Intro to Machine Learning w Sam Charrington - CTO Dose 074

Machine-generated from the episode audio and not hand-corrected, so names and technical terms may be imperfect. The audio is authoritative.

Hey, welcome to episode 73 of the CTO advisor podcast on the line I have my co-host as usual mark mark. How's it going? I'm doing all right. How about you? You know what I cannot Complain, it's Friday and I own my own business. So that means that that means absolutely nothing Another day at work tomorrow, so Sam we have Sam on the line Sam you own your own business, too Do you do you actually get a weekend, uh, you know, I do look forward to Friday's I think he's just like you bursted a bubble for me Keith, and I'm not sure I like it So as we're talking about businesses Sam help us out.

What business do you run and and and Help give the audience insight into exactly why we had you on to the podcast Yeah, so I am an independent industry analyst. I founded my firm cloud pull strategies in 11 so coming up on 11, so six seven years old now Focus pretty heavily nowadays on Trying to keep tabs on what's going on in the machine learning and AI Marketplaces, but over those a few years I've been I've you know followed cloud and big data pretty closely as well About a year and a half ago I started a podcast called this week in machine learning and AI which everyone should go check out It's Twiml AI comm and also, you know any place you want to find a podcast And that's been going really well.

We We bounce in and out of the top 40 Apple technology podcast lists all the time And we've had been you know, the beneficiaries of you know, some great guests and an awesome community And I tell folks all the time now like I don't know which is my day job You spend a lot of time we haven't got a full-time producer helping me out on the podcast Spend a lot of time on that and so fortunately you know as you noted running my own business and The podcast and the analyst stuff are very complimentary.

So it's good to do it all Yeah, that's actually pretty swimmingly as they say I started the CTO visor thinking that I'm going to do mainly strategy work and And subsidize it with content stuff Yes, trying to find the actual time to do strategy work Been a challenge so, you know surprise surprise and it's not like you fell into machine learning I think as long as I know you I've only known you for a couple years, but during that time period I've always kind of picture you as a big data and you know machine learning kind of goes hand-in-hand with quote-unquote big data It's not something that we separate it out Can you talk to kind of like the separation between?

Big data and then machine learning. Oh, yeah, I think It's true. It's been something that I've been looking at for a while. In fact At some point I went back and realized that the very first Article that I wrote that got published outside, you know, and I think it was on read, right? After I started the the firm was about machine learning and machine learning platforms and So it hasn't been something that I've been following for a while and and I think you know One of the things that has made, you know, what we're seeing around machine learning and AI so popular nowadays is Really comes out of the fact that you know, we've got so much compute readily available In part, you know due to cloud but in part just because computes gotten, you know cheap and ubiquitous We've got lots of data and as you know, you know machine learning is based on you know Creating it is based on training algorithms to identify patterns in data and That requires a lot of data particularly for the kind of things that we're most excited about now You know things like deep learning computer vision, etc So all of these things kind of go hand in hand So mark, let's talk a little bit about kind of the applicability of machine learning you run into much machine learning in your day-to-day You know, I hear a lot of vendors talk about machine learning and AI But every time I drill down by and large most of them Doesn't seem like what I would consider anything different than they did, you know ten years ago But there are a few that I run into where like, okay, I kind of get it But it all seems pretty early on for a lot of people I think the people I hear it most from when it comes to vendors are security vendors Sam.

What was you do? This is your day-to-day. So Bring us kind of into your world. Where what do you see the most use cases for? machine learning like in the real world tons of applications pretty much look at a you know, look at a An org chart or look at a you know a breakdown of kind of the you know The an organization's business processes and if the organization has enough scale to generate enough data chances are you can identify? pretty interesting machine learning problems One of the ways that I tend to think about that is, you know in any organization of any scale, you know There's going to be some you know Big massive spreadsheet that the organization is kind of running things on and you know Often there's a cell or some number of cells in that spreadsheet where you know, people are really guessing, you know Whether that's forecasting guessing the percentage Likelihood that a deal is going to close for example, you know Whenever you've got that thing where people are that cell where people are guessing that's an interesting machine learning problem Potentially.

Yeah when I was a management consultant, I actually just posted this on Twitter a little bit ago or LinkedIn And it's amazing how much time that we spent their really big spreadsheets and trying to glean pieces of data to either do predictions or some analysis and a lot of you know a Lot of man-hours for how you pay really smart consultants spent a lot of time just doing stuff You know as now that I look back of it back at it, you know There was an algorithm for some of that stuff.

It could have really Really saved us a lot of work and kind of up level the work of a bit mark What some of the areas kind of your day-to-day that you wish that you know, what machine learning could really? Benefit your you know, not just professional life, but personal life. Oh You know, especially around the house when it comes to home automation when we're talking personal life, you know If it could predict what I'm going to do better That's obviously great If I didn't have to tell it to record Game of Thrones if it just knew I would want to watch Game of Thrones You know because I'm me and it should know me You know, I think Back to the professional stuff.

I think you know being in the insurance industry. I think there's a ton of stuff we're already doing with a ton of data, you know, whether it's crop insurance or we take GIS data and you know all weather data and we do something with it. That's something you know could easily Improve, you know with machine learning. I mean, I think it is already kind of starting to but We're just very much on the fringes of that still well, in fact, I venture to guess that even in that example you use the The weather you've got this raw weather data but ultimately you want to predict what the weather is going to be at some future point in time in some future place and you're There's probably already machine learning happening in that process You know, maybe even from you know, getting it from a company called, you know company like climate company, right?

That uh, you know specializes in applying machine learning to that type of data Yeah and I think it's it'll be different when we take Somebody else's data that they've already done some work and then apply machine learning to data that we've collected internally You know to get some very specific recommendations Mm-hmm. So guys for example, a lot of insurance companies are starting to look at the idea of personalized insurance for personalized auto insurance where they're hooking up devices to the into the ODB port in the car and monitoring the driver's driving patterns and You know essentially giving discounts for you know, fewer miles driven or safer driving habits things like that So Sam and Mark, I'd like to get your feedback on this on what well first let's talk about data set sizes My and my thought would be that the data set sizes for proper machine learning can vary from a Few hundred Meg to depend on the data set to obviously petabytes of data.

Is that a correct assumption? If the Range is pretty broad and it's it's hard to even kind of compare it like that a few hundred Meg could represent You know one image or it can represent, you know, hundreds of rows of you know, relatively simple time-series data So rather than give it like that. It isn't more of a volume of data regardless of size It's it really depends right and it one of the things that it depends a lot on Depends a lot on is the algorithms that you're using.

So we've talked We've talked more about machine learning versus AI You know to date and in some ways, you know, these are interchangeable in a sense that Yeah, most of the interesting AI is based on machine learning But the traditional, you know the simple machine learning Algorithms like, you know linear regression where you're predicting some data point based on Uh, you know features based on data points that you've already collected and and have labels for um, you know, those can be pretty those algorithms can be pretty accurate with relatively few uh data points, I mean depending on the the Data that you have and the number of you know features or dimensions of your data Can be you know hundreds, uh of data so like variability of data so I can predict This is something that we when I was a consultant we did practically use You know, we want to understand how much where the depreciable assets of an organization was And based on relatively small data sets we could then create a model to project Where the depreciable assets were, you know the sure you'd have outliers but more or less that was pretty accurate So it really depends on the nature of the data set Yeah, and I mean and this is stuff that we've been doing forever.

You can do it in excel, right? It doesn't need to be anything. Uh super fancy, you know on the other end of the spectrum There's a crop of algorithms that are broadly referred to as deep learning uh, they're based on neural networks and What you're doing is you are feeding data into one of these deep learning networks and Uh by feeding that data in you are Essentially training the network how to recognize uh the things that you want to recognize so uh features in an image for example, uh and to do that kind of Machine learning takes much more data.

Um You know hundreds of thousands of images for example again, depending on the complexity of your data set and what you're trying to um, you know how many Types of things you're trying to recognize all these things play a role, but uh, that's where we're talking about You know volumes and volumes of data Yeah, I guess one of the things that I really wanted to get out of you today is you know, we we We didn't get a chance to run into each other. I don't think it aws reinvent, you know 43 45 000 people and the chances of two people that know each other meeting is pretty rare But I I saw you on twitter and you did this thing a hot dog.

Not a hot dog Yes You know I like hot dogs. I'm a chicago guy So, you know for me like if it's not from chicago and if it has ketchup on it, it's not a hot dog That's pretty that's pretty straightforward. See I know exactly what you mean. I'm a new york guy and I feel exactly the same way about pizza Yeah, there you go. So The explain this hot dog not a hot dog experiment that you guys did at aws reinvent, okay well, the the reference for those who don't know is from the show the hbl show silicon valley where uh, one of the guys invents this Image detector and it turns out it was only good at one thing identifying hot dogs Um, and so at reinvent aws, uh released, uh a product actually pre-released a product I guess you could say called aws deep lens and this is a uh, you can think of it, this is a gross oversimplification, but you can think of it as like a kind of a raspberry pi with a camera attached to it kind of built into uh, uh, you know in a one-piece device that is designed to get developers excited about and Give them an opportunity to play with uh image-based uh machine learning and ai problems so computer vision stuff and they had a they did a lab there, um, Where you got to play with this device use it with another new Tool that they released called sage maker, which is a a data science platform, uh in the cloud Build models to in this case detect the hot dog and then deploy it down to this device the cool thing about this device is that it's You know, it's I mean, I guess it's not all that cool If you know what a raspberry pi is, but it was a computer like the monitor was plugged into it.

The camera's built in uh, you're connecting out to the cloud building a model and then using other Aws technologies you can kind of automatically deploy that model back down to your edge device uh, and You know start putting up hot dogs in front of it and see if it can figure out. Uh, If they're from chicago or not Now was that using tensorflow or like aws recognition or Uh, I think the models that we did in that. Um In that workshop were based on mx net aws has an affinity for mx net, but I think they Uh, I think the the product will support.

I think it's cross platform So let's talk a little bit about you know, you mentioned mx net tensorflow This is especially google. I think announcement maybe about a year or so ago with tensorflow What's a tensor like what? ml 101 What I see that term all over the place. What is tensorflow? yeah, you can think of a tensor as a matrix or an array of numbers and uh in machine learning and ai for the most part everything is representative at Represented as as these vectors or arrays or matrices of of numbers and values and so Uh, you know tensorflow is essentially a way to process these vectors um and Uh in in essence, that's what you know deep learning is in practice It's you know doing a bunch of vector multiplications and things like that to you know, get you to a prediction So would it be fair to say something like a tensorflow provides the uh fundamental capability to to analyze large data sets Um, yeah, I wouldn't say it's necessarily about performance.

Although the various frameworks that are out there One of the ways that we do judge them one against the other is performance uh, but it's more about you know if you think about um You know if you think about kind of the progression from assembly language to cc plus plus to java, right? There are these layers of abstraction that you know as we advance in our, you know, thinking about problems We can kind of move to higher levels of abstraction and the same kind of thing Uh is happening on the machine learning side, right?

So Um, you know, if you take a class like, uh, you know, andrew ang's famous, you know stanford online machine learning class Right you're doing You know these fundamental uh algorithms like gradient descent and back propagation and the like You're writing these by hand with you know for loops and and stuff like that, right? Okay, and that's that's hard um, you know What tensorflow is doing is trying to allow you to do that at the next higher level of abstraction, right? A vector of numbers an array of numbers, you know, you tell it the operations that you want to do on that vector of numbers And then you tell it to go to town and there are even frameworks that operate it even, you know Higher levels like keras is one Um that you know, you're not looking at these array multiplications anymore But you're looking at higher level operations in neural nets like convolutions and the like So do you think that people So do you think that people will be using features that are already baked into amazon like look at amazon recognition right where it can detect an object in a scene and The facial analysis where it'll try to tell you if you're happy Uh, or your eyes are open are people largely going to use the can stuff that The amazons and the googles of the world have already baked in do you think people will be developing more of these on their own?

Uh, you know, I think I think both will happen and I think in a lot of ways, um You know that one is kind of an on-ramp to the next. 0 to to market Um, that's great But at some point you might need something else for example um You know, you might find a big issue in um In machine learning and ai is the the biases that are present in the data, right? So amazon is training on a specific data set, you know that data set, you know may have biases in it Or it may not be biased enough, you know to your use case.

Um, if you're trying to um Create an application that you know can tell, you know screws and bolts and and nuts from one another aws is off the shelf data set might not have enough examples so, uh, you may need to build your own, uh, and Train it with your own examples to get the kind of performance and accuracy you're looking for so we could go on a lot of pretty actually for a couple hours on this because this is obviously an area that deserves a top 40 Podcast entry.

There's just so much to digest. We haven't talked about data injection How do we actually uh store and utilize this data the compute needed to run against this? Uh these data sets, uh, do we own that stuff or do we run it out from cloud providers? Who's the cloud providers best suited for? Uh our type of ai that we want to perform or machine learning that we want to perform Let's recap where can people find out more about ai? ai.

com is the my analyst firm And i'm happy to give folks specific pointers. ai Uh, that's that's also really good Uh, and of course i've got uh a hundred some odd, uh interviews with folks that are kind of pushing the envelope with this Uh technology on the podcast that you know folks can Take a look at or i'll take a listen to more accurately All right. So mark, where can folks find you in your your your machine learning deep learning? Uh lessons on cincinnati chili Um, you know anytime you're in cincinnati.

We'll give you any chili Offer stands. Yeah, it's chilly but not chilly exactly. com is the site for the podcast and blog At cto advisor on twitter until then, uh, you know what? Let's get some more machine learning around the podcast going going to your mom's uh pod catcher subscribe to both formal and to the cto advisor talk to you next cto advisor podcast