Rob Baker - AI is for Everyone

Welcome to "AigoraCast", conversations with industry experts on how new technologies are transforming sensory and consumer science!

AigoraCast is available on Apple Podcasts, Stitcher, Google Podcasts, Spotify, PodCast Republic, Pandora, and Amazon Music. Remember to subscribe, and please leave a positive review if you like what you hear!

Rob Baker is a Section Head in the Data Science and Artificial Intelligence group at P&G. His group consists of Data Scientists and Data Engineers supporting R&D for all of P&G’s businesses. Rob holds a Master's degree in Statistics from Purdue and has 30 years of experience with consumer research methods.

Rob on LinkedIn

P&G jobs

Transcript (Semi-automated, forgive typos!)

John: So, Rob, I'm really happy to have you on the show.

Rob: Thanks John, I'm very happy to be here.

John: So, Rob, as someone who has been on the kind of quantitative side of consumer research for, as we just said, more than 30 years. I'd really love to hear your take on how things have evolved. You know how when you first came into, have you been in P&G for the full 30 years? Is that where you started off? Oh, wow. Okay, so maybe you could start up by talking about how is consumer research when you first start off in the field and then talk a little bit about the evolution. How you've seen things change, and that'll lead us into a discussion about the future.

Rob: Sure. Again, this was 30 years ago, so things were quite primitive. The tools were very primitive. The capabilities were very limited. There were limited numbers of people actually doing this kind of research. So basically things took a lot longer than they do today. And very none of it was obviously align at that time. So it was really, you know, a completely different world than today.

John: Yeah. I remember when I when I worked, actually, when I was in college in the 90's, I worked at the Institute for Perception, my dad's company, doing data analysis. And even then we would do some analysis. Put it on a floppy disk and put it in the mail. And then 3 days later, we would have a phone call with the client to talk about the results. So, like, can you talk? I am sure you love to do that yourself. So when did you start to see things really kind of change and take off in terms of the use of computers?

Rob: So I remember the days of the true in a five and a quarter inch floppies. So we're not even the three and a half or whatever they were. So, yeah, really they've just especially in the last five or 10 years, things have changed tremendously in terms of actually when any particular point. I can't say that there's any one particular point that I would say was that like an inflection point or something that really changed things. I think it's been a progression that is accelerating now with technologies that are being developed in other areas and being immediately reapplied in our spaces. It's been quite amazing and the future is even much brighter in this space. So.

John: Right. So, I mean, for me, what I kind of see and what I've seen since started, Aigora, is there's kind of two paths with these new technologies for you know, us in consumer research. One is kind of doing what we're already doing, but doing it faster, higher volume, faster turn, you know, quicker turnaround, that kind of thing. Sometimes it's called the fast caterpillar where, you know, you've got this idea of caterpillar and a butterfly. So fast caterpillar is kind of the automation path. I see. And then there's this idea that there's something qualitatively new that we can do and some people might call that artificial intelligence. The idea that there's some emerging there's some emerging capabilities that we now have that we don't have. So I know you have some thoughts on automation and consumer research, so maybe you can share with our listeners how you see automation coming into use in kind of everyday practice at P&G or just more general.

Rob: Much like you said, I see it the same way. There really are two different facets here. One is automation of what we do today. So to make it a lot more efficient and faster. So along the lines of better, faster, cheaper kind of things. But the much more, I think aspirational and more exciting part for me is, you know, the area of artificial intelligence where, you know, things that were never even possible or even dreamed about, you know, just a few years ago now are readily available to anybody that takes a little bit of time to learn the tools and gain access. And again, even the learning to be able to explore those tools is available online for free in many cases. So it's just a step change in that regard. I would say with a number of different AI technologies.

John: And can you give us an example maybe of some new things that you're seeing, some new capabilities that we wouldn't have had, say, 10 years ago or 5 years ago?

Rob: Yeah. Let me just tell you a story, which for me kind of summarizes a lot of where we are. So probably about a dozen years ago or so, I was on the SPSS Customer Advisory Board and I was at a meeting in Chicago. And you know, again, this about 12 years ago. So the first iPhone really had just come out. Feature phones were far more popular than smartphones. And over lunch, I was seated with the CEO of SPSS and we had been spending time on improvements to current tools, but nothing super, super huge. So I asked them, you know, what do you see as the future of consumer research in about 10 years? And he quickly reach into his pocket, pulled out his smartphone. There really weren't that many available at the time. Instead, it's image and video. And keep in mind also that YouTube was just a year too old at that time. So all of us at the table were a little bit surprised. Today, it seems like, well, I should have been obvious, but, you know, 10-12 years ago, these were not routine. It wasn't obvious really to many people at all. So I would have to say that image and video is one of the largest changes I've seen not only in consumer research, but in research general. So there's tons of examples of how we can use those.

John: Okay. So if we give us an example, maybe you have...

Rob: Yeah, One thing to keep in mind is even unsolicited from companies, consumers now will actually post images or videos of using our products on social media. The bottom is part of your rating and reviews. They will sometimes send them directly to companies. But let me just share just a couple of different examples that I've seen first-hand. So, in consumer research in particular, categories like, let's say in oral care, think of like floss, you know what percent of people actually floss. Well, when we do consumer research, it's important that we have actual consumers, you know, actual users of the products enrolled in our research. If not, they can't give us any kind of meaningful input if they don't use it. However, you know, there's motivations for some of these folks to say, yes, I actually, you know, I am a user so that they can be involved in the research, but not necessarily so they can get, you know, paid for the research. But they may or may not actually be floss users. So the idea is, and this is something we've been sued in several different areas, is to actually confirm that they are indeed users of the product. We actually have them go take a picture of the product in their house and upload it. And we develop the algorithms to confirm that this is indeed a floss and it is indeed.

John: Alright. To say you don't even have humans looking at those pictures, you just...

Rob: So the algorithms automatically detect. We've trained these algorithms on thousands of images to be able to distinguish floss from something else. And people are pretty creative. So one thing they do some task is they'll just go to Amazon. Copy and paste an image from Amazon. So this is what I use. So we have the algorithms also are smart enough to be able to detect that. And I've got to you know, we've got a very strong group here in terms of being able to to apply these kinds of methods. So we know real time we won't even allow them in the research unless they can upload an image of a product taken in their home, or at least not one that they grabbed from the Internet. And that can make a huge difference in terms of the quality of the, you know, the research and the results in some categories like maybe, you know, toothpaste. Well, pretty much everybody's got a toothpaste in the household. But some of these other categories that can make a big difference in terms of what you get. And so that's just one real, you know, concrete one, I'd say.

John: They have to prove they have children, hold up a picture of a child.

Rob: Well, you're right. I mean, so those methods could be expanded to anything that you need them to be expanded to. Again, you know, you just have to have enough images to do the training and that kind of stuff. So let me give you a one more a little more maybe aspirational example in some ways. And this is from our it's an app that we developed called Olay Skin Advisor, which uses AI to deliver a smart skin analysis and personalized product recommendation. So the way it works is that consumers basically take a selfie of their face. Our algorithms run in the background again, pretty much real time, and they give you a prediction of your skin age. Now, that prediction is is developed by tens of thousands of images that we have already evaluated. You know, that provided input for our model. Then identifies areas of the face that are most impactful to the age prediction. So maybe it's under the eyes. Maybe it's the crow's feet area. That kind of stuff. But then products are recommended that address those particular areas. So this is great for the consumer because they get their product recommendations specifically for them. It's great for us as developers because then we can learn more to try to meet her needs with better products and more targeted products or so. That's probably a little higher order. And that one took a little bit more effort for sure.

John: Was that a neural network that you train for that?

Rob: Yes. Yeah. So when I mentioned I want to talk a little bit about artificial intelligence, it really in many cases what we're talking about is so when you think of the high levels, artificial intelligence, a subset of that is machine learning and a subset of machine learning is deep learning. And the deep learning models really have only been available for last. Well, in terms of the computation ability really is only been around for, let's say, 10 years or so to be able to fit these things pretty know pretty easily with, you know, computational power that's available to us. A lot of the success in AI is due to deep learning, very much in particular.

John: Anything, that's perceptual, basically, like anything that would be like voice.

Rob: So there's a lot of different applications of deep learning. Basically, the idea behind AI and deep learning is many of these kinds of situations is to do things that humans can do. Like you, I recognize something from an image, for example, or like I said, translation from one language to another. Or just think of how many devices. When you think of, you know, really how many different devices you have in your house that probably do things. You know you've got your smartphone, you've got you maybe maybe have an Alexa device or Google home device.

John: It's always wrong one talking to I'm in the kitchen and I'm speaking to Alexa in the kitchen and the one from the other room is the one that answers. So they got to figure that out.

Rob: It's bad for me too because I have one room that has both a Google home and an Alexa device. So it's kind of it's like probably overkill, but it's so convenient. No matter where you are, you have access to this. So, I mean, really deep learning is behind all that. Even the voice to text, even like Google searches. You know, they are Google searches are so good, they're using deep learning, you know, very heavily. And another thing that I think is another great illustration of the value of deep learning is, you know, if you're using your smartphone and you're typing an email or does even typing a text message, it has recommendations. I don't know if you've noticed, but for me, they've gotten so much better in the last year or two to be able to predict what you sometimes the auto-fill is exactly what you're going to say. So basically with a fair amount of data which has been collected over time and these algorithms, you know, they can do a fantastic job for a lot of these kinds of things that traditionally had required humans to do.

John: Now, rather than just kind of turn down on a little bit of that, and obviously, if you ever feel like you, there are things you can't say because, you know, proprietary information, please just let me know. But the thing that I think has been most interesting from what you said so far is that P&G is actually training its own neural networks. And that's something that I have in my experience. I've found that generally speaking, in sensory consumer science, most companies don't have that luxury. They don't really have enough data. Would you say that that's that's true or do you think that it isn't actually? Do you think that other companies, it is within their reach to train their on their own networks.

Rob: Yeah, I think it is. So here's the thing. There's certain applications that you need very little data. Some applications that are unique to a particular company might need a lot of data. So there's a whole area called transfer learning, which is basically the concept of if a model is already developed. That's not exactly your application, but similar. And they had a lot of data to develop that model. Then you can essentially reapply the key elements of that bigger model and tailor it with a smaller amount of data unique to, you know, to your application and have very successful results. That's very common now. So with transfer learning, really, in many cases, there's something similar that you can reapply and and it is very much available. Well, you know, one thing that really has allowed that is really cloud computing. And for folks, you know, everybody's heard about it. Some people know more than others, but basically it's on-demand access to computer services that somebody else manages. So there are many different cloud providers. But Amazon, Microsoft and Google are the biggest. Now, one of the really exciting things, I think about those three and somewhat others as well, is that they really want you to use their services. So what they do, they all have extensive data science or artificial intelligence groups. So they bring to you. They develop their leading edge developers of some of these deep learning methods in other advanced machine learning methods. And they make them available to you, you know, to encourage you to use their cloud services. So they make the code available broadly, but they put those capabilities right adjacent to their cloud storage and everything. So think of things like auto machine learning, AutoML. Amazon, they have a product called SageMaker, which does a similar kind of thing.

John: And maybe for our listeners who aren't familiar with AutoML, maybe you can just briefly describe.

Rob: So really what it is, it's the name tries to describe. So automating the machine learning prrocess, so AutoML. So really what it is, if you are using their services and to some extent, even if you aren't there are ways. But you can actually just very directly reapply models that they have already developed. And again, using transfer learning, you may or may not need any of your own data or much of your own data. So really, what used to require you to buy hardware and have people to maintain everything and you're now in your company or in your office there? Now, you can just very quickly just go to one of these big providers and actually get whatever resources need. And if you need a bunch for a month, well, it's no problem. Just you know, it's elastic. You know, you can use more for a month. But then if you know that certain months, you don't need very much, well, then you just scale it back. So you're right sizing the, you know, the resources at a very very low cost. And it's being maintained by somebody whose job is to do that at scale. So you feel confident that it's being done well.

John: And do they support the transfer learning? I mean, do they have neural networks that are trained up where you can take and tap few layers off?

Rob: Yeah. That's exactly what they do. Typically like based that just like one layer. No, you would be training for your unique application. But that's exactly it. So you could you know. Yeah. It just never has there been a time when those powerful algorithms can be brought to bear on problems by people who really don't have as much, you know, that much experience or expertise with this. They're made available to quite user friendly. Because what they're really trying to do is to try to encourage you to be using their other cloud services.

John: So even a medium-sized company could potentially use AWS or as you have access to these models.

Rob: Absolutely

John: Conduct, transfer learning and with...

Rob: As you say, even more small ones, you know, with a little bit of outside help. You know, there are certainly companies that are quite capable to kind of help set you up with these different services and get you get you started if you're so inclined. But it's a great opportunity. I truly think that, you know, eventually these you know, this will be just part of doing business, that if you're not doing these kinds of things, you probably will not be able to keep up with your competitors in a lot of these areas just because they're so powerful. Yeah, I think they're going to be required.

John: Yeah, I agree with you completely, and I think that it really is. I mean, there's a message here just already. It's companies shouldn't imagine that this is only the domain and very large, like well-funded groups that even small or medium-sized groups can't get involved in this kind of machine learning revolution, especially even the deep learning.

Rob: Absolutely. And it's easy to learn about them, too. I mean, you've got all these different platforms. You could even go to YouTube and find a lot of different training. But, you know, some of these different, you know, online courses like, you know, Coursera, Udacity, some of these different platforms that have all kinds of trainings. So on each of these different you know, if you're you want to use, you know, GCP, you know, Google's version and the cloud platform, you can use that. There's training on Coursera, extensive training on Coursera. If you're, you know, more of a Microsoft kind of person than there is, your training is available, you know, readily as well. So it's easy to get started as well.

John: Now, that would be, I would say, for data scientists and engineers. Would you say it's also true for sensory scientists that they should be getting their feet wet a little bit?

Rob: I think they should definitely be aware of really what these different capabilities are, you know, just try. So I think a lot of people would be amazed that how easy it is to be able to fit some of these models. Some of them are made so user-friendly that you could just do it in an afternoon. Now, again, it's not going to be specific to your business, but they, you know, intentionally have made a lot of these things, you know, super super simple so that people get a sense of what they can do and then they can start thinking, "Okay, what does that mean how can I leverages my sensory kind of role? How can I leverages my consumer research role?"

John: Okay. Well, that's really good, Rob. Yeah, I mean, I think that. People who may really need to ask themselves their wishlist right now. Now is a good time for people to ask themselves if I could do anything, what would I do? Because there may be tools out there to support it. You were talking about holding up in a floss, making sure that your panelists really have floss. That they are, in fact floss users, right? I've been working something I've gotten involved in more recently, Alexa administered surveys. I don't know if this is something you are getting involved in or fear at liberty to say, but there's tremendous advantages to having a smart speaker inside someone's home while they're evaluating a product and allowing them in real time to give you their opinion of the product. Right? And this is another example where, like you said, AWS or Amazon makes it fairly easy to design Alexis skills. These are the sorts of things that can be.

Rob: With Lex and Polly. So we definitely do use those kinds of assistance for interactive consumer research. It can go both ways as well as, you know, collecting data as well as providing feedback. I mean, there's other ways as well. You know, in terms of there's the equivalent of Lex and Polly in these other platforms as well. But, yeah, exactly. That's exactly the kind of thing.

John: And then you get time. Yeah.

Rob: No, I'm just going to say, I mean, in so we the first skills we built for Alexa, we're probably within six months after it had come out because Amazon was quick to actually put out their skill development kit basically.

John: That's something I like speaking of Amazon and just to kind of get your opinion on this, it seems to me that as Amazon moves into the consumer goods area. They have a huge advantage over a company like P&G because they have all the sales data, because they've all these people. Huge numbers of people buying things. They can see what's selling and they can make recommendations. What steps would you say? What leverages does P&G have available to it? That Amazon doesn't.

Rob: Yeah. So I think in terms of AI generally whoever has data really has the, you know, a leg up on competition basically. Data actually informs these models. So I think in general, other forms of data that might be available to a company like P&G, you know company like us, really are the competitive advantage we have in terms of over, you know, ones that are, you know, companies that are more just directed at, you know, sales. So, you know, we do extensive consumer research and we have a lot of insights there. I'd say that wouldn't necessarily be obvious to other people that are other organizations that are more just on the, you know, the purchase kind of side.

John: Right. There's always a complete behavioral side. Whereas you have deeper information than just what you thought you have. Yeah. And what are you able to talk a little bit about how you are accessing your historical data? That's something that is very interesting to me right now, is the use of graph databases to take all of the sensory data that a company has organized it and then start to go through it to conduct queries or search for patterns, anomalies. This kind of thing to either try to answer research questions without having to field a study or conversely have the database suggest research questions that might be interesting based on things that maybe when you put all your data together, you see things that you couldn't see from just one study by itself. Is that something you have experience that we can talk about?

Rob: Yes. So we clearly do that in terms of, you know, doing meta-analysis across the research. One thing that I'm really excited about for automation is that it is forcing us to standardize our data in a way that it's integritable because, you know, if you look historically you know it just it's very hard to say. Okay. Was that the same question? Was that the same product? Because people call them different. You know, just the data. Those data aspects, you know, can be really complex and could be barriers to doing this automation, you know, forces a close look and consideration of those data elements in enabling and facilitating meta-analysis on the back end. So that's exactly. Yes. We do a fair amount of it, not enough because of data issues. And I think people don't necessarily understand how critical that, you know, that data engineering role. That's why, you know, a large part of my group is more on the data engineering side, which is ensuring these data flows or just even just individual data sets are going to be able to be leveraged efficiently in the way that we ultimately need them to be.

John: So, yeah. I spent a fair amount of my time actually trying to, I'm sorry I have to say this, rectify the past sins. So we got about five minutes left here. Rob, one thing I definitely want to ask you about, before we have to jump off the call is I know you were just recently add an AI accelerator summit, is that correct?

Rob: Yeah, actually. When you think about some of the different applications of, you know, for folks who are familiar with the Internet of things, basically where you have connected devices within your home, like Alexa, for example, or maybe your you know, your thermostat actually is connected or maybe you have a video doorbell basically with all those things being connected. A lot of times there are sensors of different types involved as well. Temperature sensors motion many different kinds of sensors. One area that I've seen AI starting to take off more is more behavioral kind of research where you're observing behaviors. Again, this is research. So people are, you know, are volunteering for this research and they get you know, you're paid to do the research. But basically, when you have you know, you can think of different areas in the household, you have a pretty comprehensive evaluations of what's going on. So, for example, you know, if this take, for example, you know, the kitchen, if you're into, you know, hand dish, you want to actually understand exactly how well people doing. So you want know what products are using when, what time, what are the conditions under which are they using them? It just gives you good insights into consumers. Well, the whole area of doing computing locally that is in that household so that you don't have to, for example, stream images or videos which include people's face which, you know, obviously consumers get a little sensitive about that, there's PII, if that data were intercepted by somebody, so edge processing really is a way to do some very comprehensive, including advanced machine learning. Right there on a small device, right in the household, so that all you need to send is a limited amount of data, including whatever you've captured from the you know, from the image or video. So what I confess was talking about many different uses of edge computing, which are many are, so, for example, Uber and Lyft were doing a lot with the self-driving cars because obviously you can't have a lag time between you know, you need immediate computation that's getting more and more available as well a lot of work going on there. And again, we'll be leveraging that even more in this hap's practices kind of space. So I see that happening.

John: Right. Right. Yeah, Okay. And any last words like what would you recommend to, you know, sensory and consumer scientists kind of going forward in the next five years? You know, in order for us to do well like the fourth industrial revolution is underway. So what do I need to do to make sure for doing well.

Rob: From my viewpoint, I would start looking at some of these case studies and ask myself, okay, which ones are most kind of relevant to me or which ones are similar enough to the kind of work that I can do and then penetrate a little further to say, OK, well, how can I get involved in that? Because it may well be the case that you have people around you that have capabilities to be able to access these tools. A lot of them are actually easier than you think. And if not, like you mentioned, there are certainly companies that, you know, for a modest fee can really help get you going in that space and get you set up. So I think awareness, first of all, of these things, because many people are not aware and I see for these things that have been around for several years now and it's awareness is the most important thing. Then just start, you know, looking. I personally love online courses, which kind of tell you a little bit about how you can actually use AWS or Google cloud platform or, you know, any of these different cloud resources for your particular applications.

John: Right. Right. Yeah. That really resonates with me as well. And to be aware that even if the first off is worth trying. I think to see what you can do yourself. But even if you're not going to ultimately be the one building the models, at least be aware that maybe people near you, maybe someone in I.T. has the ability to go and learn from these resources you've mentioned and then greatly upscaled the whole company in terms of the company's capability.

Rob: Absolutely. I mean, I think people would be shocked. You know, we we do actually in our area, we did a hackathon that we opened up pretty broadly within our organization. And the winners actually, you know, weren't necessarily in the organizations you might think they were from. So you might be surprised that the capabilities and interests and passions of people that have somewhat different roles, but that can very naturally apply these methods to benefit your company.

John: Right. So don't be afraid.

Rob: Just try and jump in. Yep, absolutely.

John: Yeah, that's great. Alright. Well, on that on that note of perfect agreement, because I completely agree with that, I'd like to say thank you. And if people want to find you. I suppose that someone listens this is inspired to apply to work for P&G, how would they get in touch with you? And what would be the right avenues for that?

Rob: So they could just send me an email on baker.ra@pg.com or look for a different, we have typically have a number of different openings in this space, both in our area, which is R&D or more on the commercial side or the production side and that's pg.com for careers.

John: Yes. Okay. And are you on LinkedIn also? Are you active there? Can people connect with you on LinkedIn? So it's just Rob Baker on LinkedIn. Yeah. Okay, we'll put the links and the kind of notes for this call so alright, Rob. Thank you so much and appreciate your time. And I look forward to talking to you in person in the near future at the next ASDM.

Rob: Thank you, John. Nice talking to you.