What a whirl! As we wrap-up Knowledge Connexions and look forward to the holidays, we’d like to say a big thank you to all of our speakers, sponsors, attendees, and of course, our partners at Connected Data London. #KnowCon2020 was a great event, and we couldn’t have done it without you.
We are very pleased with the caliber of presentations, the wide range of topics presented, and the new conversations that were sparked. What stood out to you? What did you like, and what could we improve? Please let us know by filling out our attendee survey.
And yes, the recordings are now live! You can find them in your personal Knowledge Connexions account. Just enter your details in the pop-up form and watch away. If you’d like to purchase additional access to any of the masterclasses or workshops, we have a few ticket options: a pass for access to the entire library at 100 GBP, or individual passes to any workshop, masterclass, or the entire presentation day, each for 30 GBP.
SHACL is relatively new, having become a W3C recommendation in July 2017. Veronika showed a much more clear and concise introduction to using SHACL than I’d ever seen before, including examples in a public GitHub repo and also the super helpful online validator SHACL Playground.
The idea of using shapes constraints is to apply a “closed world assumption,” then test your graph data to see whether it conforms to an expected schema, and if not, show which elements violate the assumptions. This is useful in auditing/compliance, data quality checks, “unit testing” of knowledge graph applications, means for implementing human-in-the-loop feedback about annotations at scale, etc.
In the session talk, Veronika showed a case study of her projects applying SHACL for the Norwegian Maritime Authority to audit requirements among applicants for the Master Mariner certification. The statutes for this are quite complex, and of course the client understands little about advanced math topics like algebraic topology or predicate logic. Fortunately, SHACL provides a way for those of us working on knowledge graphs to specify clear rules that domain experts can also read and understand. First thing after the conference ended, I leveraged pySHACL to validate graph data for a current project using Veronika's slides as the best reference, and her work has enabled this as a standard part of our team’s KG practices. Highly recommended.
Dan McCreary is a Distinguished Engineer in AI and Knowledge Graphs at Optum Healthcare. Dan gave one of my favorite session talks, “Graph stories: How stories and metaphors can help you promote Enterprise Knowledge Graphs”. Consider this the high-level tour through the key points of enterprise knowledge graph adoption, circa late 2020, along with excellent pointers for deep dives into the pragmatics and emerging trends as well. Dan helped build one of the world’s large healthcare knowledge graphs, and he’s ideally situated to talk about the realities of enterprise data management, machine learning applications, hardware innovations, and so on. In particular, Dan’s talk focused on two themes.
One was using stories to help people who aren’t from technical backgrounds understand the value of "highly connected datasets" in finding solutions for enterprise problems. What is storytelling in this context? How does it help illustrate the important distinctions between graph-based approaches and what can (or can’t) be done in more “traditional” relational database management systems? The stories he shared were elegant and effective, like timeless haiku. Highly recommended if you need to convey these ideas within your team.
The other theme was about hardware, how we’ve had to adapt to generations of CPUs that were never quite optimized for graph data and graph algorithms – although that appears to be changing rapidly. To wit, if you’re working in computer vision and training deep learning models on image data, that is relatively “dense” data and well suited for current hardware design trade-offs. Conversely, if you’re working on healthcare systems, that data is relatively “sparse” due to its connected nature. Optum has been working with field-programmable gate array (FPGA) accelerators in areas such as graph embedding, and Intel and other vendors have new chip lines coming to address needs for graph-based processing at scale. By all means, check out Dan’s steady stream of excellent articles on Medium at https://dmccreary.medium.com/ – for example “Rules for Knowledge Graphs Rules” and “Understanding Graph Embeddings” are two recent gems.
There’s a general category of leading-edge work in natural language called question answering, which has much interest: can you take a large collection of data (e.g., documents) and begin to ask questions that an AI system can answer using that evidence? The idea of reinforcement learning is to use agents to train policies: instead of learning to predict from patterns in the data, RL agents learn how to navigate an environment and make relatively long-term decisions that pay rewards over time. In the Multi-Hop project, they use RL agents to navigate through a knowledge graph (as their environment) and try to identify where potentially “missing” links in the data can be inferred to help answer questions. The implications for this general area of work lead toward more robust, resilient AI solutions, capable of working with “real world data” in all its inherent messiness. Salesforce has had several intriguing RL research projects recently – for example, check out AI Economist. I look forward to seeing Victoria’s research moving into production applications.
Nariman Ammar is a cross-disciplinary computer scientist and Research Fellow at the University of Tennessee Health Science Center. Nariman’s presentation (“Leveraging the Semantics of Adverse Childhood Experiences for Explainable AI Recommendations”) laid out a perfect case for what knowledge graphs can do, particularly as one part of a broader AI system. Coming from a background in sustainability, I was blown away to see such a complex social dynamic – adverse childhood experiences – represented in an ontology. That definitely sparked an “aha!” moment for my own ideas. Now that the recordings are available, I look forward to poring over her presentation to better understand the architecture of the system.
Gary Marcus is a well-known cognitive scientist with strong opinions about what it will take to develop trustable AI--which he explores in the startup he cofounded, Robust.ai. What I found most interesting in Gary’s talk (“Rebooting AI: Adding Knowledge to Deep Learning”) was his emphasis on including the “right-brained” fields of social science in AI research, like linguistics. Just as silos form a barrier to organized data, academic silos bar the path towards AI general intelligence. How can these voices mix together? I think that’s where communities like ours come in: creating opportunities for cross-pollinating ideas and discussion to build more robust innovation.
Maureen Teyssier is the Chief Data Scientist at Reonomy. The company supplies "property intelligence" to the commercial real estate industry using data analysis and machine learning. I highly recommend Maueen’s thorough, succinct and accessible overview of the knowledge graph development process (geared particularly towards a production environment). Her presentation (“Unprecedented Products in Commercial Real Estate with Automated Generation of Large Knowledge Graphs”) laid out a pathway to building a product uniquely capable of reorganizing the way market information is collected and distributed. If you’re interested in building a knowledge-graph fueled startup, this is a great place to start.
“What were the most positive parts of your
experience at the event?”
“Meeting the community and seeing how active it is.”
“The masterclass was great– very practical and insightful. The workshop format helped me meet attendees. The presentations were also highly insightful. Having people who are successfully leveraging knowledge graphs in their products and day-to-day work really helps bridge the gap between theory and practice.”
“Learning new concepts, approaches, and semantic products along with gathering with other semantic knowledge colleagues.”
“It was an excellent platform for a deep dive.”
“Demonstrations of the technologies in application.”