Machine Learning seems to be getting all the interest and hype these days, and some are even saying that it’s going mainstream. There are even dedicated conferences and summits for ML just like the 2021 AWS Machine Learning Summit. For ML to go mainstream, in my perspective, there are still real-world lessons we’ll need to translate ML into production for businesses, and I was hoping to get some takeaways from this summit. I listed here some parts that made the most impact on me. Hopefully, you’ll find these useful when you are planning to apply ML:
- Pizza with the right amount of cheese example with Swami Sivasubramanian
- Computational humor talk with Yoelle Maarek
- Getting your hands dirty chat with Andrew Ng
Since the talk has been organized by AWS, the summit is leaning towards the use of their ML services, but the takeaways here can also be applied to other cloud computing platforms like GCP that offer their own ML services.
Pizza with the right amount of cheese
A possible indication that ML is going towards the mainstream is its use in all kinds of businesses including the pizza business. Swami Sivasubramanian presented Dafgårds as an example of a business that can apply ML even with a limited training dataset. They wanted to use ML to improve the quality and efficiency of placing the right amount of cheese in each pizza dough. Yes, what a nice way to start with things I like: ML and pizza.

They used visual quality inspection with few shots learning that can learn a specific task with just a few examples. I think this is a pertinent use case for an ML application with a limited training dataset. Data collection – obtaining the right kind and amount – is usually a pain point when starting with ML projects. Going further, I was intrigued at the AWS Lookout for Vision that was conceptualized from customer feedback together with this few shots learning idea.
There were also your other companies and businesses mentioned that you’ve probably heard of, like BMW and New York Times, but that pizza idea stuck around.
Computational humor
For ML to be mainstream and to stick around, there’s also the importance of looking at user experiences. And, for quite an interesting twist, AWS is seriously looking at computational humor spending hours and hours in this research. I did like their approach of immersing the ML team into the product team. In this example, the computational humor group has been working alongside the Amazon Alexa Shopping team.
In finding ways of delighting the customer, I just can’t help think of a recent refreshing advancement with AI chatbots. Having conversations with inanimate objects like Pluto and paper planes in Google’s LaMDA program in Google IO!
Now, back to computational humor, the team was not just making the AI funny, but the team wanted to know if the customer is funny and how the AI should react. In this case, the customer is the one taking the initiative.
For the detection and possible response, the team even tried to look at cultural references, sarcasm, or if the customer is in the mood to be playful. I like how the talk also included references say to mixed-initiative theory, relief theory, and a few more.
For example, for the Nintendo Switch Gray Joy-Con product, one customer replied indicating some cultural references: "Can I use this to hack into the matrix and save humanity?"

Another customer replied showing sarcasm on a very expensive water cooler: "Will this thing make me fly? It seems due to the price that has to do something special."
Yes, the current ML status can already do so much, but there is still a lot to explore and novel ways to use it.
Getting your hands dirty
And then, there’s the fireside chat of Andrew Ng and Swami Sivasubramanian. There were quite a lot of interesting points in that less than 30-minute chat but the highlight for me was actually the first few minutes!
The first part was giving some notes for companies, addressing their leaders like CEOs and CTOs, that want to jumpstart their machine learning. When you’re leaping towards ML for the first time, you have to get your hands dirty.
Taking too long to plan and get started is likely a mistake. For the first project, the data can be messy but there could be workarounds, and the business should be able to start with small pilot projects or proof of concepts and have a quick win. Learn from those to grow and scale the project.
And I did like the idea of trying out a pilot project before forming the long term strategy for machine learning. The learnings from the POC will help generate a strategy that aligns more with the company goals and the people culture.
I did mention in a previous article on Reinforcement Applied to Business Problems that the biggest obstacle in RL lies in the problem formulation and the next steps would be shaped by that formulated problem. So, care should be made in the problem formulation. But even having said that, it shouldn’t prevent the people from getting their hands dirty and still try out. In the long run, there could probably be an ML platform but it’s always those first POCs and quick wins that provide insights on the direction of the ML journey.
And now, even for non-leadership roles, when starting an ML project, for a completely novel application, it may be difficult to create target metrics to check if the ML has been successful or not until the ML team has done a proof-of-concept or if there are related projects in the literature to help define a base-level performance. The key here is likely to go for a quick and dirty prototype system, iterate, and learn from those experiences.
Parting notes
An increasing number of companies are adopting ML, but this growth isn’t slowing down anytime soon. It may have started with tech companies but ML projects are now seeping through other companies including those pizza companies. The key part and recurring theme in the summit has been in translating ML into production.
Addressing pain points like having a large training dataset are crucial, so having techniques like the few shots learning is a welcome change. And for these ML projects to succeed and not just be all hype, customer experiences should be given priority. Investing in research that addresses customer experiences like those with computational humor becomes not just a side-project but an essential aspect of an ML product. And it’s just refreshing to see different perspectives in the world of ML from time to time. Considering ways of working such as embedding the ML team into the product or development team may be a key part in the success of the project.
Translating ML into production would also mean going beyond the theoretical, and especially for companies that are starting their ML journey, getting their hands dirty with small pilot projects and have those quick wins can help provide some direction on a longer strategy for ML in the company.