In the latest episode of All Jargon Aside, our host, Graham Wilkinson connects with Ben Royce, Head of Performance Data Science at Google and Nik Bear Brown, Assistant Professor of Game AI, AI and Machine Learning at Northeastern University. Nik consults in computational biology at the Broad Institute of MIT and Harvard, and Ben works primarily with YouTube advertisers and uses artificial intelligence to find patterns that optimize creativity. In this interview, the three got together to discuss the state of Artificial Intelligence, bias in AI and the future of AI – both in and outside of the marketing ecosystem.

Since Ben and Nik couldn’t join us in person for a traditional podcast recording, we conducted this interview via the phone and have produced the following transcription. Please note that this interview has been edited and condensed for clarity.

Graham Wilkinson:

Our obsession with recreating human intelligence can be traced back as far as ancient Greece and ancient China, through medieval times, the Renaissance and right up to the modern day. These efforts to reproduce the key components of living things through the use of machines can be seen
as the precursor to attempts to create intelligence. The field of artificial intelligence, as it is today, has a great number of illustrious contributors from the fields of math, computer science, philosophy and neuroscience, just to name a few. Names like Leibnitz, Churring, Lovelace, Shannon and Newell were just some of the pioneers and key contributors.

So, with artificial intelligence having confidently moved from the realms of science fiction to reality, and with the ever-growing applications and awareness of this technology around the world, do we really know what this term means? And what are the implications for the advertising industry – both good and bad? I’d like to start by asking: what exactly is artificial intelligence?

Nik, are you able to break this down for us?

Nik Bear Brown:

Humans are generally good at a lot of things. I think AI, in its current state, can be really good at very specific tasks, like driving cars or doing taxes. I don’t particularly like to do taxes, so personally I think it’s great if I can automatically upload a little data and have AI complete a task for me.

AI as it currently stands, has a lot of very good algorithms to do very specific, often very mundane tasks. That said, we’re very far away
from a lot of people’s conception of AI. I think a lot of people see AI as terminator robot who acts just like humans and are comparing AI to human intelligence. I don’t think that’s where we are at all – right now AI is very powerful and has very specific applications.

Graham:

What’s particularly interesting about this, is that when people started pioneering in artificial intelligence, they started trying to solve problems like: how can we win at chess? These are problems that humans felt were really complex, but actually, we figured out quite quickly that winning a chess game is not such a difficult problem for a machine to solve. And actually, some of the most difficult problems are the things that we take for granted and do in very autonomous fashions. So, Ben how do you think about artificial intelligence?

Ben Royce:

I think of artificial intelligence as a broad umbrella term. A number of years back there was a big debate in the community about whether we should even call this technology “Artificial Intelligence” because it may have set expectations a little too far out. And I think that’s probably still true.

I look at AI as kind of big buzzword that encompasses everything we’re talking about, but if you narrow it down you get into practices like Machine Learning, where there are more specific technologies. But in general, what we’re trying to do is teach computers how to learn, not just follow rules.

Graham:

You brought up another term that’s used a lot in the same breath as artificial intelligence, and that’s Machine Learning. What is the distinction between AI and Machine Learning?

Nik:

Most people consider AI a subset of Machine Learning. But I like to think of the term “statistical learning” rather than Machine Learning.

Statistical learning means primarily using algorithms like linear regression, tree-based methods, etc. to create very powerful statistics to make predictions. That, with the advent of deep learning we’re no longer just developing a model, but rather showing a machine how to do something. So I would say AI is a subset of machine learning, but I don’t think people use the word statistical learning to separate using statistical approaches versus using other kinds of AI.

Graham:

So Ben, when you’re speaking to clients from a Google perspective, how do you distinguish between AI, Machine Learning and Deep Learning?

Ben:

We debate a lot about what language to use. Outside of the academic community, we decided that AI is our general term, even though there are nuances to that. When we talk about Machine Learning, we’re talking about cloud computing and machine learning APIs. For example, Google Assistant or the Google Home where there are inputs and outputs, you’re interacting with this machine versus some of the more traditional methods such as production or natural language processing.

We talk about machine learning, especially in the cloud space. Controlled vocabulary is very important in these spaces and sometimes the excitement has over overridden the official terminology that may have come out of research.

Graham:

You brought up, Google Home and Google Assistant, and if we add in something like Google Translate, how do you classify those technologies as being intelligent? Are they a narrow form of intelligence, simulated intelligence, instantiated intelligence?

Ben:

Compared to previous technologies these things are pretty intelligent, right? Are they as good as human intelligence? No. We’re not anywhere close to that because we don’t fully understand the human brain.

Google Assistant uses a combination of technologies that makes it work. That includes speech processing, language recognition, or accent recognition for example. It also has to understand the question and the demand, which requires different technologies. The thing that makes it feel intelligent, ironically, is the persona. Google Assistant is supposed to be like a librarian. She’s supposed to be kind of funny, a little sassy. And the funny thing is, while that makes her feel intelligent, a lot of
that design is not done by machines, it’s done by people. But it’s the combination that makes it feel like you’re talking to a person, or at least an iteration of someone.

Graham:

That kind of goes back to how I started off the interview talking about automatons and our obsession as human beings with ascribing intelligence as a human trait. As a consequence, we like to think of artificial intelligence ultimately culminating in in some sort of a humanoid body. From your perspective, Nik, that’s not how you think about artificial intelligence on a daily. Correct me if I’m wrong, but you aren’t building robots in your research and actually, artificial intelligence lives inside of a non-humanoid shell.

Nik:

Once you go down the route of “intelligence”, I think you’re entering a domain of philosophy. We have issues like defining machine learning versus statistical learning. And I’m just a very practical person – if there’s something that I’d be willing to pay a human being to do, like, translate something, do my taxes, drive a car, but a machine can do it, I consider that intelligent.

Graham:

It’s important to have this conversation about intelligence, whether it’s a human trait or not, because ultimately, we are going to talk about bias. Bias is something we think of as a human trait. Ben, can you talk about bias in machine learning models and deep learning?

Ben:

In deep learning, it’s important to understand that bias can happen and it’s not necessarily because of the technology. Bias comes from the data, almost all of the time. When we train a model on something the data is the raw ingredient. The training data has to be unbiased or it’s going to produce bias in the future.

For example, say you’re trying to train a deep learning model to identify shoes and you want this model to answer a simple yes or no question: is this a shoe or not? If the model is trained using only images of men’s shoes, and you’ve never input high heels or any kind of women’s shoe, then it won’t be able to identify a high heel. That would produce what is considered a bias towards male shoes.

That can have all sorts of repercussions, so it’s important to remember that the source of the training data is often the problem and that requires data stewardship to prevent that from happening. These systems aren’t one-and-done often and require a lot of upkeep. Updating your models and showing diverse examples will teach a machine to be very, very unbiased.

Graham:

Nik, I’m interested to know if this is something that you see in your daily research – do you find that data is the root of the bias? Or is there evidence to show that bias exists in the machine itself?

Nik:

I agree with Ben. Bias comes from the data, but it also comes from the humans labeling the data. A lot of these training data sets are labeled by human beings and there often isn’t diversity in the people looking at and labeling the data. This produces bias. What you have to understand is these algorithms, particularly the supervised algorithms, learn what you teach them and the human beings who are labeling this data are themselves biased. This means that unless you’re explicit about labeling something the machine won’t learn it.

The bias also comes from how you collect the data. In a classic example, there was a group using a modern interpretation technique and teaching a machine to discriminate between images of wolves and dogs. The machine did very well, with 99% accuracy, but when they used techniques after to reveal the bias, they found out all the computer was doing was identifying if there was snow in the picture. It just so happened all the wolves were shot outside way up in Canada, but none of the dogs were.

So bias creeps into the data a lot and there’s a big push in model interoperability techniques and other techniques to reveal the bias algorithmically, but you can’t really avoid having a bias because human beings themselves are biased and that creeps into the data.

Graham:

We’ve been talking about image recognition but what you described is an inherent problem with the technology and its limitations around creativity. Humans can act spontaneously, and that randomness is the same thing that allows us to think creatively. In advertising, there’s obviously a huge creative engine that is comprised of an army of creative human beings. Is there a future where we think machines will have the ability to creatively problem solve, or is that really something that will always be in the realm of a human being?

Ben:

I think the holy grail for creative in artificial intelligence would be to have a machine say: ‘I can create an ad, just tell me what you’re selling and how you want to sell it and I’ll build the ad,’ and we’re nowhere near
that right now. The targeting capabilities are incredible with artificial intelligence but on the creative side, I actually think it’s going to be
a long road to any sort automation. I don’t ever actually think it will
be totally automated because creativity is usually the mixing of two completely disparate things, which a machine can’t really do. On top of that, creativity is usually net new and unique and the definition of unique is very, very different from person to person. So, while I do spend all my time working with machine learning and creative, I’ve noticed that it’s better to, for example, do pattern or trend detection, ad targeting or segmentation versus creating a net new piece of art.

Graham:

Nik, have you done any work with creatives in your research of deep learning and machine learning?

Nik:

Yes, we work with a technology called GANs to automatically generate content and what that does is not necessarily creative, but we can literally generate millions of images and predict what your images are likely to be, but then you still need a human being to look at those. It’s more like if you were to give a creative person 100 designs, those 100 designs would give that creative person ideas.

Graham:

Ben, I’m keen to get your thoughts on how currently AI and ML is being used in the advertising space and where do you think it could be used more?

Ben:

One of our top engineers, Jeff Dean, said that one of the best use cases for AI is advertising, not necessarily for ad creation, but for finding the right audience and understanding their signals. For example, if someone is doing some research and trying to find something to buy, that’s the perfect time for advertising to come in, in a privacy safe way. To that end, we use AI for finding audience segments, we use it for detecting fraud, spam prevention, things like that.

I think the most underutilized way that we could use AI is in the insights generation process. You can go from data to information very easily with artificial intelligence and that could get rid of a lot of the work that advertisers do when looking for a great insight or something that they can latch on to and build a campaign with.

Nik:

I just saw an interesting thing with Converse where they’re doing a wedding line. So, for this project, I think Converse could utilize AI that would generate content that looks like this potential Converse wedding line. That way, Converse could literally tweak its designs without producing them and then see how people respond. When people respond, they could then go back to the insights they’re getting and make the shoes accordingly. I’m not seeing a lot of that yet, but it seems to me that there’s a huge opportunity just to throw stuff out there before it becomes a real need. And then, if people are responding, you could produce the most popular design or product fairly quickly given current manufacturing techniques.

Graham:

The way that I would interpret that is that it’s a more systematic approach to interpreting insights. But, to come back to my original point – can machines truly be creative? It gets into the argument that once you make something a system, it really isn’t that creative anymore you are just giving it more and more rules that it can potentially apply. So, I think for the time being, people in creative can rest assured that their jobs are safe and sound.

And before we go, I’d like to ask you both one more question. Where would you like to see AI or ML being used in advertising that it currently isn’t already being used? Nik?

Nik:

To expose how advertising is done for people who are consumers, that is for them to understand how those decisions are being made today in the market.

Graham:

Okay, and you Ben?

Ben:

I’m genuinely excited to see spokespeople that are developed. As in, their generated voices, their personalities and their faces. I’d love to see a digital spokesperson as you know, sportscaster or something like that.

Graham:

So…should I be worried that they’ll be some sort of AI powered podcast host to replace me in the future?

Ben:

I think with your personality we’ve got a long way to go before we get anywhere close to you.

Graham:

Thank you Nik and Ben for joining us.

For more episodes and updates on All Jargon Aside, follow Kinesso and our host, Graham Wilkinson on social!