Prosus AI Marketplace
On organizing a virtual event for our global Data Science community
As one of the largest technology investors in the world, Prosus has a diverse portfolio of companies spanning across many industries (classifieds, food delivery, fin-tech, etc.) in markets around the world. All these companies have data science teams building machine learning into their products. As the Prosus AI team, we have the privilege to work closely with all these teams on various machine learning initiatives. On many occasions, we see that these organizations address similar problems, struggle with the similar issues, and are interested in the same science. The Global Prosus AI Marketplace is the platform to connect hundreds of data scientists across the Prosus portfolio of companies. Given the COVID travel restrictions, we held our first virtual Prosus AI Marketplace.
We called the event a ‘marketplace’ as we aimed to recreate the vibrancy of a marketplace, where people could make spontaneous connections, showcase their work, share their expertise and learn from each-other.
The event took place from September 8 - 10, 2020. You can catch a short glimpse in our wrap-up video here.
In this post we share a few insights on what it takes to organise a global virtual data science event successfully.
1. Pick your Tech Stack wisely - especially for a virtual event
How to decide what tech stack we would use to host the virtual event? We needed a robust tool with functionalities for video streaming, breakouts, networking, chatting, polling, talent directory, event and knowledge management, etc. We tried out some new all-in-one virtual event platforms that sprung up with the onset of COVID-19. Though they were promising and had some advanced features to simulate the experiences of a physical event, in our tests, failed on fundamental things like video quality (especially for more than 50 people). Because of not knowing an established leader in that space and shortage of time to experiment with many such platforms, we decided to skip it in favour of tools that we are already familiar with.
Plenary: Live Streaming via Studio
We wanted to do something better and different than just hosting another meeting over video conference. So we created a real studio at Prosus HQ, from where the event was moderated and live streamed. This contributed positively to the experience of the participants as they were lead through the program by a dedicated event host and received high quality content. Of course, the studio was run only with a small team, and we took social distancing, masks and other safety precautions very seriously.
Interactivity: Zoom and Slack
Besides plenary sessions, we organized many interactive sessions as part of the marketplace. And plenary sessions (e.g. Keynotes) were followed by a Q&A from the audience at the end. For these sessions we decided to stick to our tried and tested solutions: Zoom for breakouts and Slack for interactive Q&A. The flow between and within the tools was orchestrated by a dedicated technical team, that managed all transitions from live streaming to breakout sessions and interactions.
Event Portal: A custom Event Website as one-stop solution
In our effort to make it simpler for participants to find all the information and links for various sessions, we built a website for the event to have one place to access all information related to the event. Everything could be accessed from this dedicated site, from streaming, to zoom sessions, to slack channels to polls. To enable real-time interaction with participants while they are watching a live session, we embedded slido polls on the same screen. This feature was useful for participants, as they did not need to switch tabs or do any sort of multi-tasking on a second screen or phone.
2. Content is king: high quality speakers and a diverse format
Data scientists are in general, content hungry and are known to be an extremely hard audience to please and we like it that way. We designed the event in a way that different sessions addressed various goals of the event - hone data science skills, connect and learn from each other, learn from experts in the field, and showcase use cases and applications.
- Kaggle competition to stress the importance of bias checkers in ML models. We used data from DonorsChoose.org to predict which school project will be funded. Models were ranked not only on accuracy, but also on bias and fairness of predictions. We formed cross-company teams to encourage interaction even before the event. The top 3 teams won tickets to their favourite ML conference.
- Panel Discussion on ‘AI applications in Food Delivery’, with panelists as data science leaders from Swiggy, Delivery Hero, iFood and Prosus. The panel addressed the common challenges and solutions of food delivery across the globe, as well as the local specific deployments; this is a great example of the global-local approach of Prosus. The session was moderated by from TwiMLAI.com (the most well-known ML/tech podcast out there!) and will soon be released as part of the TwiMLAI.com podcast series.
- Breakout sessions to encourage data scientists to connect with like-minded professionals across the group, share experience and address important themes related to AI, moderated by topic experts. Infrastructure for ML, Data Science Career Talk, Expanding NLP Capabilities, Computer Vision, Personalization, Embeddings, Augmentation and Return on Investment of ML projects are few themes we discussed.
- Fireside chat with Prosus CEO about his views and expectations for AI at Prosus. This offered an opportunity for participants to directly interact with the CEO and ask questions on his AI vision for Prosus.
- Demo Booths for people to showcase their work and learn from colleagues. We had 20 demo booths on a broad range of topics from running AI-driven Innovation projects, to Knowledge Graphs or Active Learning. This session made it absolutely clear to everyone about the collective knowledge we have as a group and the amazing portfolio of machine learning applications across the group.
- Debate on MLOps to have our internal experts debate contrasting views on MLOps, like ‘MLOps is just glorified DevOps’, ‘MLOps is more organizational rather than technical’ etc. This session was moderated by from MLOps Community.
- We invited Josh Tobin to give a keynote on best practices for production deep learning, and Hilary Mason to share her views on trends in ML for 2020 and beyond.
- Finally we held two Technical, hands-on Workshops on Bias & Fairness with Pedro Saleiro and Deep Learning without the bugs with Josh Tobin.
3. Any good event needs some swag!
The fact that the event was virtual, did not deter us from shipping some cool marketplace swag to our participants. And the icing on the cake was the surprise food delivery we sent to our participants around the world, of course using our own portfolio food delivery companies - Swiggy, Delivery Hero, iFood, wherever possible. As you can imagine, logistically it was not an easy task, but our event management team pulled it off. We realized that especially in times like these, such small gestures go a long way to revive the feeling of belonging to the community.
4. Cater to all timezones
One big challenge in organizing a live virtual event across the world, is the timezone difference. We had participants and speakers from Brazil, Europe, South Africa, US and India. Brazil is almost one working day behind India. For inspiration, we looked at some recent ML conferences that were hosted as a virtual event. Here’s a nice analysis from ICML. After considering options like pre-recording videos or organizing same session twice, we decided to spread the event over 3 days and found a compressed time window of 4.5 hrs that was still reasonable for all geographies.
We all know data science is a team sport and this event brought together a large group of talented people, which together make a world-class data science team. In hindsight, there are many real advantages to hosting this event virtually. Some are obvious (no jet-lags, shorter times away from the daily obligations, lower costs per person). Others are less obvious, yet important like the ability to open up the event to every data scientist across the group independent of their seniority or the proof that intense collaboration across a large group can be achieved while working from home.
Last but not the least, it takes an enormous amount of effort from a lot of people behind the scene to pull-off a successful virtual event, much more than organizing a face-to-face event. So we would like to thank all of them for their valuable contribution.
If you have any thoughts on how to organize a virtual event, or how to engage with a global community, especially in the context of data science, we would like to hear from you. Please feel free to reach out to us at datascience@prosus.com, or comment directly on the article.
