Thu, Jun 16, 2022
When starting to organize this Canada trip, we weren't much familiar with Toronto's tech and Machine Learning ecosystem and community. We even thought we didn't have much network and connections there since we believed Tryolabs' name hadn't made it to the Canadian market yet. But, the moment we started to put our eyes on Toronto -the 4th largest city in North America- we realized globalization is real and that if the world is a small place, the tech Machine Learning world is even smaller. When we landed in Toronto, we were surprised by the number of welcoming friends and partners that already knew Tryolabs AI & Machine Learning expertise and were eager to show us Toronto's opportunities. Thanks Dave, Rima, Joaquin, Fabi, Daniel, Amy, Andrew, and everyone for this fantastic stay and experience.
Participating in the MLOps World conference confirmed that the Canadian market is already significantly big enough to invest in and explore, and that we underestimate it while focusing on the US market.
The Canadians have been doing fantastic work to position themselves as the competing Silicon Valley. The Toronto Machine Learning Society- thanks to visionaries and leaders such as David Scharbach- is part of these incredible efforts proving great value to Canada and the world.
The MLOps World Conf started showcasing precisely this. With Jason Carter’s opening remarks, we were all reminded of the vast indigenous history the Canadian land has.
And Tomi Poutanen's keynote speech hinted why Toronto is so important and considered the place of birth of deep learning.
Tomi Poutanen, a former student of Geoffrey Hinton, the godfather of neural network research, shared his takeaways on how this thriving AI ecosystem has started. According to Tomi, Toronto has always brought people together from different facets and backgrounds. While in 2017 there was almost no concentration of AI/tech startups in Canada, this has changed rapidly thanks to multiple factors.
Among other efforts, the AI Ecosystem in Ontario was shaped by the loyalty of a few pioneers and experts from the University of Toronto and Waterloo. They created a diverse, collaborative, and welcoming community integrated with government support. Now, this AI ecosystem is nurtured by:
All of which generate an ideal ecosystem to develop impactful companies and solutions that are changing the world with the support of a diverse community from different backgrounds, regions, genders, and cultures.
Tinder, Zoom, Meta: they all tried to give us the idea that the online virtual world could substitute the in-person experience after the pandemic. But once again, being able to experience the in-person experience reinforced the belief that there is no way to replace it and that we shouldn't try to do so. Beers, after parties, coffee breaks, lunches, and informal conversations with speakers and attendees were the most valuables moment of the conference. Thanks MosaicML, Aporia, Arize & H20.ai, Qwak, Featureform, and Modzy, for putting all this together.
This is the way communities and long-lasting relationships will continue to flourish
When entering the sponsors' booth room, our first impression was to feel overwhelmed by the number of companies apparently doing very similar things. There are MLOps platforms and products for:
With the rise of AutoML and the increase in companies developing products to accelerate AI/ML development, the dilemma of building vs. buying gets more relevant every year. Should a company invest in building their own MLOps platform, or are there any pre-built solutions to buy? I think Machine Learnings engineers are, by nature, doers and builders. They enjoy the heavy lifting of Machine Learning development, and often because of their roles and responsibilities, the balance is biased toward the building side. Also, they generally do not have the proper incentives or time to understand, benchmark, and critically analyze all the latest solutions that are out there. But when looking at each product, the question is: how does each product differentiate from the others? Is there any particular out-of-the-box solution better than the others?
After talking to many of them and hearing the experience of big players in the industry, it dawned on me that these 15 - 20 solutions will not necessarily compete with each other, or at least they shouldn't.
But they will end up complementing each other. And that a robust MLOps Platform such as the one presented by the Loblaw Digital team is a mix of Building+Buying.
In general, at Tryolabs, we still refer to Data Scientists and Machine Learning Engineers indistinctly. The main difference could be that Data Scientists generally have no formal engineering backgrounds (Economists, Linguists, Marine Biologists, Chemistrists, etc.) vs. Machine Learning Engineers (Software Engineers, Electrical Engineers, Mechanical Engineers, etc.) who do. However, during the MLOps World conference, the main difference between these two roles was reinforced in several talks and even had its panel.
In the panel, What do Engineers Not get About Working with Data Scientists, Azin Asgarian (Georgian), Asmita Usturge (Microsoft), Sarah Sun (Scotiabank), Seder Akinli Kocak (Vector Institute) joked about the long-existing rivalry in between the teams and their personal characteristics.
Also, Demetrios warmed up the discussion by posting the controversial question of Jupyter notebooks in production or not? But the definition that generated the most controversy was that Data Scientists are stubborn and very curious by nature, and ML engineers have a higher ego and are practical doers.
But leaving jokes and generalizations apart, they all agreed that the unicorn Data Scientist who is an expert in Math/Statistics, Programming, and Business is and should remain a mythical creature. As both the Data Science and Machine Learning engineering fields and practices matured, they have moved each other apart, which could be beneficial or not. Having this division in roles, skillsets, and characteristics was supposed to generate higher specialization, allowing everyone to focus and grow in their specific roles. However, having Data Science teams and Machine Learning Engineers teams speaking in different languages and using different tools to build the same systems has, in many cases, threatened many organizations and AI / Machine Learning development.
The recipe for success in all cases was making sure they all speak the same language and have a clear understanding of the MLOps best practices. One could think that MLOps is part of the background and toolkits only Machine Learning and Data Engineers should have. However, to have PoCs that evolve into successful production systems, DS must follow and understand the same language and, at the end of the day, have the same toolkit.
MLOps can be thought of as the latest field in Machine Learning. The first MLOps World conf took place in 2019, and all the sponsor companies have less than 3-4 years in the market. So predicting the future of this very early staged field is hard to tell. But we can surely say that MLOps is here to stay and that the hype for MLOps platforms, tools, conferences, and expertise will continue to grow.
As we have seen and presented in MLOps: we are all doing AI-ish projects, but many of these projects make it to production. And a model that never makes it into production is incapable of producing value for a business or organization. Therefore, the need to crack MLOps and generate actual business value will be crucial for future years.
So these MLOps solutions, best practices, and tools will be the future, and it's better to ride the wave. Thanks, David, for the invitation and the opportunity to learn from many MLOps experts. Thanks, Toronto and everyone at the conference for such warming welcome. Lastly, thanks Fabián and Diego, for your company and contribution to this summary of takeaways.
Very much looking forward to the next MLOps World Conference 🙌 and to keep exploring the Canadian market.
See you in Collision - Toronto 2022 next week!