Download

Basecamp Networks uses Machine Learning to diagnose easily what kind of sickness or parasite a crop might be suffering. They’re powered by Google Cloud Platform, and their CEO, Craig Ganssle, is here to tell Mark and Francesc all about it.

About Craig Ganssle

Craig Ganssle is the Founder and CEO of Basecamp Networks. With over 20 years in the technology industry, Craig has extensive experience developing and deploying wireless networks and designing predictive learning solutions for complex problem solving.

Craig started Basecamp Networks in 2007 providing global wireless network infrastructures as well as creative solutions for difficult and time-consuming IT issues. As a partner with Google since 2008, Craig was one of the original Glass Explorers, Google’s original beta test group. In early 2013, Craig advanced to an elite small team for this innovative technology. Since then, under his leadership and Basecamp’s agricultural focus, Craig’s vision for Glass led his team to develop intelliSCOUT®, the world’s first wearable application offering farmers a truly hands-free solution, enabling agricultural problem-solving by collecting actionable insights from the field in a fraction of the time. intelliSCOUT® technology has been demonstrated, in conjunction with the Basecamp Networks’ offering, to dignitaries throughout the world, including France, where Craig was personally requested to present this technology to President Francois Hollande. In October 2016 Craig and Basecamp won the Atlanta Telecom Partnership (ATP) Technology Service Provider of The Year Award (in addition to his achievements in agricultural technological advancements).

Prior to founding Basecamp Networks, Craig was recruited by Verizon Wireless in 2001 where he oversaw U.S. Southeast Operations in the network engineering division and was later tasked with deploying LTE in the Southeast United States. Craig’s advanced innovative solutions are currently in use across Verizon’s entire company footprint today. During this time Craig also received a bachelor’s degree from Muhlenberg College in Business Administration, and a Computer Science degree from AIU.

In addition to his contributions in the public sector, Craig served honorably in the United States Marine Corps as an intelligence communications operator with the Joint Special Operations Command (JSOC) 2nd Force Recon Battalion. He was later assigned as a detached “special operator” under the Department of the Navy, to Naval Special Warfare Unit S.E.A.L. Teams as a “joint operator” before being honorably discharged in 2001 for medical reasons. During his six years of service, Craig was highly decorated with various commendations and medals for his service and valor.

Craig is very active in mission work with his church that includes providing internet services in rural and remote locations globally. Basecamp Networks is headquartered in Alpharetta, GA where Craig resides with his wife and children.

Cool things of the week
Interview
Question of the week

How can I learn machine learning for free?

Follow this courses:

  • Machine Learning by Stanford University coursera
  • Deep Learning by Google Udacity
  • CS 20SI: Tensorflow for Deep Learning Research Stanford

And more:

Where can you find us next?

Francesc presented at Gopherfest and the video is already out there! Next he’ll be teaching at Onboard Buenos Aires and running a workshop at QCon New York

Mark is currently at Nordic Games Conference, and while he won’t be there, if people are in San Francisco they should head over to the Playcrafting & Extra Life 24 Hour Game Fest where we are raising money for the UCSF Benioff Children’s Hospitals.

Transcript

FRANCESC: Hi, and welcome to episode number 77 of the weekly "Google Cloud Platform Podcast." I am Francesc Campoy, and I'm here with my colleague Mark Mandel. Hey, Mark, how are you doing?

MARK: I'm doing very well. How are you doing, Francesc?

FRANCESC: Pretty good, pretty good. Very excited about a very cool episode with Basecamp Networks.

MARK: Yes, super cool episode. Some people I met when we were at Cloud Next not too long ago, having a great conversation about machine learning and agriculture.

FRANCESC: It sounds a little bit like, if you remember, we had this episode a long time ago with Descartes Labs and-- because they were also doing things with agriculture. But this is actually completely different. Rather than having satellite images, they use, basically, sensors and cameras from your phone directly to take pictures and tell you what's going on with plants. It's really cool.

MARK: It's really, really cool. After that, we're going to have a quick chat about learning resources for machine learning.

FRANCESC: They're all free.

MARK: They're all free.

FRANCESC: Yeah. So that's a very cool thing, if you have time. That is the only thing that is missing for me--

MARK: That's true.

FRANCESC: --to go through all of these courses. But if you have the time, there are some really cool resources we're going to be sharing with you about how to learn machine learning. But before that, we have the cool thing of the week-- cool things of the week.

And the first one is a new region. We just announced on May 10 a new region in Northern Virginia. And actually, I did not know where Northern Virginia was. So for--

MARK: You had to look it up?

FRANCESC: Yes, I had to. For all of those out there, Northern Virginia is somewhere in the East Coast near DC. And that's as much as I'm going to say.

MARK: That's fair enough. I don't know any better. I'm Australian.

FRANCESC: Yeah, there you go.

MARK: I have no idea. There we go. It's on the East Coast. So it is US East 4. So that's pretty cool. We're kind of on a bit of a roll with regions, I think.

FRANCESC: It's very cool. You're going to see that, so for Washington, DC, New York, Boston, Montreal, and Toronto, it means that your latency in return trip time will be reduced in between 25% and 85%, which is pretty awesome. If you have some customers in there, around East Coast, there's a huge, huge benefit for you.

MARK: Yeah, I'm just looking through the product supported. It is listed on the blog post that we will have in the show notes, but it's pretty much almost all of them. Everything from App Engine, to Compute Engine, to Dataflow, to Datastore, to Cloud DNS, and Cloud VPN.

FRANCESC: The one that I don't see is Bigtable, for instance, and Spanner. But I'm sure they will be coming soon. So if you're not using those, go check it out.

MARK: And speaking of those products, specifically on the compute side of things, general availability is now there for 64-core CPUs.

FRANCESC: That is cool.

MARK: So if you have massive workloads, or just like making really big machines, then we can definitely do that for you at a general availability level, which is totally cool. And speaking of which, if you go to the blog post, there is a quote there from our friend Tim Kelton from Descartes Labs talking about what they do with 64-core CPUs.

FRANCESC: Yeah, and if you want to learn more about what Descartes Labs do, check out the episode number 41 where we interviewed Tim Kelton, which is one of their cloud architects. Definitely an interesting episode, and we talk about how to use very powerful computers like this one, but also how to make it cheaper by using preemptible instances. There you go.

Cool. And we have one more cool thing of the week. It is one blog post that comes from--

MARK: His name is Alexander Holbreich. I'm going to assume that's how we pronounce it.

FRANCESC: Holbreich. Yeah.

MARK: Yeah.

FRANCESC: Sure. I'll go with that.

MARK: I've actually done a decent job this time, possibly. Really love it when we can find cool stuff from the community-- either people send it to us, or it gets re-tweeted by teammates. And if you have stuff, please send it through.

But he wrote a really cool article about using Terraform with Google Cloud Platform-- sort of the steps that you need to go through to get it all set up, some of the gotchas that can kind of catch you out during the process. And basically, get you set up and ready to go if Terraform is your automation tool of choice, if you want to create machines, and storage, and all sorts of other good stuff.

I'm all for automation. Anything that automates, I'm all for.

FRANCESC: I was going to say that I'm not shocked that you're interested in Terraform, knowing that your favorite product ever is Deployment Manager. So yeah.

MARK: As long as things are automated, I am happy. I don't mind how people do it.

FRANCESC: There you go. So yeah, go check it out. It's a very interesting article. So thanks, Alexander Holbreich, for writing it. So I guess it's time to go talk to our friend Craig Ganssle from Basecamp Networks.

MARK: Let's do that then. So a few months ago, in March when we were at Cloud Next, I had the distinct pleasure of meeting up with a gentleman by the name of Craig Ganssle. Craig, how are you doing today?

CRAIG: I'm doing great. How are you?

MARK: Very, very well. We had a super interesting conversation when we were on the show floor at Cloud Next. So I basically was like, you need to come on the podcast and tell us about all we do. But before we get stuck into the company that you work for and the really interesting stuff that you get up to, can you tell us a little bit about who you are, and what you do, and the company you work for?

CRAIG: Sure, absolutely. I'm Craig Ganssle, as you said. I am the founder and CEO of Basecamp Networks. And Basecamp Networks is an innovative technology company that works with wireless infrastructures and software development using artificial intelligence and machine learning, specifically in agriculture. But we do work in some verticals outside of agriculture.

FRANCESC: What were you doing at Cloud Next? How do you use the cloud? You're talking about agriculture, you're talking about plants. What is the technological part of this?

CRAIG: Sure. We built a product in 2013, an application called intelliSCOUT. And intelliSCOUT is an artificial intelligence platform that uses image recognition to detect pathogens in plants, on crops, in agriculture in real time. The entire thing lives in the Google Cloud Platform. And we love it living there. It has had great results.

And being that it's in the cloud like that, it allows us to scale dramatically around the world. intelliSCOUT is being used in multiple countries around the world. And it's used highly in R&D environments with large corporations to do better trialing and better production for seeds and the performance of crops, I guess we could say.

When seeds go in the ground, and we plant, and then everything grows in the middle, and then we harvest in the end, when everything's growing in the middle, people are trying to capture data from the field before we could potentially have a crop that gets affected by a disease, or we lose that crop. And so intelliSCOUT takes the actionable insights from the field that would normally take weeks, or a month even, to diagnose and get back into the hands of a farmer. With intelliSCOUT, it happens in minutes.

FRANCESC: So some time ago, we interviewed Descartes Labs. And they do something really cool where they are also able to identify some problems, like forests and stuff like that, through satellite images. Is that how you do it? What is the input to your application?

CRAIG: So we don't do anything with satellite right now. What we do is we use mobile devices, what people are already using. And actually, as well with Google Glass, which is how my entire vision for intelliSCOUT began. I was one of the early people allowed-- asked to come up to New York and get Glass. And that's really kind of how it started.

So Google Glass-- iOS right now, coming soon on Android, and also soon on drones.

FRANCESC: Nice. So you're one of those Google Glass explorers?

CRAIG: Yeah.

FRANCESC: The pioneers. Nice.

MARK: All right, so just so I get this right in my head. So basically, you have people out in the field either using their mobile phones or Google Glass, taking photos of crops and then sending that back to some services that run in the cloud, or Google Cloud Platform, that then process that and determine information about it such as yield, or disease, or other such things. Is that pretty much correct?

CRAIG: Yeah. That's pretty much it. It has some additional automations in crop monitoring. There are some things-- like with corn, they will count the kernels on an ear of corn to help predict yield. And today, a lot of times they do that with a Sharpie. Well, if you've ever stood out in July, in a field, in Georgia, it gets really hot.

And so, what was one of the first things we did with Google Glass is now you can take an ear of corn, hold it up in front of Google Glass, or an iOS phone or tablet, and snap a picture. And it counts the kernels with over 97% accuracy in about two seconds.

MARK: Wow.

FRANCESC: That is probably higher than humans.

MARK: Yeah.

FRANCESC: When it's hot especially.

CRAIG: The average percentage that people were accepting in the ag industry was about 72% to 86%.

FRANCESC: Oh, nice.

MARK: Wow.

CRAIG: So we've increased that, and we've narrowed the time window in which it happens. So it takes the picture, sends everything up to the cloud, runs through the process, comes back down in about two seconds with results.

MARK: So is that-- what powers that? Is that TensorFlow? Is that something else? Is that some kind of machine learning thing? What's the deal there?

CRAIG: Yeah, so we started with kind of like OpenCV, and now we are using TensorFlow. There's a lot of components that we're doing. Firebase-- I think we're using Pub/Sub and Kubernetes. We're using a lot of the Google Cloud Platform, the full stack.

FRANCESC: So let's concentrate first on the part that for me is the most exciting one, which is the machine learning side of it. You said that you started with OpenCV. Could you tell a little bit to the audience that doesn't know what OpenCV is? And why were you using it, and what made you migrate to TensorFlow?

CRAIG: Sure. I will answer that as best as I can without my chief developer. But OpenCV is Open Computer Vision. And that was a bit more of an open source sort of practice that we just started messing around with. TensorFlow-- it was just faster, the way the algorithms sort of compared the data that was coming in. And it was able to train the machine to learn the results quicker, if that makes sense.

So taking a cataloged library of imagery and capturing multiple pictures, you have to train the machine. So first you have to start with your zero base. Hey, this is a healthy plant, right? And then this is corn leaf blight, and then this is corn leaf blight pretty severe, and this is corn leaf blight really bad. And TensorFlow managed to take those images and sort of process them a little bit more seamlessly that the results became far more accurate when we took our algorithms, mixed them with TensorFlow algorithms, and really got the entire sequence running.

FRANCESC: Yeah. I think that it is pretty interesting, because OpenCV is not really machine learning. I've used it before. And it allows you to do things like counting the kernels in the corn, for instance. It would be able to do that.

CRAIG: Right.

FRANCESC: But you need to write it. Detecting something like how healthy this plant is-- that sounds actually pretty amazing.

MARK: I was going to ask, did you have a catalog of imagery of diseased corn or something like that to train this against? Did you have that prepared, or did you have to go find that?

CRAIG: [LAUGHING] That's a funny story. No. Actually, the answer is no we didn't. And we searched. We searched agriculture universities, other universities, large seed corporations, even the Library of Congress. Nobody had properly tagged images of crops, or the pathogens, or pests that could infect crops for all the ones that we started working on.

So we had to really start creating them ourselves. And that's been a very daunting task. And that is also where TensorFlow helped, where we were originally having to capture thousands of images to train the machine. Now we're down to-- we can probably get away with 40 to 60 images to get the machine trained to have really good accuracy results. But nobody really had anything.

I mean, people had images. But they weren't cataloged properly, they weren't tagged with what is this, and what pathogen is it, and what's the severity level? And severity level has been really hard because that's very much dependent upon the subject matter experts. So one pathologist might say, hey, this is 30%. Another might say, no, I think it's more like 45%.

And so, a couple of different countries, including the USDA and multiple companies, have been looking at intelliSCOUT to possibly be that first benchmark to set a standard in an accuracy setting, kind of a benchmark setting in the industry-- because of the way it's capturing information, and we're sort of analyzing it.

FRANCESC: So something that really surprised me was you said that, well, to train the model, at the beginning you had to use thousands of images. Which is what sounds normal to me, really. When you are doing deep learning, you really need big data. But then, you say that you went down to 40, 60. What magic is this?

CRAIG: That's some of the algorithms and the ways that we internally have come up with training the machine.

FRANCESC: That sounds like top secret.

CRAIG: Kind of, yeah. Yeah, that falls under the intellectual property clause.

FRANCESC: That's fine.

MARK: That's fine.

CRAIG: It also is very dependent upon what disease we're looking at. Some are very easily detectable and some are not. And for the ones that are not, we still do-- sometimes need to get up in higher images. But it's really not in the thousands anymore. It can kind of just be in the hundreds. We built into the platform, both in Google Glass and on the iOS app, a reticle design.

So when you launch the camera, there's a reticle. And when you line up the leaf, or the crop of the plant in that reticle, it filters out a lot of the background noise automatically. So you're basically certain to get a good picture right from the get-go.

MARK: Cool. So I know nothing about agriculture. But now I'm just kind of curious because it popped in my head. Do you do like insects and stuff? I don't know, are locusts still a problem? I have no idea. Is that a thing?

CRAIG: Yeah, so that falls under the category of pests-- bugs and insects, and working with entomologists. We have, in corn, you have sweet corn army worm. In cotton, you could have boll weevil. Different pests that infest the crop at either an early stage or a late stage.

MARK: You talked a little bit about some of the products you're using at GCP. Can you talk at least a bit about some of the platforms you're using. You mentioned Kubernetes, your Firebase. Well, let's stick with machine learning. Did you use Kubernetes to help train the TensorFlow models? Did you use Cloud ML Engine? What did you use there?

CRAIG: Yeah, so the developers are using Kubernetes to satisfy kind of the definitions automatically with TensorFlow. Load balancing, health checks, and instance groups, Cloud SQL. Pub/Sub for production, depending on some of the system architecture. Some of the others, I think I'd have to get with the developers to find out specifically how they're using everything. But yeah, we also are working with the Lagom, a bit of the Lagom framework.

But TensorFlow, probably one of the biggest that we have.

FRANCESC: Sorry, what's the framework you mentioned?

CRAIG: Lagom.

FRANCESC: I do not know that. Could you tell us a little bit about it?

CRAIG: It allows us to sort of push things to production in real time.

MARK: So I'll ask this question then. So why did you end up choosing Google Cloud Platform as your platform? Were there particular features, or tools, or platforms, or things that made your life easier for what you were building?

CRAIG: So I have three definitive reasons for that. The first one being, we actually started in AWS. And it was not user friendly to work with. And it crashed on us a couple of times. So we moved over to Google Cloud Platform. The interface is just far easier to use and navigate through. The connective framework of things like Kubernetes, and Firebase, and the other things that kind of plug into GCP were a whole lot easier to use.

And Basecamp Networks as a company has a longstanding partnership with Google, with the Glass development team for Google Glass. We're a G Suite reseller and integrator. So those kinds of things, it just made sense to just keep it all in the Google house.

FRANCESC: Great. I'm wondering, since you cannot share the magic about going from thousands of images for training to going to 40 to 60, which is amazing, and I will ask you once we've done recording-- could you share any other best practices of pieces of advice for other companies that might be considering doing something similar to what you did?

CRAIG: Sure. Our chief solutions architect, Dr. Christian Kennedy, he has a really good way of explaining to people. This is how we are explaining the machine learning. You're training a machine, OK? And so, that data, whatever you're training it with, is coming from somewhere. In our situation, it's coming from the expert knowledge of pathologists and entomologists in agriculture, and how they see pathogens and pests affecting crops.

So as we train the machine, we have to go off of the best knowledge that we get from them. The methods of how we're taking pictures, and part of that reticle design to make sure that we're getting good pictures, making sure that they are properly tagged, they are properly classified so that our classifiers have the best accuracy possible going into the original scan of training that machine from the very first time. And I think that's really a big key with any kind of machine learning, and even deep learning.

So sure, we're doing the if this, then that, right? Component with machine learning in agriculture-- if you see this, then it's corn leaf blight. But then, what do you do? Then we get into the deep learning of what is next. And this is what you should do from a spraying an herbicide or a pesticide. If we catch things earlier, we're able to spray less herbicides and pesticides. That means the farmer can have better crop outcome.

That means, globally, we can impact climate change because we're spraying less in all of our farmland around the world. But all of this comes from getting great data to start with. And so, that's why intelliSCOUT, we did not start by putting this in the hands of just all the farmers as kind of a $5 app you could download and subscribe to.

But more starting with the large corporations who have the pathologists, the PhD talent in-house, to give us really good data to start with so that we know we're training the machine at the most accurate levels of severity, and getting the best possible classifiers for identifying what we have, and what we want the machine to recognize. So it's all about really good data to start with.

FRANCESC: So keeping on the data topic, what kind of training have you been doing? Do you train the models regularly? Did you train them once and they're good enough now? I guess that you keep on getting more and more images. Do you also use those images that people are using to detect if there's some illness in their crops to improve your own predictions?

CRAIG: Absolutely. So that's kind of the big factor with TensorFlow is the more you use it, the smarter it's going to get. So we kind of train everything once and really get it going. And the more that it gets used, the more it's getting trained. And so, it's getting smarter the more it gets used. Because even if somebody takes an image, and that image they want to have it classified, or have it identified-- what am I looking at here-- that image is also then repurposed and put into the system as a classifier.

FRANCESC: Great. So one last question on the same topic because I'm very curious. So what happens if someone takes an image, and you're not able to classify it correctly? Does that happen? Do you have some experts in-house? What is the flow in those cases?

CRAIG: It sort of casts it off and alerts us that we have an anomaly. We have a situation here with an image that is not fitting with the classifiers that we've programmed in the machine. So if you're out scouting corn for army worm, and you take a picture of, I don't know, southern rust, the system is going to say, that's not army worm, but I don't know what it is. We need to take a look at this image.

Then once we program and train it on southern rust, it will say, that's not army worm, but that's southern rust. So if it's something new, it's going to cast it off and alert through the system and say, it's not army worm, but I don't know what it is.

FRANCESC: Great.

MARK: Awesome. So now I'm also super curious from like a practical level. What do you see as being-- what has been the biggest change in regards to how agriculture or farming-- that you've had under this program has changed from what it was previously, if that makes sense?

CRAIG: How have we impacted farming, kind of?

MARK: Yeah, exactly, exactly.

CRAIG: Going back to whether it's the app on a mobile device or utilizing Google Glass, part of what we want to be able to do with intelliSCOUT as it gets smarter and smarter and gains more knowledge, and more crops, and different diseases, and different pests-- is really enable smallholders of the world. We want to be able to take this technology and share it with people in other countries.

So for example, in some far remote areas of Africa where their land is not as rich as maybe it is here in the United States, and so every square acre of land is very valuable to them. And they want to maximize what they can grow off of that land to its fullest potential. So by being able to catch diseases or pests far earlier using intelliSCOUT, by sending them something like a pair of Google Glass.

And they put it on, and they launch the intelliSCOUT application, and they simply scan the field and take pictures. And the results come back in near real time with, here's what's going on. And then, into that deep learning of, here's what you need to do. And so, if you look at something that would normally take weeks or a month, until you get those results, depending on weather patterns, your disease could spread far more.

And back where it originally started, that could be just dead and gone and lost crop and lost yield. So if we can capture things in more real time, and we can enable smallholders with better knowledge, they're able to grow more, grow better food, become much more sustainable. So that's one way is to just kind of enable the education of smallholders around the world with better technology.

The other piece of it is just, as I said with the image library, there wasn't anything for us to start with. And so, we're building that catalog library now that will exist in agriculture.

MARK: So you said something really interesting there that I think you kind of slipped in, which was once you've got sort of the diagnosis, giving suggestions on what people should do. Is that something you're doing now, or is that a plan for the future?

CRAIG: It's something we're doing now to a small extent. We want to correlate even more data from other data points on the farm. So intelliSCOUT would be wherever you're standing when you take a field report, it tells you where you are and where-- where you were and what the weather was when you took that report. So if you're standing in a field, and you take a picture of corn, and you're looking at a particular disease-- we want to correlate data that came from the machinery from the day it was planted.

So the John Deere planter, or the AGCO planter, or whatever-- when intelliSCOUT takes the report, it says, yeah, but I want to go further and see when this was planted. What was the spacing? What was the compaction? And from then until now, what's the irrigation been like? Have we had a lot of rain?

Have we had to manually irrigate through pivots or other methods? What's the nitrogen levels been from then until now? And really start pulling data from other data points on the farm to make intelliSCOUT smarter and smarter and get into that deep learning. That then allows us to get into predictive modeling. And the predictive modeling could potentially change the game for agriculture.

FRANCESC: That's [INAUDIBLE] amazing.

MARK: Yeah.

FRANCESC: So unfortunately, we're kind of running out of time. But I wanted to give you the opportunity of letting us know if we missed any interesting topics. Everything you've said was pretty amazing.

MARK: Yeah.

FRANCESC: But is there anything else that you'd like to mention?

CRAIG: I don't think so. You guys asked some really good questions, and it kind of explains it really well. We enjoy being in the Cloud Platform. It's working out really well for us. We do a lot with Google in other areas. We're glad this whole project started with Google Glass, which a lot of people are like, oh, it's dead and gone. And it's been doing some wonderful things in enterprise.

And we're proud to be one of those companies who've kind of kept it going. So thanks for having me. I really appreciate it. Glad we could kind of tell the story of what we're doing here. And we'll see where it goes.

MARK: Awesome Craig. Well, thank you so much for joining us. That was super, super interesting.

FRANCESC: That was amazing. Thank you.

CRAIG: Thanks a lot. Appreciate it, guys.

MARK: A huge thanks to Craig from Basecamp Networks chatting with us about what they do with machine learning in the agriculture industry. I love the super cool practical application of Google Glass, actually. I thought that was quite cool.

FRANCESC: That was pretty cool. Yeah, for me, the coolest thing is like all the potential that this technology has to actually help people and help countries. The fact that your plants get sick can actually create real problems. So this technology is amazing.

MARK: Awesome. Well, since we're talking about machine learning, why don't we segue nicely into the question of the week?

FRANCESC: Very nice. Let's segue the--

MARK: Yeah, thank you. It's always good when we explain segues. They work so much better.

FRANCESC: So yeah, so the question that we received was, I want to learn machine learning. Where do I get started? And especially, oh, just go to university, go to college, whatever. Those are options, but they're, in general, pretty expensive. So instead, what we found was three different courses that are completely free and available on the internet, and that you should check out.

The first one, the one I went through on Coursera, and it is a very cool one, especially because the teacher is Andrew Ng. I think I pronounced it correctly. He used to be at Google. Now I think he's at Baidu. He's one of the geniuses of machine learning. And you can see him explaining how machine learning works starting from zero.

Like understanding what are the basics from the point of view of mathematics, and moving into deep learning, and neural networks, and all of that. It is really cool. And it's a very good place to start. That's for sure.

MARK: I would second that. I've done a quarter of it.

FRANCESC: Yeah, I mean I didn't finish it.

MARK: I ran out of time. But it is excellent. It is absolutely excellent, totally free. It's on Coursera. I have a strange feeling, I could be wrong, that it's actually the first course they had on Coursera.

FRANCESC: I am not sure, but it is possible. It is possible. It's a pretty old one. But it's actually really, really good.

MARK: It's amazing. It's really from fundamentals-- really enjoy it. You have a second one here that is from Udacity. I haven't done that, but it's also a deep learning course.

FRANCESC: Yeah, It's a deep learning course by Google. And basically, the idea is if you already have the basics of machine learning, and what you want to learn is, so how do I do deep learning, this course is by Google. And it is taught by--

MARK: Oh, I'm not pronouncing that last name.

FRANCESC: --Vincent Vanhoucke.

MARK: That seems reasonable.

FRANCESC: I'm going to stay with that. And he's a principal scientist at Google. He works in the brain team. So you're going to see things like deep learning and, specifically, how we use it at Google. Also talking about how TensorFlow works. It is a pretty long course. It says the timeline is around three months. But again, it's completely free, and it's on Udacity. So check it out. We'll have it on the show notes.

And finally, there is one more from Stanford. And I have not taken this course from Stanford, but I took one long time ago. They're awesome. You have all the videos, all the notes, all the slides. And this one specifically on how to use TensorFlow for deep learning research. So this one is definitely not an introductory course. Don't go there trying to be like, hey, I know some math. I just want to do--

MARK: And it'll be fine?

FRANCESC: Yeah. It will not be fine. It will definitely hurt. But if you already know what machine learning is and what you want to learn is what are the basis of TensorFlow-- like what is a tensor, and how it's implemented-- this is the course you want to go through.

MARK: Cool. So it sounds like, all right, if I don't know anything, I should go to that Coursera course. If I'm sort of mid-level, I go to that Udacity one. But if I'm already like, man, I've got mad skills in machine learning, then I go and do the Stanford one instead.

FRANCESC: I'd say that that's correct. Yeah.

MARK: Awesome. All right. That's great. Now I'm going to go stare at these courses and potentially do them maybe at some point in the far future.

FRANCESC: I know. The cool thing is, Udacity and Coursera, they have this mobile app that allows you to download and watch them later. So yeah, you can actually watch that while you're on the plane. So yeah, that is really cool. But yeah, then you need the time to actually watch those videos.

MARK: Now, we should also make note that episode 71, we did do an episode on Cloud ML with Yufeng Guo. And we also did a TensorFlow episode with Eli Bixby--

FRANCESC: True.

MARK: --that you're going to look up the number for while I'm talking, which goes through TensorFlow and how it works, and how it can be applied to machine learning as well.

FRANCESC: Yeah, and it was episode number 31. TensorFlow with Eli Bixby, very cool episode, very interesting-- and actually, I think the first one that we actually put on YouTube.

MARK: That is pretty cool.

FRANCESC: Yeah. And also, since we're talking about that, it's worth mentioning this could have been a cool thing of the week. The GP used for Cloud Machine Learning Engine, which is what we discussed with Yufeng Guo on episode 71, they're now general availability. So a little announcement. So if you want to check used GPUs, now you can even with general availability that comes with SLAs, SLOs, and [INAUDIBLE].

MARK: All right. Well Francesc, thank you again for joining me. But what are you up to, what are you doing, where are you going?

FRANCESC: So I'm finally back to San Francisco. And by the time this episode comes out, Gopherfest will have been a huge success on Monday, or at least I hope so.

MARK: I expect so. I would expect no less.

FRANCESC: But after that, I will be going to Argentina. And I will be speaking at a Go meet-up in Buenos Aires. Buenos Aires pronounced with Australian accent, probably. That will be on June 1, and on June 2 I will be there for Google Cloud OnBoard Latin America, which is a really cool event.

It is a training event. You can go there and learn about all the Google technologies. And the cool thing is that it is in Buenos Aires. But it will be also live streamed to all Latin America. So we'll have a link on the show so you can follow the live stream. And the live stream will actually be in Spanish. I'm going to be giving talks in Spanish.

MARK: Nice. That's cool.

FRANCESC: Yeah. What about you? What are you up to?

MARK: When this goes live, I will be literally, probably almost on stage, at Nordic Games Conference talking about containers, and game servers, and stuff. Then I'll come back for a little bit and go on a bit of a vacation. But I definitely want to do a shoutout to one particular event that we'll be hosting at the Google launch pad in San Francisco.

They actually do things all across the United States, as well as worldwide. They're a program called Girls Make Games. They basically run game jams, as well as mostly summer camps, for girls that are of the age from--

FRANCESC: You said 11 to 17 before.

MARK: 11 to 17. Yes. Girls from the age of 11 to 17. So they're going to be running the SF1 in the Launchpad for a short period during that. It's basically a full program in which girls get to learn how to build games, deploy games. It's really, really cool. It's over the summer period around July 10. They have a variety of locations across the United States for the summer period when kids are off school.

FRANCESC: Cool. We'll have the link in the show notes. Maybe if you are a girl between the ages 11 to 17, check it out. And if you know one of those, check it out too, and let them know.

MARK: No, it's a great program. I can't recommend it highly enough.

FRANCESC: Well, I guess it's time to say goodbye. But before that, why don't we remind all of our listeners how to get in touch with us?

MARK: Certainly.

FRANCESC: So we have web page.

MARK: At gcppodcast.com.

FRANCESC: We have an email.

MARK: Hello@gcppodcast.

FRANCESC: We have Twitter.

MARK: @gcppodcast.

FRANCESC: We have Reddit.

MARK: /r/gcppodcast.

FRANCESC: We are on Slack.

MARK: At the #podcastchannel@slack, which can be found at bit.lease/gcp-slack.

FRANCESC: And I'd say that's it.

MARK: We have Google+@plusgcppodcast.

FRANCESC: I should never forget Google+.

MARK: Excellent. Well, thank you again, Francesc, for joining me for yet another week.

FRANCESC: Thank you, Mark. And thank you all for listening.

MARK: And we'll see you all next week.

Read full transcript... Hide full transcript...

Continue the conversation on: Google+ or Reddit