Democratizing AI: It’s anyone’s game

“Humanity’s history is a story of toolmaking, and AI is a tool like any other. It’s not a replacement for people. What it is, is a very specific tool — a very powerful tool that, like the computer, will create a lot of opportunities.”

Google Cloud chief decision scientist Cassie Kozyrkov is on a mission to democratize AI. At the recent Rakuten Technology Conference, she joined Rakuten’s own chief data officer Ting Cai and global head of the Rakuten Institute of Technology Ewa Szymanska on a panel exploring everything from the very fundamentals of AI to its modern applications.

“It’s important to know exactly what AI is a tool for: It is a tool for writing software,” Kozyrkov continued. “It’s actually another approach for telling computers what we want them to do for us. And this is very simple and very powerful.”

A new way to write software

At the core of every AI is a model, Kozyrkov began.

“What is a model? It’s just a fancy word for recipe — a recipe for converting inputs into outputs. Now that is what computers do for us — that’s what they’ve always done. And that’s what they do in traditional software engineering as well.”

How this model is created is what sets the two disciplines apart.

“Now where does that recipe or model come from in traditional programming?” she posed. “It comes from a human programmer thinking really hard about the problem… Kind of handcrafting that recipe and explicitly expressing to the computer, here is what you do with the input to get the output. Step by step. That’s traditional programming, that’s traditional code.”

Kozyrkov employed the example of training a system to recognize an image of a cat. To do this traditionally, a programmer would need to go through the painstaking process of defining the essence of a cat: “Triangles for ears, ovals for eyes — what about cats with flat ears?” she ventured. “It’s going to be pretty hard for you to come up with a set of rules that capture catness.”

With AI, however, this near-impossible task becomes unnecessary.

“What if we had another way to communicate our wishes?” she proposed. “What if we could explain not with instructions but with examples. That… is what machine learning and AI are all about. That is the power of it. Instead of having to come up with the instructions, you now come up with examples. You say, here are a bunch of cats, here are a bunch of not cats, and from there, you figure it out.”

AI is a communication problem, not a technical one

“Notice that this isn’t about robots or any of this kind of fancy stuff. This is about communication. This is a second way to tell machines what you want them to do.”

Kozyrkov reminded the audience that this method of teaching is something that humans have been doing forever.

“We as humans already have these two modes of communication when teaching one another. Sometimes if I want to teach you how to do a task, I’ll give you the exact steps,” she pointed out. “Sometimes I say, here, watch me do it a few times and then you figure out the patterns and come up with the instructions yourself from there.”

It’s AI that makes this teach-by-example approach viable with computers.

“With machines, traditionally we only have one way to program it: explicit instructions. Now, we’ve got both. We can use examples or data. This means that even if we can’t say how we do the task, we still have a chance at automating it anyway,” she shared. “And that is extremely powerful and that’s why you should be excited.”

Are companies investing in the wrong kind of AI?

“Now, why do businesses fail at making use of such an incredibly powerful technology with so much opportunity? I think that the secret here is that there are actually two machine learnings — two AIs.”

Kozyrkov challenged her audience (of techy conference-goers) to raise their hand if they knew how a microwave works well enough to build one themselves.

“How do you trust that thing? You don’t know how it works. Why do you use it to heat your food? Because knowing how it works isn’t the way to trust. The way to trust is making sure that it delivers for you exactly what you needed it to do.”

Kozyrkov believes that companies need to approach AI in the same way.

“There are two AIs out there, two machine learnings,” she repeated. “They are the difference between building new microwaves and using microwaves to build new recipes… Two completely different disciplines that both go by the same name. And this creates a lot of confusion. I think that where businesses fail is that they think they want the one — they hire for the one — but actually they want the other.”

Today, there is no shortage of AI-powered tools that can be used to innovate on all manner of problems: “Warehouses upon warehouses of microwave ovens that other people have built,” she quipped. “That is the new AI. That is the applied AI. And that is the new revolution in software.”

Democratizing AI — it’s anyone’s game

“There is a lot of space for a lot of different players here, now that we’re able to apply AI to many different interesting business problems.”

The field of AI has come a long way, and today it’s more open and beginner-friendly than ever. No longer is it the case that only AI researchers can make a significant impact.

“Now, anybody who understands the problem that they’re solving, who understands how the business and the real world works around that, who has technical skills and who is interested in constructing technology and solutions for solving those problems — anyone with any perspective on that whole ecosystem of stuff is very welcome and very needed in AI,” she stressed.” All these tools are available for use because the research history is long and it’s been packaged up and ready to go. You don’t need to know how the microwave works in order to use it.”

Kozyrkov encouraged the audience to explore the many tools already available.

“Go play. Have fun. Try out these things that already exist and have been built for you. And put them together towards the goal of solving your business problems,” she encouraged the audience. “Put your own data through them and see if you can get your problems solved with what already exists. There’s a lot of space, just dive in. You don’t have to read a stack of textbooks first, I promise.”

AI is collaborative, human-oriented

Kozyrkov also warned developers not to forget who they are building their systems for.

“It’s very important to consider the context for how it will be used and to make sure that you’re thinking about how to enrich someone’s experience,” she stressed. “How to save time; how to improve lives; that you’re really putting the user first — because fundamentally, this is going to be a very human thing that has a lot of potential to enrich people’s lives.”

AI systems are, in the end, built by humans.

“To understand the people that you’re serving, it helps to have a lot of different views represented on the team. A higher probability of overlap with the people that you’re actually building for.”

She reminded the audience that everyone in their organization will be involved in the final result. “What I would like people to focus on is this notion that you — your whole team, which would include everybody on it — the business leaders, engineers, the statisticians, the designers, everyone — that you are serving as a teacher.”

“Humanity’s history is a story of toolmaking, and AI is a tool like any other. It’s not a replacement for people. What it is, is a very specific tool — a very powerful tool that, like the computer, will create a lot of opportunities.”

Cassie Kozyrkov, chief decision scientist, Google Cloud

In this context, Kozyrkov doesn’t believe that there is any one type of person who is well-suited to working in the AI space.

“When I’m asked the question, Cassie, what is your ideal kind of AI employee or AI person? That is my least favorite question I can get about AI. The reason is simple. It’s not just a one-person thing. It’s not a one-person job. It’s a team that usually consists of many different people who are playing different roles and who are working together and bringing their shared disciplines and perspectives. So it’s very interdisciplinary, it’s very collaborative. And I often find that where you’re more likely to fail is not on the technical stuff — it’s on that interdisciplinary collaboration.”

It’s this collaboration that will allow AI to flourish, Kozyrkov stressed.

“Take the perspective of this being a collaborative thing that you are designing and that is how you build that brighter AI future.”

Rakuten’s winning combination

Kozyrkov had praise for Rakuten’s rare combination of research and application.

“Often you have either good research in isolation, or you have application-based research that’s not done in-house,” she lamented. “Both of these are very important things. They rarely exist together with the kind of excellence that your institute represents. So this is really phenomenal.”

Rakuten’s own chief data officer Ting Cai revealed his impressions after having joined the company earlier this year. “I have never worked at a company where everyone talks about data and AI. So that makes me feel very excited.”

Cai also praised the high quality of the data at his disposal.

“I feel like Rakuten has a lot of unique data,” he continued. “We have a lot of transaction data, offline data. That’s fantastic data for us to apply. And using that data gives us unique opportunities to create unique solutions and original algorithms.”

Global Head of Rakuten Institute of Technology Ewa Szymanska also encouraged employees to use the assets that Rakuten makes available to innovate.

“Look again at the breadth and variety of businesses that Rakuten has and just imagine the wealth of data and the possibilities that it opens up by combining them to build things that don’t exist today.”