In this week’s episode of In-Ear Insights, join cofounders Katie Robbert and Christopher Penn as they tackle the basics of machine learning and artificial intelligence.

  • What is machine learning?
  • How is it different from artificial intelligence?
  • What are good and bad use cases of machine learning?

Listen in as they tackle these questions and many more.

Download the MP3 audio here.

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for tuning in.

Christopher Penn
This is In-Ear Insights, the Trust Insights podcast.

In today’s episode of In-Ear Insights, the Trust Insights podcast, we are talking all things machine learning and machine learning. What a one, it is one of the hottest spaces the hottest discussion topic, machine learning artificial intelligence, everybody is saying they do it. Everybody is saying they have it, companies are being bought at ridiculous valuations. So let’s dig into this. Where do you Where do we want to get started with this discussion,

Katie Robbert
we should start with at a high level defining what is machine learning, and then sort of dig into the differences between machine learning artificial intelligence and look where everything fits in. Because I think there’s a lot of confusion about even what it is. And then when we think about the marketing space, where does it fit? How is it being used, and where do I get started

Christopher Penn
machine learning, I think that that’s the best place to start machine learning is

if you think about traditional software, you write the code, and then the software spits out data. So you fire up Microsoft Word, right? There is code written, you interact with the code, it spits out a result, right, a lovely 20 page document, like a term paper, all those term papers we loved writing in college, that’s traditional software, machine learning is the opposite. You feed machine learning software data, it right, its own code, and then it spits out new data and new results once it’s done. So it’s like a, it’s like a food processes that can learn from the foods you give it. And then and, you know, process them in new and different ways. Yeah, as appropriate, as other was best analogy. But it is it is for data sets that, frankly, are too large for people to be able to write static code against or four days, it’s a change so much that again, people, human beings just can’t scale to that level of data, at least not without like devoting our entire city working on one problem.

Katie Robbert
So are you saying that if you feed a large data set to machine, a machine learning algorithm, or a piece of software, that it will tell you what the data is saying? Or do you have to tell the machine learning algorithm what kind of output you’re after or both,

Christopher Penn
it’s a bit of both. So if you were to look in an interface, really good example of one is a IBM Watson studio, if you look into its machine learning interface, particularly the neural model, or there’s a, there’s a little Dropbox you drop in that says, here’s the data I’m putting in. And then there’s sort of a set of boxes, drag and drop boxes, you can assemble for, like, will digital Legos that make the data move in certain ways, and then have a target outcome of some kind, it could be a report, it could be a table, it could be a reduced data set, it could be a summary, and so on, and so forth. And what’s happening is, in those cases, boxes in the middle, that’s where Watson is writing his own code. So it could be doing soft Max is it could be doing eigenvectors can be doing all kinds of mathematical operations, you as the human don’t need to write any code you as a human architect, dude, you’re the conductor of the orchestra, you do need to know which boxes you want. and in what order in the same way that a conductor of the orchestra doesn’t necessarily need to be an expert, violin player and an ex ex, what cello player but they do need to know when those those talents are used. And so it’s sort of a human conducted machine, piece of software that takes the data in process, it transforms and gives you a human usable output.

Katie Robbert
So in that example of the conductor, the conductor needs to know at least what kind of music is being played. So the same thing with machine learning, you need to understand at least what kind of questions you’re trying to answer, even if you don’t know what the answer is. So, you know, if I couldn’t just be a piece of data to machine learning and say, Tell me what it says I would need to have to give it some sort of direction. Just like as a conductor, I wouldn’t say play something and then be mad when they played jazz when I was really hoping for classical kind of a I guess.

Christopher Penn
Yeah, it depends.

There’s two broad types types of machine when this supercapacitor machine learning where are you saying machine, I want this outcome, tell me how to get to this outcome. And then is unsupervised machine Tell me what’s in the box. So a really good example of supervised machine learning is,

as a marketing application will be SEO, tell me out of all of these bought, you know, pages of text, what keywords are most closely related to analytics, for example, and you would use all these different machine learning techniques, like vector ization and LGA and all that stuff to tell you these are the words that most closely related to the word analytics. And so you might see like Google Analytics and marketing analytics and that you would you would build your SEO strategy based on that. So that’s supervised, unsupervised is when you ask the machine to categorize data that you can’t make sense of so in that same example say I download blast thousand blog posts from the marketing crops website and I say machine Tell me what the top keywords are in this I don’t know I don’t know what they are and it will go through and you can use rake Oregon vector ization and unsupervised mode to say okay, so b2b is a top keyword in the marketing process, website and email marketing as a top clustered phrase and but influencer marketing isn’t. And you would look at this, these clusters that the machine has come up and spit out, and then you can say, Okay, well, now, now that we know what’s in the box, is that what we want? So this kind of interesting cycle of like, just like in serving, there’s qualitative and quantitative in machine learning, that can be a cycle of quantity of supervised and unsupervised. What’s in the box? Is it valuable? If it is, let’s do more of that, and then go back and looking at is now what else is valuable? And what’s not

Katie Robbert
in a roundabout way? You did answer my initial question that you have to give the machine learning algorithms, some sense of direction. So in both of those examples of supervised and unsupervised you use SEO. And in both examples, you told the software, here’s the direction I’m pointing you in. And so for the queue, so for the supervisor, it sort of told you all the conducted keywords, and then in the unsupervised you said, I’m looking for XYZ. So what I’m getting from that is that you still need to know roughly what you’re after, even if you don’t know where it is, how to find out what it’s going to be. So what I what I’m interested in is knowing sort of this marketing space, this ever evolving industry with a move towards technology, whereas machine learning already being used in where are some places that would benefit from more machine learning,

Christopher Penn
every place could benefit for more machine learning. So

Katie Robbert
think about it in terms of where to start.

Christopher Penn
So let’s, let’s, let’s actually start at the end, what are the benefits that you want? Right? So generally speaking, we’re going to want three things, right? We want as better, faster, cheaper, right? That’s what humans always want better, faster, cheaper,

when you look at acceleration, that’s one thing, machine learning does really well, let’s get let’s do the thing faster. Instead of having, you know, an intern, googling for results? And like, Oh, these are the top pages for the search term Google Analytics. So have the machine do it, we’ll get it, we’ll get it done much, much faster than the human accuracy is the second, right, again, same intern googling they may they, they they get partying too late at the Red Sox game, and other little under the weather. And they make a lot of mistakes, machines, properly trained, will come up with a much better answer like these are the terms are related to to Google Analytics, or these are the top photos on Instagram, or these are the things people said about your product on G to crowd and the third one is automation, what stuff isn’t worth having somebody do

you know, we are we are members of a startup company, there are only so many hours in the day, what are the things that are not worth our time but have to get done and are generally highly repetitive. Those are things that are candidates for machine learning. So machine learning is really good at faster, better and, and better use of time. Or cheaper, I guess is is the consumer way of putting it. When you look at marketing, how much of marketing is repetitive? Oh, God, we’re loading up social media posts again, oh, gotta write another blog post about this again,

every time you find yourself doing something repetitive, you should be thinking how can I get a machine to do this

Katie Robbert
without getting too into the weeds? With the answer?

Christopher Penn
I think that there’s a lot of confusion between things like machine learning artificial intelligence, robotic process automation, because they’re all very similar in in a way, but can you give us at a high level what the differences are between at least machine learning and artificial intelligence because my understanding and maybe I’m wrong is that robotic Process Automation is part of each of those things that it can stand alone. But it’s really part of both robotic Process Automation thing is a cousin because in a lot of cases for our PA, you’re still you may not be writing physical code but you are still recording macros essentially that are then replayed, there’s no learning involved in in hitting the play button on a macro, right. So it’s related. But you could use that same data to train machine learning software to develop new or better a fast ones. But you’re the artificial intelligence question is a good one. Because that is that is a hotly debated semantic topic. So artificial intelligence is the process of teaching a computer any aspect of human intelligence. So if you’re listening to this episode, and you can distinguish the words that I’m saying from noise, right, which means they have meaning you’re doing what’s called language processing. Now, you learn this as a as a baby, you know, between zero and five years old in a lot of cases, and obviously improves over time until you become a teenager. But generally speaking, that’s a that’s a cognitive process that is very difficult for a living organism to do. And even more difficult for machine to learn. Machine learning is a subset of AI is a subset that is that is about helping machines learn through data. There are other as all aspects of all aspects of machine and learning artificial intelligence, not necessarily all aspects of artificial intelligence or machine learning, but it is they are functioning from a marketer’s perspective. They’re functionally interchangeable terms, they are small distinguishing characteristics on the mathematic side.

Katie Robbert
But to that point, as a marketer, it’s not that it doesn’t matter which one you’re using, it matters that you’re using something to help you do your job more efficiently and make less errors. So let’s talk about some examples where there’s a very concrete examples of where marketers can use machine learning what’s available to them today that perhaps maybe they’re not using?

Christopher Penn
That’s a loaded question. And the reason is a loaded question is because it depends so heavily on your skill level, almost every machine learning technology is available to everyone for free of financial costs. Because open source code is incredible. And the community has done an amazing job of making it available. But that’s like saying, anyone can buy surgical instruments off of Amazon,

you probably should not be doing open heart surgery unless you’re qualified surgeon.

Katie Robbert
So then let me reframe the question then, yeah, let’s say I am a mid level few years of experience behind me perhaps some a marketing manager, where can I start with machine learning, what can I start to use

Christopher Penn
if you’re a mid level marketer and you’ve got and you know, the processes, you know, what it is you’re trying to, you probably know how the work is done, the where the where to start with machine learning to me is looking at the existing vendors you’ve already got and seeing are their vendors that have machine learning technology built into their products that will do a better job than the services you’re currently using. So real simple example. Suppose you’re doing social media monitoring, and you’re monitoring for your brand name. And your brand name is something very popular number. years ago, we worked with a company called tango tango is a messaging app, but is also a very popular style of dance you in in the old world, one of the things you would do is you would write up of noxious long queries to try and isolate your brand name contextually from people talking about dancing, very difficult to do. And it takes you a very, very long time to do that you instead might look for a machine learning vendor that says, Let me train we provide some training data to the tool that says this is these are the kinds of social posts I do care about these kinds I don’t care about and have the vendors machine learning software, essentially learn from that. And then over time, it will get better at identifying those mentions, even without your training, you’ll you’ll continue to quality check will say, Yep, this is this is actually about Tango, nope, this is not about Tango are our partner talk Walker has that built into a software. And so it’s very, very good. Most of the time, a marketer is not going to be doing the implementation of machine learning themselves that is in but in much the same way that you can learn first aid, you probably should not learn open heart surgery unless you are going to become a surge.

But you should be looking at your vendors. And you should be looking at a lot of consumer technologies that may be way assistive in some of your tasks. So you and I both put up a ton of social media posts, we both take a ton of photos, one of the simplest ways to manage your company’s photo libraries stuff, all your photos into Google Photos account, it’s machine learning, its deep learning will categorize and tag your photos and understand what’s in them. So that when you need a photo for a blog post, like two people eating pie will spit out the appropriate images that you’ve taken. And you can then manage your your own inventory much, much more easily.

Katie Robbert
So I want to go back to something that you mentioned that it was it, we went by it so quick that I want to go back to the topic of this notion of quality control. So I think that one of the other misunderstandings about machine learning and artificial intelligence is that it’s a set it and forget it thing, you can’t just put something into an algorithm and hope for the best and say, well, the machines are taking over, and they’re going to figure it out. So in your example of putting your photos into Google Photos, and the machine learning starts to categorize everything, the user still needs to go in and do that quality control, because there is a good chance that some of your photos are categorized incorrectly. So can you speak a little bit about

the need for that human quality control with all of this new emerging technology, specifically, machine learning with again, without getting too in the weeds, because we could probably talk forever about this topic,

Christopher Penn
there’s two things go wrong with machine learning broadly. One is the algorithm trains incorrectly, like you’re mentioning photos are tagged incorrectly, and stuff like that. And, and, and that is a problem. And, and that is what IBM calls human in the loop someone quality checking the algorithm and saying, yes, this is what we want to know is what we don’t want. Amazon just had a very public failure of this with their HR recruiting, because they tried to automate it. But they the algorithm was heavily biased against women than a human in the loop didn’t catch it in time. But the bigger problem in machine learning is that the data you’re putting in isn’t correct, right. And so the machine will, machine learning means the machine learning from the data we provide it, the data is crap, garbage in, garbage out,

in your example of photos. Suppose and this is gonna be a very humorous example, suppose you accidentally had some adult material on your computer, and it was ingested by Google Photos, guess what, Google doesn’t know the difference, its algorithm doesn’t know that, if it’ll start categorizing that and scanning that. And then when you go to pull up, you know, pictures of someone who looks like this, this would be a hope you’ll have a photo of, you know, here’s a photo of your CEO, and it’s like, Whoa, where’d all that other this stuff from, like, well, your data input was wrong, your data training data set was was incorrect. And as a result, the algorithm learning correctly now you’re pulling up, you know, embarrassing photos in front of your boss, which you should never do.

But this shows up in so many ways, we’ve had previous discussions about bias, that is where a lot of bias comes from, not from the algorithm, but from the data feed it and that is probably the most important thing for a marketer to realize is that what you go with goes into the machine drives what comes out in the same way that if you swear all the time in front of your kid, and then you like, why is my kids swearing? Well, what you trained your child on was what you say,

Katie Robbert
I actually had that conversation with a good friend of mine, she was very confused as to why her child had what we’ll just call the trash smells. And I looked at her and I said, you know, that the, your child got this from you. And so it was, it was an interesting, but I think that that’s a really good way for people to understand the concept, you know. And the other thing that I find interesting is that it always, regardless of what we’re talking about, whether it’s machine learning, artificial intelligence, robotic process automation, or anything else, it always comes back to the same place, you have to start with a solid foundation, you have to get your house in order, you have to have a plan, you have to have a goal those things regardless of where technology takes us, that’s a consistent that will never change. And whether you have some sort of technological assistance to get your house in order. And when I say that, I mean, your analytics infrastructure, cleaning up your data sets, making sure that you’re collecting data consistently, that’s not going to change regardless of where the technology goes. And I think that that’s a really good place for us to start to wrap up is machine learning is technology that can only benefit marketers in the long run, because you can have your job done better, faster, cheaper, those are some of the benefits, but you can’t start to do machine learning until you get your infrastructure straightened out. And I toss this to you on any additional thoughts on that the

Christopher Penn
biggest mistake people make, and it’s one that you alluded to earlier, is that machine learning as somehow magically just put everything in, you know, just drop off the box, the machines doorstep and it’ll come back with, you know, magical unicorn blessed results later on doesn’t work like that. Machine learning isn’t magic, it’s math, it is based on statistics and probability. Those are the fundamental underpinnings of all machine learning. Is this a picture of a hot dog or not? Is this a picture of a cat or not? Is this key word related or not? It’s all probability. So the data that you put in governs the the the training and governance the output of the machine. If your data is garbage, your machine learning will be garbage, there is no way to make that untrue them like you said, they will follow the follow the basis royole focus on what doesn’t change strategy. Having a strategy never changes, you should have one that we go gold never changes and what people want out of their marketing never changes either, right. On the b2b side, people always want to save money to make money you save time and then I get fired consumers always want better faster cheaper I everybody wants better faster, cheaper and so keeping those things in mind as your as your guideposts as your as your lanterns and the dark will make machine learning a lot more easier to understand and make your decisions about what technology more clear because you can focus on will does this tool or technology or data or outcomes doesn’t help get me closer to better faster cheaper or is it just a shiny object and that’s that’s really is the I guess and always the final word

Katie Robbert
if you the listener, want some help or have questions you can certainly call contact us at trust insights.ai. You know, we know a lot about machine learning. It’s one of Chris’s big passions in life. So give us a holler. Email us, tweet at us, Facebook us you know, whatever, whatever your platform of choices, you’ll probably find us on there. So

except facts well, you know if I need to set up my fax machine I will

but you know, it’s it’s doesn’t go along with this whole advanced technology thing. But yeah, give us a shout. We love to talk machine learning.

Christopher Penn
And as always, please subscribe to the first insights YouTube channel. Now with even more interesting tutorials and our newsletter at trusted sites.ai will talk to you next time.

Thank you for listening to In-Ear Insights, the Trust Insights podcast please ask a co worker or colleague to follow our show on Google podcasts. Apple podcast wherever you listen to your shows got a question like us to answer wants to help solving your data and analytics challenges. Visit us at W www. Trust insights.ai today.


Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, Data in the Headlights. Subscribe now for free; new issues every Wednesday!

Click here to subscribe now »

Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new 10-minute or less episodes every week.


Also published on Medium.

Pin It on Pinterest

Share This