowledge graphs) can give us a way forward, towards machine learning systems that can explain their results, that need less data, and that generalise better outside their training set.


--


Frank van Harmelen leads the Knowledge Representation & Reasoning group in the CS Department of the VU University Amsterdam. He is also Principal investigator of the Hybrid Intelligence Centre, a 20Μ€, 10 year collaboration between researchers at 6 Dutch universities into AI that collaborates with people instead of replacing them.


--


Slides available at: https://www.slideshare.net/slideshow/systems-that-learn-and-reason-frank-van-harmelen/267008886


---


Subscribe to our YouTube channel for more gems from the vault:


https://www.youtube.com/@ConnectedDataWorld

What do graphs have to do with novel hardware architectures for AI workloads?


Graph processing is the key to unlocking new architectures, as much as new architectures can boost execution of graph-oriented workloads.


As machine learning-powered applications are proliferating, the workloads that are created in order to serve their requirements are taking up an ever increasing piece of the compute pie.


An IDC study found that Data Management, Application Development & Testing, and Data Analytics workloads represented more than half of all IaaS and PaaS spending already in 2018. IDC notes that this was driven in part by initial adoption of artificial intelligence and machine learning capabilities.


The rise of generative AI means that as adoption grows, data and AI workloads will dominate. This is why we see NVIDIA earnings skyrocket, as well as a renaissance of novel hardware architectures designed from the ground up to serve the needs of data and AI workloads.


More specifically for data analytics, understanding relationships among data points is a challenging but essential capability. Graph analytics has emerged as an approach by which analysts can efficiently examine the structure of the large networks and draw conclusions from the observed patterns. This is why DARPA set out to develop a graph analytics processor with the HIVE Project.


Furthermore, all machine learning models are best expressed as graphs. This is how machine learning libraries such as TensorFlow work. Efficient processing of graph-based networks involves large sparse data structures that consist of mostly zero values, and next generation architectures should avoid unnecessary processing.


This panel explores the interrelationship between graph processing and novel AI hardware architectures. Hosted by ZDNet's Tiernan Ray with panelists from some of the most groundbreaking AI hardware companies: Blaize, Determined AI / HPE, Graphcore, and SambaNova.


---


Tiernan Ray. Contributing Writer, ZDNet


Tiernan Ray has been covering technology & business for 27 years. He was most recently technology editor for Barron's where he wrote daily market coverage for the Tech Trader blog and wrote the weekly print column of that name. He has also worked for Bloomberg, SmartMoney, and for the prestigious ComputerLetter newsletter covering venture capital investments in tech


Val G. Cook. Chief Software Architect, Blaize


Val G. Cook is Chief Software Architect at Blaize. An AI visionary and authority on the design of graphics and visual computing architectures, Val possesses two decades of experience in graphics and multimedia algorithms and software architecture. He is responsible for the Blaize Graph Streaming Processor software programming environment.


Carlo Luschi. Director of Research, Graphcore


Carlo is responsible for the study and development of algorithms for machine intelligence. Prior to Graphcore, Carlo was a Member of Technical Staff at Bell Labs Research, Lucent Technologies, and more recently Director of Algorithms and Standards at Icera Inc., which was acquired by NVIDIA in 2011.


Raghu Prabhakar. Software Engineer, SambaNova


Raghu Prabhakar is a senior principal engineer and one of the founding engineers at AI innovation platform SambaNova Systems. His research interests are in the areas of programming models, compilers, and hardware architecture for reconfigurable dataflow architectures.


Evan Sparks. Founder, Determined AI, an HPE Company


Evan Sparks, Vice President of Artificial Intelligence and High Performance Computing at HPE, co-founded Determined AI (now an HPE company). His group helps businesses get better AI-powered solutions to market faster and delivers the open source Determined Training Platform for large scale AI model development.

While mathematicians have used graph theory since the 18th century to solve problems, the software patterns for graph data are new to most developers. To enable "mass adoption" of graph technology, we need to establish the right abstractions, access APIs, and data models.


RDF triples, while of paramount importance in establishing RDF graph semantics, are a low-level abstraction, much like using assembly language. For practical and productive “graph programming” we need something different.


Similarly, existing declarative graph query languages (such as SPARQL and Cypher) are not always the best way to access graph data, and sometimes you need a simpler interface (e.g., GraphQL), or even a different approach altogether (e.g., imperative traversals such as with Gremlin).


--


Ora Lassila is a Principal Graph Technologist in the Amazon Neptune graph database group. He has a long experience with graphs, graph databases, ontologies, and knowledge representation. He was a co-author of the original RDF specification as well as a co-author of the seminal article on the Semantic Web.


--


Presentation slides available at https://www.slideshare.net/slideshows/graph-abstractions-matter-by-ora-lassila/266140641


---


Subscribe to our YouTube channel for more gems from the vault:


https://www.youtube.com/@ConnectedDataWorld

Taxonomies are the duct tape of connected data. They seem simple, flexible, and familiar. They are widely used. And they seem to work across many use cases and many domains. 


But when looked at in more detail, taxonomies turn out to be crude tools for knowledge organization that are very difficult to create, to scale, to adapt, to align, and to build on.


They don't work well for larger or more complex domains and use cases. Experienced talent and flexible tools for creating them are hard to find and to develop. Often taxonomies are built then abandoned for other, more robust approaches to knowledge organization.


It is essential to re-evaluate your connected data strategies in the context of alternative approaches to knowledge organization.


------


Mike Dillinger. Technical Lead for Taxonomies and Ontologies, AI Division, LinkedIn


Mike Dillinger, PhD, focuses on teaching machine learning algorithms about the world of work at LinkedIn. Before that, he was Technical Lead for LinkedIn’s and eBay’s first machine translation systems, and an independent consultant specialized in deploying translation technologies for Fortune 500 companies.


---


Subscribe to our YouTube channel for more gems from the vault:


https://www.youtube.com/@ConnectedDataWorld

What is Connected Data, and how is it interesting from a market point of view?


Knowledge Graphs have reached peak Gartner hype. Graph data science and graph AI are the fastest growing areas in AI. Graph databases are the fastest growing category in enterprise software.


Add to this the historical foundations of graph algorithms and analytics and semantic technology, which have been invigorated and are seeing widespread adoption, and you get the burgeoning Connected Data landscape.


While there is ongoing technical innovation happening in the domain, how does this translate to market value and opportunities for investment?


How is this market defined, and what is driving its growth?


Join us as we define and explore this landscape, discuss technology and use cases, challenges and opportunities for growth and investment, and where the future may take us.


Join George Anadiotis, Panos Papadopoulos, Bob van Luijt and Konstantin Vinogradov from our Connected Data World 2021 panel discussion as they address the following:


Key Topics


Defining the Connected Data technology and market landscapeExploring the Connected Data marketProviding an outlook for the future

Target Audience


EntrepreneursTechnical people with entrepreneurial spiritCxOsDecision makersInvestors

Goals


Define and explore the Connected Data landscape for people who are interested in it from a market perspectiveAnswer questions that matterHow is this market defined?What are some key drivers for growth?Where are the opportunities for investment?What is the outlook for the future?

---


Subscribe to our YouTube channel for more gems from the vault:


https://www.youtube.com/@ConnectedDataWorld