Discovering insights that support decision-making across the company that includes product

We're chartered with discovering insights that support decision-making across the company. That includes products; improving our cameras, our software, our cloud services, our marketing capabilities, and our work with finance – and to do that we work with engineering.

Our charter is to discover insights, in any way we can. And we're primarily using analytics and data science techniques, traditional logistic regressions, and different types of predictive modeling.

But, until now, we haven’t used AI or computer vision for this. So, this is an example where AI, computer vision, and deep learning are extending our capabilities. How do we discover insights we couldn't before? These types of techniques are a natural extension of our work.

How to use image classification to improve the performance of the product and the software team

I want to give an example. This story is about how an analytics team adapts to technology to extend our capability. And how do you drive change? How do you do that when no one on the team knows how to do deep learning, or is familiar with it, or understands how we would use it?

This is an example of the switch framework, researched by Dan Chip and Dan Heath. It's a very powerful technique that every successful change project includes. I'm telling the story of a team, adapting analytics, and using adapting computer vision to improve our insight, capabilities, and to drive change within the team.

We started with a question – we have a business problem or a question. Our users generate millions and millions of pictures and videos. And we have a general idea of what they create – lots of extreme sports, lots of cute videos of their kids, all kinds of things we wouldn't imagine.

But we don't know the percentages. How many are photos of children; is that 1%? Is that 50% of the content? That makes a big difference. If the images are indoors, then low light sensors, low light technology, and video editing software will work better due to the low-light.

Knowing that tells us what will drive improvements in our product. This is our objective. What is the content, in terms of segments, and what percentage of these types of images are being shot by our customers?

And so that will allow us, and it has already allowed us, to improve our features and performance on camera, in our apps, and in our cloud services. Feature enhancements and exploration of what could be a potential new feature to add to our products also comes out in new capabilities like 360 cameras, future cameras, and our future mobile app software.

Also in marketing, when we know the most significant percentage of our customers or group of customers, like a certain type of image capture activity, we can market directly to them with that relevant content. If you're a surfer or a mountain biker, we can cater that marketing to that group.

At the very least, we know if we're using this type of image that it's going to capture a larger market segment than if we use some very niche image or video. Also CRM advertising, so direct to consumer emails, in-app messaging that can be tailored to someone's interests, as well.

And GoPro Studio too. We have a studio team that creates a lot of launch videos and advertisements. They come to the analytics team, and they want to know what we should create, as there are creators with lots of ideas, but they’d like some guidance.

Should I be creating skydiving videos? Should we have this? This is like the dream team. They send these guys around the world and they just have fun and shoot cool videos. But they do want to know: “what should we shoot? What kind of content?” Understanding what our consumers are capturing can inform even our creative studio team.

Shifting gears back to how you drive change. We're a team of analysts and data scientists with machine learning experience, but not deep learning experience, and definitely not convolutional networks or image classification experience. But it's not a stretch. It's something we're capable of learning and doing.

We prove that on our team. The switch framework uses the analogy of a rider on an elephant going down a path, then the writer represents the rational, logical mind, and the elephant, the emotional mind, which is more powerful. You see that right away when you try to influence an individual, a team, or a company.

Then the path, where are we headed? I'll walk through an example of how I did this on the team. As the writer, you want to follow the bright spots. We started with not knowing where or what kind of technique to use. We looked at, in the company, who was using image classification techniques.

You script the critical moves, you chop them up, and make it very simple going from that question to quantifying millions of in segmenting millions of videos and photos. You need to break it up for somebody who's never done image classification before.

You point to the destination, what does that look like? Why are we doing this? Helping the data scientists understand why is a big deal. It makes a big difference. In the company, it turned out that the hardware team was actually using image classification to embed and create algorithms on our camera for scene detection, to improve the performance of the product and so was the software team.

If you haven't used our app, it's pretty sweet. It's using lots of this image classification that goes into the algorithm to help the software detect faces, smiles, and action points of interest in the videos, and it will select those interesting moments to automatically create a video that's better than most of us here could create in 10 hours.

And the Cloud team is using it to organize content. Three teams were already using image classification techniques. We found that the mobile team was using the Amazon recognition engine to feed and improve the mobile app algorithm. It turned out this is a perfect application for us on the analytics and data science team, to leverage Amazon's off-the-shelf recognition engine.

Just by understanding enough code on how to use it, how convolutional networks work, and how image classification work, we were able to figure out how to apply it to our use case. This is an example of what it looks like for us. It's not actual percentages, but it looks something like this.

Looking at millions of images, we can tell that a significant percentage of our customers create content that’s not jumping out of a plane, wingsuits, you know – it's not traditional, typical GoPro content.

My team has been able to influence the understanding of who our customers really are. And they're not just action fanatics. In fact, most GoPro customers are not adventure sports folks. We know there are these distinct groups and we know the percentages of the types of content created. So, we can then start to design, for example, indoors, like I was mentioning, low light sensors, low light technology, video software, that’s sensitive to darker images, that kind of stuff.

It helps us refine the product. It’s our goal to do this type of work consistently. But how did we get someone on the team to actually start doing this? First, we found the machine learning engineer who was willing and able to train us and found a motivated data scientist – her name's Megan, and she's extremely motivated.

If you want to keep your data scientists around, you need to keep them interested in learning, otherwise, they're out. We got the machine learning engineer to train us and help us understand how to leverage Amazon's recognition engine and document it for our use case, using our code for our content.

And then we tested it. We'll get the rest of the team training, and then it’ll snowball from there. Also, what's its purpose? We want to segment, so explaining to a data scientist who's focused on data and crunching the numbers. They don't always know the real purpose and impact of the work, so we help explain that goal. Then we can produce this type of classification similar to our machine learning engineering team on the software side.

We're going to leverage this to segment our customers from a marketing and a customer understanding perspective, not just for the algorithm in the app or the camera. The implications to applications are much broader, and even more impactful than product-level improvements.

So, motivate the elephant, find the feeling, shrink the change, and grow your people. All the research shows that we make most of our decisions, not logically, but emotionally. We feel something then we take action. We rationalize before we take action and then we think we're being logical.

The fact is, most of us don't make decisions that way. We make emotional, quick decisions and then justify them logically. Finding the feeling – this is finding the feeling and Megan, and then by extension, the team. Getting that person interested and motivated, feeling like this is a reason to stay at GoPro on this team, and learn and have an impact across the company.

This is somebody who has an analytics background, turning the cranks, who is very good at basic analytics, and lots of predictive analytics modeling. What's the next step? What's the reason for them to stay here? Helping someone understand on a professional growth level, like learning deep learning, using AI to extend your capability in your toolbox.

You're learning new skills; next level insights, high impact insights, which looks good on your resumé. You'll be able to demonstrate the actual impact of the work to hundreds of people on a product, and you can say “I made this change to the product.”

Shrinking the change is often daunting. I recommend one of these: it's an online course that Adrian's created and it’s really good. It helps you understand what deep learning is and how to do it for yourself on your laptop.

In this case we worked with Megan to apply image classification to small datasets. And then we built from there, scaling it from your laptop to the cloud to publicly available images. That’s a whole different lecture right there.

How do you actually manage and apply it on a small number of images and larger datasets? We walked her and the team through how to scale up. When we did that, it showed the whole team that a data scientist who's not a machine learning engineer can apply computer vision and expand out.

Now there are two other people on the team and we have three data scientists with the ability to apply deep learning techniques for image classification purposes.

The importance of rallying the herd and motivating the elephant

The next step to shape the path: How did we get to the three people on the team and potentially four or five more using it? What Chip and Dan Heath recommend, and what I've done for the team, is to make it a safe place.

We set up projects and goals; we have multiple image classification projects with our studio team with publicly available content, and with the awards videos coming in. We have multiple projects for the next two years that we can work on. So, it's something to look forward to: big goals.

If we don't get there, that's fine. We have fun with it, we explore and we learn something new. And that creates a habit. We have fun Fridays, where we're all learning something new and right now it's this. We spend a few hours on Friday learning and teaching each other all about how we can apply these techniques.

It creates this habit within the team where we're constantly learning in the right direction, which is basically what, once we had Megan interested, just snowballed. Because the other data scientists are like, “well, if she can do it, I can do it.”

Now there's a little bit of competition. It's what I'm doing with my team, which is rallying the herd, and in that sense, everyone's interested in moving in this direction and learning these new skills.

In summary, direct the writer, the logical person. If you hit these three points, you’ll make progress with the logical side of the individual and the group.

Motivating the elephant is really about finding that feeling. shrinking the change, and growing your people. If you find that feeling, people knock down walls and doors, they'll find a way to get to that solution or achieve that goal.

Shaping the path, it's just creating an environment within the team to learn. If you feel free and safe to learn, you'll be able to convince a team. These guys are all engineers. They're super focused, they love what they do. Thinking out of the box isn't always something they're interested in, but creating an environment where that's safe and encouraging, that's how I shaped the path. And that’s it from me.