Machine Dreams AI transcript and summary - episode of podcast Think with Google Podcast

Go to PodExtra AI's episode page (Machine Dreams) to play and view complete AI-processed content: summary, mindmap, topics, takeaways, transcript, keywords and highlights.

Go to PodExtra AI's podcast page (Think with Google Podcast) to view the AI-processed content of all episodes of this podcast.

View full AI transcripts and summaries of all podcast episodes on the blog: Think with Google Podcast

Episode: Machine Dreams

Author: Think with Google / Gimlet Creative
Duration: 00:22:07

Episode Shownotes

In this episode, we’ll talk with experts at the cutting edge of machine learning technology and learn about its possibilities. We’ll discuss one of Google’s innovative tools with Pinar Demirdag and Alexander Mordvintsev. We’ll also talk with Google experts Avinash Kaushik and Ben Jones about how machine learning can be

the marketer’s assistant. Learn more about your ad choices. Visit podcastchoices.com/adchoices

Full Transcript

00:00:00 Speaker_02
There are different types of creativity. Analytical power is creativity. Artisanal power is creativity. To be able to comprehend a complex problem and give a one simple answer in return is creativity.

00:00:17 Speaker_05
This is Pinar Demirdag, and she knows a thing or two about creativity. She and Viola Renate are a digital artist duo known as Pinar and Viola. The two often use technology in their artwork and fashion design.

00:00:31 Speaker_05
In fact, Pinar sees it as a way to push artistic possibilities.

00:00:35 Speaker_02
And technology, artificial intelligence, it can analyze data and create solution in a speed our linear human minds cannot comprehend.

00:00:46 Speaker_05
Pinar says there's a clear role for machine learning in the creative process. I see it as the ultimate assistant. Welcome to the Think with Google podcast. I'm Tess Vigeland with Gimlet Creative.

00:01:03 Speaker_05
And in this show, we're bringing you inside knowledge to power your marketing. So far in this series, we've explored diversity-focused data, personal video trends, and why you should never underestimate the power of surprise.

00:01:18 Speaker_05
Today, we'll talk with people at the cutting edge of machine learning technology, and we'll explore how marketers can partner it with creativity.

00:01:30 Speaker_05
A few years ago, Pienaar and Viola were visiting Paris when Damien Henry, the technical program manager for the Google Arts and Culture Experiment team, asked the duo to join him for lunch.

00:01:43 Speaker_02
He was quite excited to show us something, but also quite nervous.

00:01:48 Speaker_05
Over the meal, he showed them psychedelic images produced by a code called Deep Dream. He told Pinar the code was written by an engineer named Alexander Morvintsev at Google's Brain Lab in Zurich.

00:02:01 Speaker_02
And as Pinar tells it, Damien Henry went on to tell them that Alexander Morvintsev, the genie behind Deep Dream, invented a new way of image creation by using artificial neural networks, and it looks exactly like your work.

00:02:21 Speaker_05
Pinar says she was more than a little surprised by what she saw. The computer code had manipulated the colors and forms of everyday objects into something otherworldly, seemingly from a parallel universe.

00:02:34 Speaker_02
The patterns that Alexander was creating by using his quote-unquote accidental tool, I got like shivers in my spine because it looks exactly like my body of work as an artist.

00:02:49 Speaker_05
But Alexander didn't write the code at work. In fact, a bad dream about somebody trying to get into his apartment woke him in the middle of the night.

00:02:58 Speaker_05
He was spooked enough to check the garden door and make sure everything was fine, but he couldn't go back to sleep. Instead, he decided to work.

00:03:10 Speaker_01
I grabbed my laptop and spent, I don't know, like 20 minutes just writing down this code for experiment I had in mind for quite a while.

00:03:20 Speaker_05
So once Alexander had the code down, he uploaded an image to try it out. To his surprise, really strange patterns emerged, which wasn't exactly what he thought would happen.

00:03:32 Speaker_01
Well, that's not something I was expecting. I was hoping that it will improve quality of image, maybe act like a super resolution.

00:03:44 Speaker_05
He wanted more people to try the code, so he posted it from Zurich while it was daytime in Google's Mountain View offices.

00:03:53 Speaker_01
I shared the code internally, and there was hundreds of people playing with this.

00:03:59 Speaker_05
Excitement around Alexander's tool, which would eventually be called DeepDream, grew faster than he expected.

00:04:07 Speaker_01
When images started to leak outside, we thought, OK, we should really proceed fast and publish this. So we published image galleries, we published the code.

00:04:17 Speaker_05
And the images were so strange and exciting that the code itself went viral when it was released to the public. The sharing of Deep Dream wasn't a marketing campaign for Google, though. It was released so the public could experiment with the code.

00:04:32 Speaker_05
In turn, Google received more information on what exactly the code was capable of, which is typical of how Google tends to work.

00:04:43 Speaker_01
Google has a research lab currently in the world of machine learning. This research is happening in a very open fashion.

00:04:53 Speaker_05
So fast forward to Pinar and Viola at lunch. Pinar says she was surprised by the similarities between Deep Dream images and the work they'd been doing in their studio. But they decided to embrace the technology.

00:05:06 Speaker_05
Then we wanted to do something quite original. So the two of them paired up with Alexander to create a variation of the coding tool that they named Infinite Patterns. It's like Deep Dream, but its manipulation of images is more narrowly focused.

00:05:23 Speaker_02
We adjusted the tool to create self-repeating patterns, because before our collaboration, it was only making images. And then we said, if you make patterns, it can have also a purpose, because with patterns, patterns can cover any surface.

00:05:40 Speaker_05
Once you've uploaded an image, you can give the tool some directions in how you want to manipulate it.

00:05:46 Speaker_02
You curate the input and the making process, but the tool makes it for you.

00:05:53 Speaker_02
So the outcomes that you see that are very curvy, I call them delicious, juicy, pop, round-edged, that is a result of my own consciousness, my emotional taste as an individual.

00:06:11 Speaker_01
Actually, it looked a lot like abstract art.

00:06:18 Speaker_05
The code works as a collaboration, a kind of middle ground between machine and creativity.

00:06:25 Speaker_01
Obviously, machine doesn't do it on its own. At least now, there is someone whose intention drives the machine.

00:06:33 Speaker_05
So as humans are figuring out how to dream with machines, here's the big question.

00:06:39 Speaker_03
If an algorithm dreams, what does it dream of?

00:06:44 Speaker_05
That's Avinash Kashyap, who leads the strategic analytics team for Lorraine Tuhill, Google's CMO. We heard from her back in episode one. Avinash has written two books on data.

00:06:55 Speaker_05
He says data scientists and engineers are exploring ways to push machines to think more innovatively.

00:07:03 Speaker_03
which are trying to explore this other side where some people believe that what makes us unique and what machine learning algorithms might never be able to accomplish is creativity.

00:07:15 Speaker_03
These algorithms have learned from the greatest masters that we've ever had in the human race, like Rembrandt and Picasso, and you can tell them to paint like Picasso. And it is amazing how good they can replicate Picasso's style.

00:07:31 Speaker_03
Of course, they can't think like Picasso. His genius is missing from it, but they can mimic. But what exactly is this technology that data scientists are using? You can think of them as smart algorithms. A lot of people call it machine learning.

00:07:48 Speaker_03
And what happens is these algorithms are substantially more effective than us at being able to ingest massive amounts of data and be able to find the patterns that solve some very fundamental human problems.

00:08:04 Speaker_05
One example could be something like buying a new toothbrush. Avinash says he and his lab took a hard look at applying ads to people using Google's search engine.

00:08:15 Speaker_05
Say this person is searching around the internet, looking at a few different electric toothbrushes. Google's search engine comes up with an ad that's best suited for what they were looking for.

00:08:25 Speaker_03
Given the amount of data we have and given the amount of complexity in the world, an algorithm is much better at creating that ad. So we call that smart creative.

00:08:34 Speaker_03
So in AdWords, for example, now, we use machine learning in order to figure out what is the best answer to every question that Google gets asked.

00:08:42 Speaker_05
the user doesn't buy a toothbrush immediately, but goes elsewhere on the web, maybe to check the news online.

00:08:50 Speaker_03
So this is a Monday, it's late in the evening, not at work, completely different mindset, thinking of different things.

00:08:57 Speaker_05
But despite all that distraction, Avinash and his team had already told the machine learning algorithm, a form of artificial intelligence, to create an ad that would be best suited for that buyer on this news website.

00:09:13 Speaker_03
And what the AI did is assembled for us 53 million variations of ads that were perfect for every single person who saw that ad. And the fact that it could understand what piece of creative would work for each person was almost magical.

00:09:33 Speaker_05
Tailoring ads really worked. It made the selection process much easier for the buyer, and Avinash and his team saw immediate results.

00:09:43 Speaker_03
We saw a lift in downloads. We saw a lift in brand recognition. We saw a lift on all the core metrics. We tracked six core metrics. It lifted all of them. Simply put, the AI fine-tuned the search.

00:09:56 Speaker_03
It is absolutely fair to think that all these smart algorithms are there to assist us be smarter, better, more clever, spend more time with our families, find more joy from our work. But they're very much in an assist mode.

00:10:11 Speaker_05
Avinash says some marketers don't always buy into machine learning.

00:10:15 Speaker_03
My gut feel is better than your algorithm. That's a big, big one, by the way. People are like, I am smarter than it.

00:10:23 Speaker_05
And consumers tend to echo this kind of skepticism about targeted data.

00:10:28 Speaker_04
What does my data mean? Am I in control of the way that information comes to me and the information that is there about me? This is Ben Jones, creative director at Google and founder of Unskippable Labs. And those things right now are not in sync.

00:10:43 Speaker_04
We're figuring that out as a society, as a political system, as technology platforms. And so it feels a little bit out of control to us, right? Those things don't exactly sync up and we don't exactly have aligned expectations or understandings.

00:11:00 Speaker_04
So I'd say that's in a state of evolution.

00:11:03 Speaker_05
With all of the changes machine learning brings for both the marketer and the consumer, Ben says balancing AI is going to take work.

00:11:12 Speaker_04
For us, ultimately, the user is the person that's the most important. Do they feel in control of the data? Are they getting value out of it? And do they understand sort of how their data is being used? Are they comfortable with that?

00:11:25 Speaker_04
That's ultimately where we land.

00:11:27 Speaker_05
After the break, we'll continue to talk with Ben about how machine learning can help marketers across the board.

00:11:42 Speaker_00
You're listening to the Think with Google podcast, brought to you by Google. At Think with Google, it's their mission to make marketers more knowledgeable by providing research, insights, and perspectives that change the way marketers do business.

00:11:56 Speaker_00
In this episode, we're talking about how marketers can incorporate machine learning technology into their daily practices and how they can partner it with creativity.

00:12:05 Speaker_00
For more on how machine learning can power your marketing, head to thinkwithgoogle.com slash machine learning. Again, that's thinkwithgoogle.com slash machine learning.

00:12:21 Speaker_04
The most important aspect of machine learning is understanding that we don't understand it yet.

00:12:27 Speaker_05
Here's Ben Jones again. He's a creative director at Google and founder of Unskippable Labs.

00:12:33 Speaker_04
Machine learning is using massive sets of data to identify patterns based on a hypothesis.

00:12:41 Speaker_05
In the first episode of this series, we talked with the Geena Davis Institute about its collaboration with Google and USC. We learned how they're using data from machine learning to look at gender parity in film and TV.

00:12:55 Speaker_05
Hard numbers provided the Institute with ways to prove the hypothesis that Hollywood simply doesn't provide as many roles for women. These same rules apply when it comes to proving a point in marketing.

00:13:08 Speaker_04
So you have an idea, and machine learning can comb through a lot of data to say, is this true or not true, and how true is it?

00:13:17 Speaker_04
But the downside is that you need to understand how to ask the question, and you need to understand how the pattern recognition works in order to get value out of it.

00:13:29 Speaker_05
Ben likes to use a metaphor for how he thinks about machine learning and how it's helping marketers today.

00:13:36 Speaker_04
The best mental model, the best explanation that I've heard comes from a guy named Ben Evans. And he said, it's like having a billion interns and not having an Einstein. Like it's not going to be Hal.

00:13:49 Speaker_04
It's not going to be a genius, you know, human matching intelligence. It is a sort of massive brute force that if you train correctly can do extraordinary things. And that's no slight on hardworking interns.

00:14:02 Speaker_04
If you've ever had an intern, like, you know, at the beginning of the summer, especially, it's hard to know how to ask them and what they understand and what they come back with.

00:14:12 Speaker_04
And then, you know, you train them to understand what you're looking for and how to look for it, and you get more and more value. Well, you multiply that by a billion and you see some real value and some real complexity.

00:14:24 Speaker_05
Every day, Ben and his lab look at reams of content to analyze ad performance.

00:14:30 Speaker_04
My team's job is to look at a billion and a half hours a day of watch time on YouTube and say, what can we find that's valuable for creative agencies and advertisers?

00:14:42 Speaker_05
And each month, Ben and his team pull the top and bottom performing ads in the world and comb through them for a few specific metrics.

00:14:51 Speaker_04
What patterns do we see for ads in auto or in beauty? Or what ads drive store visits most effectively or increase consideration?

00:15:02 Speaker_05
This matters most importantly because time is of the essence.

00:15:07 Speaker_04
And the split really is, if I don't want to watch it, if I don't choose to watch it, then six seconds is about what I'll tolerate, right?

00:15:15 Speaker_04
So the bumper ad format that we use, which is six seconds long, came because that was the sort of magic point where people dropped out.

00:15:23 Speaker_05
And what Ben and his team at the lab are noticing is that as ads are evolving, so are consumers. In particular, how we're engaging with ads.

00:15:35 Speaker_04
We've never been so sophisticated. We've never consumed so much media so powerfully, so richly. We're the most sophisticated media consumers in the history of time, in all forms, in all screens.

00:15:48 Speaker_04
In the last year, we're spending more time on screens than we do with any other activity besides sleeping, between work and leisure, mobile, TV, desktop, et cetera. So we're incredibly powerful users of screen-based media.

00:16:04 Speaker_05
So what does that mean for marketers? They have to battle for attention through an ever-increasing amount of information and distraction. This is why being better able to connect data points is so important.

00:16:17 Speaker_04
Part of the reason that machine learning is a huge topic now is not because we've suddenly figured out it could be helpful, but because there is so much more processing power and so many more of our data sets are connected in a way that we never have before, that we can ask questions we've never been able to ask and get answers back at a speed and with a clarity we've never been able to get.

00:16:39 Speaker_05
But as Avinash, the leader of the strategic analytics team for the CMO at Google, said earlier, marketers too might have concerns about where changes in technology leave them.

00:16:54 Speaker_04
The people who are worried are worried that their jobs are going to get taken away, right? A machine's going to write the copy, or a machine's going to be able to figure out my brand. And I think that those people should not be worried.

00:17:05 Speaker_04
It's not going to write ads. It's not going to originate campaign ideas. Those are places where humans are still much, much better than machines. But it is helping to provide sort of tuning that is very helpful and very valuable for advertisers.

00:17:20 Speaker_04
Here's Avinash Kaushik again.

00:17:22 Speaker_03
We are at the cusp of a kind of transformative change, the last one of which humanity saw, perhaps, was with the Industrial Revolution, where everything changed in so many ways. Business changed, working changed, workers changed, everything changed.

00:17:42 Speaker_03
And I subscribe to that point of view, that we are at the cusp of everything changing.

00:17:51 Speaker_05
But Google wants to make sure the brands it works with are ready for the future. In a project Ben Jones and his lab did with BMW in Germany last spring, machine learning was integral in developing the best ad for an electronic vehicle.

00:18:07 Speaker_05
For the collaboration, there was a control ad, just as BMW had created it, and then two variations refined by Google. In one version, Google played with elements of the ad's narrative arc.

00:18:20 Speaker_05
And in the other, it was much less story-driven and more focused on how the car's features were presented.

00:18:28 Speaker_04
And when we ran those in market simultaneously against controls, the narrative version did 30% better than the control, and the feature-driven version did 200% better than the control.

00:18:41 Speaker_05
For Google, making changes based on data is less about finger wagging and more about making sure consumers actually watch the ad.

00:18:50 Speaker_04
If you want to make a black and white silent movie and win an Oscar, please feel free, right? There's been one of them since 1927. Same thing for ads.

00:18:57 Speaker_04
You want to open your ad with a long, slow tracking shot of the Pacific Coast Highway, please feel free. I've watched your audience disappear 99 out of 100 times before a car even shows up.

00:19:12 Speaker_05
But even with all this great information provided by machine learning, the tool is still finding its way. One of the first machine learning projects Ben's team embarked on involved a team of data scientists working within 200 codes.

00:19:25 Speaker_04
And as they looked at, I can't remember, 1,000 or 2,000 ads, the number one characteristic most highly correlated to better performance was set an ad in a living room. And you know, on its face, it's absurd.

00:19:38 Speaker_04
Like, am I supposed to write ads that are set in living rooms if I edit ads so that more living room footage is in there? Is it going to make sense? No, we had asked the interns a bad question and we got out a bad answer.

00:19:48 Speaker_05
As Ben says, part of the job of a marketer today involves figuring out how to work with the billion interns.

00:19:58 Speaker_04
I think the net of it is that it'll make the creative end of creativity better, more creative. It'll make marketing overall much more effective. It won't make all ads the same. I think it'll push in the other direction, actually.

00:20:11 Speaker_05
And this is where Avinash says marketers will continue to thrive, because marketers have always embraced the challenge of reaching their audience.

00:20:19 Speaker_03
There's nothing more fun than being able to figure out how to tell a story and change their minds. It's like an amazing challenge, right? Human beings don't like changing their minds. And I've come to realize it can't just be data.

00:20:34 Speaker_03
Then it becomes an even tougher challenge.

00:20:37 Speaker_05
Ultimately, machine learning is here to be a marketer's assistant.

00:20:42 Speaker_04
It can make a good ad better. It can't make a bad ad good, right? It's not rescue, it's tuning.

00:20:59 Speaker_05
The Think with Google podcast is brought to you by Google and Gimlet Creative. This episode was produced by Katie Shepard, Carrie-Anne Thomas, and Emily Shaw. Gabby Bulgarelli is our fact checker. We're edited by Andrea Bruce.

00:21:11 Speaker_05
Bumi Hidaka mixed this episode. Catherine Anderson is our technical director. Our theme is by Marcus Thorne Vigala. Additional music from Marmoset, Billy Libby, and So Wiley. You can find us on Spotify, Apple Podcasts, or wherever you listen.

00:21:27 Speaker_05
And if you like what you've heard, share it with your friends and colleagues. We'll see you next week.