Helping African farmers with TensorFlow

Amanda Ramcharan, Latifa Mrisho, and Peter McCloskey joined Daniel and Chris to talk about how Penn State University are collaborating to help African farmers increase their yields via a TensorFlow powered mobile app. Sponsors: Hired https://hired.com/practicalai – Salary and benefits upfront? Yes please. Our listeners get a double hiring bonus of $600! Or, refer a friend and get a check for $1,337 when they accept a job. On Hired companies send you offers with salary, benefits, and even equity upfront. You are in full control of the process. Learn more at hired.com/practicalai https://hired.com/practicalai. Rollbar https://rollbar.com/changelog – We catch our errors before our users do because of Rollbar. Resolve errors in minutes, and deploy your code with confidence. Learn more at rollbar.com/changelog https://rollbar.com/changelog. Fastly https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorship – Our bandwidth partner. Fastly powers fast, secure, and scalable digital experiences. Move beyond your content delivery network to their powerful edge cloud platform. Learn more at fastly.com https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorship. Linode https://linode.com/changelog – Our cloud server of choice. Deploy a fast, efficient, native SSD cloud server for only $5/month. Get 4 months free using the code changelog2018. Start your server - head to linode.com/changelog https://linode.com/changelogFeaturing: Amanda Ramcharan – Website http://amramcharan.com/, GitHub https://github.com/aramcharan, X https://x.com/amram_1987Latifa Mrisho – X https://x.com/tifa_mrishoPeter McCloskey – Chris Benson – Website https://chrisbenson.com, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaShow Notes: Assessing a mobile-based deep learning model for plant disease surveillance https://arxiv.org/abs/1805.08692Penn State-developed plant-disease app recognized by Google https://news.psu.edu/story/513236/2018/04/02/research/penn-state-developed-plant-disease-app-recognized-googleUpcoming Events: Register for upcoming webinars here https://practicalai.fm/webinars!

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Baiting for change log is provided by Fastly. Learn more at fastly.com. We move fast and fix things here at change law because of Rollbar. Check them out at rollbar.com and we're hosted on Linode servers. Head to linode.com slash change log. This episode of Practical AI is brought to you by hired. One thing people hate doing is searching for a new job. It's so painful to search through open positions on every job board under the sun. The process to find a new job is such a mess. If only there was an easier way. Well, I'm here to tell you there is. Our friends at hired have made it so that companies send you offers with salary, benefits, and even equity upfront. All you have to do is answer a few questions to showcase who you are and what type of job you're looking for.

• 0:45Speaker 00

  They work with more than 6,000 companies from startups to large publicly traded companies in 14 major tech hubs in North America and Europe. You get to see all of your interview requests. You can accept, reject, or make changes to their offer even before you talk with anyone. And it's totally free. This isn't going to cost you anything. It's not like you have to go there and spend money to get this opportunity. And if you get a job through hired, they're even going to give you a bonus. Normally it's $300, but because you're a listener of Practical AI, it's $600 instead. Even if you're not looking for a job, you can refer a friend and hired will send you a check for $1,337 when they accept the job. As you can see, hired makes it too easy. Get started at hired.com slash Practical AI.

• 1:39Speaker 00

  Welcome to Practical AI, a weekly podcast about making artificial intelligence practical, productive, and accessible to everyone. This is where conversations around AI, machine learning, and data science happen. Join the community and slay with us around various topics of the show at changelog.com slash community. Follow us on Twitter, we're at Practical AI FM. And now onto the show.

• 2:03Daniel Whitenack?

  So Chris, I am super excited today. I have seen a lot of amazing things in my life, but this next project that we're going to talk about is really pretty amazing. I think you probably agree with that. Absolutely. Yeah, so we have Amanda, Peter, and Latifah with us here, two joining from Pennsylvania, and Latifah joining from Tanzania. So the first thing I'd like to do is just have you guys intro yourselves a little bit. Maybe one of you can do that for all three of you. Amanda, do you want to take that?

• 2:38Speaker 02

  Sure. So, hi everyone. My name is Amanda. I am a postdoctoral research fellow here at Penn State, and I got my PhD in agricultural engineering and computer science. And I work with Plant Village to design

• 2:54Speaker 02

  models that diagnose plant diseases on a phone. And with me, I have Pete. Pete is a research technician for Plant Village, and he's my right-hand guy to help us like, train the models and test the models. And he also helps a lot with the development on TensorFlow. And with us is Latifah. Latifah is a grad student out of Tanzania. She works with us. Her field is microbiology. Right, Latifah? Did I get that right?

• 3:25Speaker 03

  Not really.

• 3:26Speaker 02

  Nope, nope. Okay, well, that's what was in the video.

• 3:32Speaker 02

  So what is your degree in, Latifah?

• 3:34Speaker 03

  Molecular Biology and Biotechnology.

• 3:37Speaker 02

  Yes, okay. Latifah is a grad student in Molecular Biology and Biotechnology, and she's been really, really helpful to us to test the model and how the app works out in the field with scientists, as well as extension workers for agriculture, as well as farmers.

• 3:57Daniel Whitenack?

  Awesome. And did you guys all meet, I guess,

• 4:02Daniel Whitenack?

  you met around this project, right? How did the project get started? And maybe just give us a little bit of an intro on what the project is and what you're trying to do.

• 4:13Speaker 02

  So the foundation of Plant Village is to make technology available to farmers. So make the latest technology available to farmers to improve food security globally. And one of the technologies right now with the highest potential is artificial intelligence and machine learning. So and phones are a very prevalent technology right now across the globe. So we wanted to combine those two technologies to create tools that farmers can access to help them grow food better. One of the major crops grown in the world is cassava. And there are many diseases and pests that affect cassava and reduce the yields.

• 4:56Speaker 02

  So we wanted to combine AI with phones to help farmers grow cassava better.

• 5:03Chris Benson?

  And for those who don't know cassava, could you could you just mention briefly what that is because not everybody in the audience will be familiar with that.

• 5:11Speaker 02

  I think I would I would I'm gonna ask Latifah, the cassava expert, to to define cassava. I think she can do it the best.

• 5:21Speaker 03

  Okay. Cassava is a root crop, just like potatoes and yams. So here we we planted to use the roots as source of starch, which you can boil, you can fry, you can mix it up with coconut milk. But also, we use the leaves. And the leaves, we use it as source of vegetables. So you can basically just boil it and put it in salt. And you can eat the cassava and the leaves. But one of the other important uses of cassava is we make the starch into flour and we cook it in a way that is almost like polenta.

• 6:02Speaker 03

  And you can have that with any other thing like vegetables and beef and stew or sauces. So it's a it's a root crop that is high content of starch and the leaves are used as source of vegetables for vitamins and other micronutrients.

• 6:18Chris Benson?

  Sounds really versatile then.

• 6:20Speaker 03

  Yeah, it is. Yeah.

• 6:22Daniel Whitenack?

  And you're making me hungry.

• 6:25Daniel Whitenack?

  And I haven't eaten yet. And I'm kind of craving some cassava now. So, yeah, you mentioned that, you know, you want to kind of help these farmers in the field grow cassava better. What is kind of the range of diseases that affect this plant? And how, you know, how are they dealing with these diseases in the in the absence of, you know, the solution that you've developed around AI?

• 6:58Speaker 03

  So mainly the way cassava is being grown in most of Sub-Saharan Africa where we grow it, people just grow it as a crop in the field. So it's a substance called crop, something that can grow it, uproot it within six months, make their flower and have food. But right now it's being commercialized and people are growing it to export and making flowers for export. So the main problem with the diseases that we have, that there are a lot of diseases, but the main diseases, the two viral diseases, these are caused by viruses. And one is the mosaic disease, we call it cassava mosaic disease. And basically this affects the leaves of the plants and the leaves become destroyed

• 7:39Speaker 03

  to a point that it does not make enough food for the plant as it should be when it's healthy. And because it destroys the leaf and the leaf, the plant does not grow very well. It becomes stunted, it doesn't grow, and it doesn't produce good roots. So when people come to harvest, eventually they find out they actually don't have enough roots or no roots at all. With the other disease, it's called the brown streak disease, it's also a viral disease. This affects the leaves too, but the leaves are not destroyed. And what it does, it actually destroys the roots. So when you harvest, you find roots, but when you cut the roots and they are rotten inside, so they are useless. So it's either you don't have product or you have useless product.

• 8:22Speaker 03

  So at the moment, there is no viable solution other than identifying the plants that are infected. And those that are healthy, so that you can plant healthy plants, or um remove those infected ones so the disease does not spread. But that's where the challenge is. Some of these uh symptoms for these diseases are difficult to identify. So sometimes people don't know if their plants are infected until when they harvest and they realize they don't have any products that is useful.

• 8:54Daniel Whitenack?

  Yeah, that must be frustrating.

• 8:57Speaker 03

  Oh, yes, it is, because you can lose the whole farm. And people depend on cassava for food or source of to sell to get income to for their livelihood.

• 9:08Chris Benson?

  So it would be pretty disastrous for for the for the harvest to fail in that way, right?

• 9:13Speaker 03

  Yes.

• 9:14Chris Benson?

  And and so as as we turn toward, you know, how you're what you're doing with this project, you know, what are the goals of the project and and how are you finding that AI has become a useful tool for you in accomplishing this?

• 9:26Speaker 02

  So one of one of the main goals is to sort of use AI to democratize access to technology. So the conventional way of diagnosing these plant diseases is for someone with that expertise to go out into the field and look at the plant and tell you what's wrong with the plant. But there are just not enough people in places like Sub-Saharan Africa to do that. So what what we want to do is put that knowledge and collaborate with the scientists who have that knowledge to input that into a model in a phone. And now farmers can just download an app and then have the eyes of an expert look at their crop and tell them what's wrong.

• 10:07Chris Benson?

  That is so cool.

• 10:09Daniel Whitenack?

  But yeah, that's awesome. So in a sense, the like they're getting that expertise via the model on the phone that they just essentially would not have otherwise.

• 10:19Speaker 02

  Exactly, yes. And um we we definitely work in collaboration with the scientists and experts to create a model like this. I think one of the um.

• 10:31Speaker 02

  Sort of negative sides of AI is people always saying that, oh, AI is going to take our jobs. But we actually can't build these models without the human experts there to tell us what type of data to collect and how the model should perform.

• 10:45Daniel Whitenack?

  Yeah, that's that's so true. And and I I definitely think this is a great example of that kind of augmentation, right? Because you're also you're not creating like a a robot farmer, right? That does all the cassava things with with AI and replaces everyone. But really, you're putting expertise into the hands of the farmers in the field. It's really allowing them to improve their their quality of life and and increase their yield, right?

• 11:13Speaker 02

  Yeah, exactly, exactly.

• 11:16Daniel Whitenack?

  Awesome. So how like in terms of I guess we can kind of, you know, steer things a little bit more on the the AI side now. I'm really interested, you know, in in some of the the technical things around that. First off, I I imagine that there was a problem of even, you know, finding cassava data to train your models, right? How did what was the process like in terms of starting out this project? And even just finding the data that you needed to to train these models?

• 11:47Speaker 05

  Yeah, so the data collection process actually, I think, took about two years. From the beginning of 2016 through October 2017.

• 12:02Speaker 05

  So that was just like a kind of mass photograph period where there was at least three trips. One before I came onto the project and then my first two trips where we just went to Tanzania with some cameras and spent anywhere between like five and eight hours out in the field just photographing leaves. Coming back and sorting them out and making sure that like what we were photographing is what or what we thought we were photographing was what was actually there in in terms of like the the correct ground truth labels. Because that's one of the more difficult things to. When you're dealing with any type of, you know, health or sciences, especially

• 12:46Speaker 05

  plant sciences, correctly labeling the data is really, really difficult process. And because of the different symptoms, looking differently on different varieties and at different stages of the infection, they look different, as well as having like co-infections. So how do you deal with that? And so that I would say the data collection process was probably the biggest, most intensive, hardest part of this project because there is just like no publicly available, especially cassava data, but generally like plant disease data sets are not publicly available. And then, so we had to build our own and then go through all of the, you know, the troubleshooting

• 13:29Speaker 05

  and difficulties that are, that are come along with building your own data set.

• 13:33Chris Benson?

  So a couple of follow-up questions on that. First of all, what kind of sample size did you have in terms of how big was the data set that you used of those leaves?

• 13:40Speaker 05

  So the data set that we used for the final model, or that we're using now is made up of about like 2,200 images, I think. But like our entire cassava data set, I think, has somewhere between 15 and 8, 10 and 15,000 images. But like, like I said, like not all of those images are useful because it could be incorrectly labeled or because they're just a bad style of image. So, so we really only have about like 22 or 2,500 really high quality images that we can build a model with.

• 14:17Chris Benson?

  It was, was it challenging, given the fact that it takes that expertise? You know, this is a domain expertise that, that most people don't have for the labeling. So whereas in, in other projects, not like this, you know, you can kind of outsource some of that labeling if it's something that, that your typical layperson could, could address. But I imagine that recognizing the disease states took some level of expertise. How, how did you approach that?

• 14:43Speaker 05

  So that's where IITA comes in and they are the cassava experts for us. And so we, in October, sat down with, I think, three cassava experts and went through the dataset. And so they gave us the, so when we go out and collect the data, we go out with a cassava expert and they point us to like, okay, this is what this virus looks like. All of these leaves are infected with this. So during the whole process of the data collection, the IITA cassava experts are there with us, making sure that we're getting the right data.

• 15:16Daniel Whitenack?

  And, and the-

• 15:17Speaker 02

  So IITA-

• 15:18Daniel Whitenack?

  Yeah, go ahead. I think that was my question.

• 15:20Speaker 02

  Yeah, so IITA is the International Institute for Tropical Agriculture and their East Africa office is based in Dar es Salaam, Tanzania, and that's where we met Latifa.

• 15:31Daniel Whitenack?

  Gotcha. Yeah, so in a sense, you're, I mean, it wasn't like you went, you guys went and then just took a bunch of pictures and then shipped the pictures off to some people to identify disease plants, but you kind of did it on the fly in the field, saying, you know, this is a disease plant at this stage, I'm gonna take a bunch of pictures of it so that I, so that I know those are labeled correctly. Is that kind of right?

• 15:54Speaker 02

  Yes. Yeah. Exactly. It's part of a bigger, bigger role of Plant Village to sort of have a dataset of images of all the plant diseases of the world.

• 16:06Daniel Whitenack?

  Awesome. Are there other existing datasets that kind of fit into this, the same category of, of disease plants? I'm just not, not familiar enough with that field.

• 16:20Speaker 02

  There are, there are a number of different international organizations that keep databases like this, but they're not open access.

• 16:28Daniel Whitenack?

  Gotcha. Okay. And is, is this cassava data that you're producing, is, is that gonna be available publicly?

• 16:35Speaker 02

  That's sort of yet to be determined. We're definitely sharing it with researchers for research purposes, but whether it'll be publicly available, I'm not sure yet.

• 16:45Daniel Whitenack?

  Right.

• 16:46Speaker 02

  We're definitely open to sharing the, the model and the weights on the model.

• 16:54Speaker 02

  So we can, because we're making the model openly available and free?

• 16:59Daniel Whitenack?

  That's awesome. Yeah.

• 17:00Speaker 02

  Yeah.

• 17:01Daniel Whitenack?

  Yeah, that's, that's great. So even if, even if people, you know, wanted to integrate, you know, Kasaba intelligence into other apps, at least that, at least that model would be out there, right?

• 17:12Speaker 02

  Exactly. Yeah.

• 17:13Daniel Whitenack?

  That's, that's awesome. And in terms of, I mean, you guys mentioned the importance of mobile devices in this process, you know, could you explain a little bit more, maybe like, why, why mobile devices were the target device for, for this work? And then maybe someone can share as well, like what, how you eventually landed on TensorFlow as, as the framework that would allow you to kind of target that device?

• 17:43Speaker 02

  Right. So mobile is sort of, is emerging as the most ubiquitous platform for technology. It's also, the great thing about an app is that you can run it offline. And in a lot of these farms and these locations, there's no internet access. Farming is done in usually rural locations. So we had to sort of set our restrictions there for, for meeting the population of users that we wanted to use this technology. So that's why we chose mobile.

• 18:15Speaker 02

  And then what was your second question?

• 18:17Daniel Whitenack?

  Yeah. So, so yeah, my, my second question, that's, that's great to, I definitely figured that would be the, the answer as far as, you know, the, the access to computing devices, but also kind of the connectivity and, and all of that. What, what, I mean, given that you knew that you were targeting the mobile devices, in terms of the actual, like, you know, enabling the, the model to run on the mobile devices and all of that, what were the particular challenges and solutions that, that you found to those? You know, I, I know, like, maybe battery usage is, is a thing. I know I can, you know, I can run, you know, Slack and watch a couple YouTube videos on my phone and then I have to, I have to charge it, right?

• 19:02Daniel Whitenack?

  So maybe there's like, what, what were the particular challenges that, that you found with that?

• 19:08Speaker 02

  Yeah, so definitely first model iterations that we did when we took it out in the field, they phones just got so hot so fast. And then the inference times slowed down considerably and again, the battery was draining. So we had to kind of step back and think, okay, we need a different model design to to work in these conditions. And that's when we moved from classification to object detection. Um, with the new object detection models, the inference times stayed uh low and they mo- their phones didn't heat up as fast. And the battery didn't drain as quickly. So that was a big change that we had to do.

• 19:48Speaker 02

  And what was really nice about the object detection was it when you can really see how the model is working because it draws these boxes around what it's is classifying as a specific disease or problem. So in a way, it's also teaching people using the app. This is what I'm looking at and this is what this disease is.

• 20:09Daniel Whitenack?

  Yeah, that's that's super interesting that, you know, you started out actually thinking, it it sounds like you started out thinking like, oh, the thing to do is just really to tell them like disease or not like classification, but really you found a lot of value both operationally and like in terms of of teaching the people more and and bringing bringing that intelligence to them as far as just detecting detecting objects and disease portions of plants. That's that's super interesting.

• 20:38Speaker 02

  Yeah, it was a win-win.

• 20:41Chris Benson?

  One of the things I was under, I was curious about is as you is you got out and did this and had to select the the devices and uh and your your computation and and memory and stuff on it, what did you run into from an inference standpoint in terms of like, um, you know, not every mobile device can can support inferences and in terms of computationally, uh, what, how did you go through the process of saying these mobile devices are in or these are not, or you selected a particular one? How did you approach that? And what were the constraints that you faced?

• 21:12Speaker 02

  Yeah, so um, earlier this year, we went out and tested a number of different mobile devices to see how they did. And we were actually pleasantly surprised that. So the app that we have is compatible with Android 5.0 and above. And um there was a range of performance between the the older mobile phones to the newer ones. But there weren't as wide a range as we expected. Uh so they so they did a pretty good job. So we kind of further tested the app with a mid-range phone. Um, we did see a lot more variability in with the the response of the camera to how they did.

• 21:51Speaker 02

  But one of the variables that we're interested in is whether the app is misdiagnosing or not diagnosing at all. Because if it's not diagnosing at all, that's okay, because we can design around that. But if it's misdiagnosing based on different phones, then that's a really big problem. So then we start to get more nuanced into, okay, what are the metrics that we're going to use to really evaluate? Um, the app on different phones, because uh it's not all equal and it's not going to be perfect. Uh so so what do we want to sort of aim for? And we we were aiming for more false negatives as opposed to false positives. In the way the app performs. And Pete, you you do you want to add to that a little bit?

• 22:37Speaker 05

  Yeah, it's just kind of along the lines of we we would like the model to be more conservative in its predictions. And and and instead of saying the wrong thing, it will just say, I don't know, or I'm not confident enough.

• 22:48Speaker 02

  Or look at more leaves.

• 22:49Speaker 05

  Yeah, it's can show me another leaf that maybe has more symptoms.

• 22:53Speaker 02

  Because when a human goes out into the field, they don't just look at one leaf and tell you your problem. They're looking at the whole plant, the whole field. And people tend to be more conservative about how they, you know, tell somebody bad news. As opposed to being like, well, I'm just going to tell you all the possible problems you have out there and I'm not sure. So we wanted our model to to sort of replicate being conservative.

• 23:19Chris Benson?

  So as you were having to go through and decide exactly what the architecture of your model would be, and you know that you're deploying it into an area where you have a lot of mobile constraints around it, both from the device itself and maybe from the environment. How did that affect the decisions you made on what architectures to go with? Whether you could utilize transfer learning or not? How many layers? That kind of thing. How did the environment affect your modeling decisions?

• 23:49Speaker 02

  So definitely, we had to use transfer learning because there was no way we can collect that many images to train a model from scratch with these specific classes, disease classes. There's just not enough data out there. It'll take a really long time. But then, we were really constrained to the MobileNet architecture because the inference time is the shortest, it ran the fastest. We've also tried, and we're continuing to try different ways to shrink other more complicated, deeper models. But we really got a good performance with the MobileNet architecture. Like, it really worked for our test case.

• 24:29Speaker 02

  So we didn't have to do a whole lot of extensive testing of different architectures for this specific Casava problem.

• 24:39Chris Benson?

  Did you have to remove layers or anything, just given the vast number of layers in MobileNet? Or anything? Or were you able to just take it as is, transfer that over, and then train on top of it?

• 24:50Speaker 02

  We did exactly that. We took it as is, transferred over, and trained, and did transfer learning on the MobileNet model. And it's vastly better than when we first started off with the Inception V4 model to do classification. It was much more robust on a mobile device.

• 25:09Daniel Whitenack?

  Yeah, that's awesome. And that's great to hear. You know, it definitely seems to be there's more and more. You know, I remember there's a lot of pre-trained models out there. There's a lot of different architectures. You know, used to, it was pretty hard for me to kind of go through that process of just getting a kind of off-the-shelf model and having it perform well. But I think, you know, that is changing. And especially with, you know, a lot of stuff I heard around TensorFlow estimators and other things at TensorFlow Dev Summit. I know that there's a lot of emphasis on that sort of application, which is really great to hear.

• 25:53Daniel Whitenack?

  So kind of, yeah. Speaking of TensorFlow Dev Summit, which is where I met you guys.

• 26:02Daniel Whitenack?

  Just to be real here, if Jeff Dean featured my project in his talk, I think I would like flip out. What was that like?

• 26:14Speaker 02

  All right, full disclosure, I was like, wait, who's Jeff Dean? What's happening? And then Pete's jaw just dropped. He's like, what is wrong with you?

• 26:26Daniel Whitenack?

  You guys are just so wrong. Like, you're all about, you know, saving the world with AI. And Jeff Dean is just tagging along and trying to.

• 26:40Speaker 05

  It was actually pretty funny for me because so my brother is working at Google now. And he is working in the Google Brain, you know, part of it with AI and works several levels underneath Jeff Dean. And I was, so I guess they were streaming the summit around the Google offices. And my brother was texting me like, oh, my God, I'm so jealous. You just got a shout out from Jeff Dean. Like, that's so crazy. And then I was like, yeah, this is awesome. And then we actually got to talk to Jeff for like a solid 20 minutes after sometime, you know, during the summit. And that was just like unreal.

• 27:20Speaker 02

  Yeah, it was super validating because we were really able to talk to him, get all geeky about the models and how the models perform and the performance metrics. And we were able to relate a lot to the work that we were doing because he does a lot of work with health data.

• 27:38Speaker 02

  So yeah, so it was like, okay, like, we're on the right track.

• 27:42Daniel Whitenack?

  That's awesome.

• 27:43Speaker 02

  If we can hold a conversation with Jeff Dean, great.

• 27:46Daniel Whitenack?

  Yeah, I mean, just like incredible stuff. And of course, we'll put the, we'll put the link to the video, the featured video in his talk in the show notes for everyone to see. But that's just super cool. And so it was so awesome to see all of you guys there. And are you getting, is Google kind of, other than kind of,

• 28:10Daniel Whitenack?

  you know, featuring you and kind of bringing, bringing some publicity around it, were you able to kind of like foster some collaboration with, with Google on the, on the project?

• 28:21Speaker 02

  Yeah, definitely. There, some of the engineers at Google are a bit, are like our mentors. When we run into problems and we want to, we have different challenges. There are definitely a handful of engineers there that I reach out to, to get their thoughts. They're super helpful all the time. They respond really quickly. They're really, they really like this project and like to help us out.

• 28:44Daniel Whitenack?

  Yeah, that's, that's great. And I, I don't necessarily bring that up to, you know, to advertise Google, but I think it's great that they're supporting projects like this, which, like you mentioned before, there's a lot of emphasis on, you know, malicious AI and all sorts of things. But this is so awesome to see, you know, see a project like this and see, you know, kind of at least some level of support, you know, in, among major industry companies around, around this sort of thing, which is, is just really encouraging.

• 29:16Speaker 02

  Yeah, I definitely want to shout out to Pete Warden. Pete Warden's been super helpful and is a really great resource to us here at Plant Village.

• 29:24Daniel Whitenack?

  Awesome. Yeah.

• 29:25Speaker 02

  So we really appreciate his, his expertise.

• 29:28Daniel Whitenack?

  Yeah. Awesome. Awesome work. And Latifah, I want to follow up with you. I was wondering if you could just kind of describe a little bit, how, how is the, the app currently being used in the field? And what are, what are kind of the next steps to get, you know, cassava farmers using, all using the app? Which I don't know, does the app have a name? Um, maybe you can share that as well.

• 29:51Speaker 03

  Yes, um, the app is called Nuru, which is um Swahili word, name rather, female name. For um, it means light. So we use Nuru as a light that the farmers can use to um, look at their disea- I mean, look at their plants and their farming. So, yeah.

• 30:15Daniel Whitenack?

  Awesome. Yeah, and are are people using uh, using the app now in the field? Or is it kind of in uh, um, alpha beta version? Uh, do do people have access to it now?

• 30:26Speaker 03

  No, not yet. We are sort of in our beta version. And ideally, um, we will start, the thi- we can't get to every farmer, because not every farmer can afford a smartphone at the moment. So we start working with extension offices, agriculture extension offices, up to village levels who actually work with farmers. So if we get to one extension officer, we'll probably be able to get to ten to fifty farmers, depending on the size of the village. So our aim is to actually get it to the extension offices. And um, as more times go and farmers be able to

• 31:08Speaker 03

  get smartphones, then we'll be able to now get to farmers directly.

• 31:13Speaker 02

  So Latifa, um, they recently did a, an experiment with testing how Nuru did compared to scientists and extension officers and farmers in um, a district in Tanzania. So I don't know if, if Latifa, you wanna kind of share those preliminary results? I think that was really exciting to see that.

• 31:34Speaker 03

  Um, so it's quite interesting, actually, because they're experts. Um, so I did, um, we did a study with about ten experts, so groups of ten. And the experts, even though they're experts, their all, their ability to identify the different diseases also varies. And mainly varies on how exposed they are to the um, to the symptoms, to the variety, and um, how long they've been working in the area. And then when you go to extension offices, the agricultural extension offices is more or less the same. If they have been exposed to these diseases, then they're more inclined to know and understand them.

• 32:13Speaker 03

  But those agricultural extension offices who have not been exposed to the diseases, they actually are sort of clueless. And more or less the same for the farmers. So having a tool that it can um educate and aid the extension worker to be able to do his job and help the farmers. It's going to help a lot more than just helping them to identify the diseases. It's also teaching them and giving them more expertise while they are learning and using the app.

• 32:45Chris Benson?

  That is super cool. Um I guess as we start finishing up here, I wanted to ask all three of you to kind of hop in and and you know, we we we hear so much these days in AI about people's concerns and worries and and and uh the kind of the downside of of what AI may bring to uh to humanity and into the world at large. But you know, you guys are in the middle of doing some pretty amazing stuff. I just wanted to get a sense of what it feels like to know that you've been successful using AI in these contexts um and how how people react to that.

• 33:20Speaker 02

  Yeah, I think so. Yeah, I can start. I think it's, I mean, we're all so surprised, like, what? It worked?

• 33:30Chris Benson?

  Did you think it was a, did you think it was a long shot when you started out?

• 33:34Speaker 02

  I just didn't know. Nobody else, nobody else sort of did this very specific project. So a lot of the research data sets in computer vision have, you know, a thousand classes or eighty classes. And we were really honing in on a very specific classes, a very small data set, and then putting it on to a phone and testing it in a field. So it was very much applied and real. You couldn't get more real world than that. And I'm like, I don't, I don't know if it's it's really gonna work as well as all these super high metrics that people report in the research. So it isn't, in the real world, it isn't as high as ninety percent accuracy, but it's still pretty good and it's still useful.

• 34:15Speaker 02

  So um you know, that was, that's just great news. And I think that I'm I'm I'm just naturally a cautious person. I'm always sort of looking for the nuance. Okay, where the model does well, that's great, but where does it not do well? And where can we continue to improve and um? I would like that as we talk about AI in society and AI for good that we we start having much more nuanced arguments than, oh, it's gonna like destroy our lives or it's just gonna like make every change our entire world. I think we we need to just get more nuanced into how we think about technology.

• 34:52Chris Benson?

  Pete or Latifah, what do you think?

• 34:54Speaker 05

  Yeah, from my end, I I think I'm still I still have a streak alive where every single person I've ever told about this project has responded with, that's the coolest thing I've ever heard. I can't believe that like you're you're using your skills and your resources to like help the, you know, help these people in Africa and not just like creating cool filters for Instagram, you know, and so that's been like just the most rewarding thing for me to to know that like what I'm doing i- is really, you know, meaningful and like could could eventually like change the lives of of millions of people. But and

• 35:37Speaker 05

  I I think it's really important to to like reemphasize that like we are we're building a tool that helps people be better at their jobs and we're not building a robot to take over their jobs and do all the farming for them. We're just trying to help them produce more food and and live better lives that and and and sort of adapt to that. And like our our end goal is really to just like teach farmers and spread knowledge. And and so we'll, you know, they'll use our app, you know, for a couple of months or a year, a couple of seasons, and then hopefully they know, you know, what the disease symptoms look like and they don't need our app to show them. And so then we're like educating

• 36:18Speaker 05

  farmers and and eventually putting them off on their own. And so they don't, you know, rely on us anymore and they can provide for themselves.

• 36:26Chris Benson?

  That's super cool. Latifah, how about yourself?

• 36:29Speaker 03

  Um, for my side, this is like quite new, particularly in Africa, where when you talk about AI, technology, mobile devices, the mass are more aware of Facebook and Instagram. And when you have a tool that um can be used in a positive way and has a lot of impact in people's life, it's a plus plus for everyone. And what Nuru is doing at the moment is helping us as researchers and scientists share our knowledge with the farmers, with the end users,

• 37:12Speaker 03

  because we actually need them to see what we see. We need them to see the diseases that we see this. So we are using Nuru to help us portray our knowledge into the farmers so that they can be able to improve their farming practices using the research that has been done and has been proven. So when you have a tool like that, a tool that can integrate what the experts know, what the end users need, and improve their way of life, and it actually fits in in the way they lead their life, it's um it's a tool that can actually work and will be able to sustain in the system.

• 37:55Speaker 03

  And it paves a way into other researchers using AI and the same technology to also follow the same practice to generate products that can help people. Because at the end, you might have good products, but they do not fit in the way in people's lives. And now people will start getting scared that, oh, do you want to replace us? But now we are fitting the technology into people's life, and that gives them a bit of confidence, a boost, and also comfortability in actually using the tool. So I think we are moving to the right direction, and I'm so excited and looking forward to see where this can go.

• 38:32Daniel Whitenack?

  That's awesome. And I'm super inspired. I think I should just take that recording of what Latifah just said and make sure that everyone in industry listens to that at least once. So such a great and inspiring challenge and encouragement as far as, you know, the talents that we have and the way that they can make an impact, a real impact on people around the world is just super encouraging and inspiring.

• 39:08Chris Benson?

  And I think that person making the Instagram filters is crying right now.

• 39:14Daniel Whitenack?

  Well, maybe they'll just make an Instagram filter now that, you know, detects cassava plants and labels them.

• 39:24Daniel Whitenack?

  So the next tie in is the Instagram plugin. But I know, so just, we're kind of wrapping up, but I know before we jumped on the recording here, I know Amanda, you're talking, you're just super excited about next steps and, you know, really just viewing this as the beginning. Do you want to kind of close this out by just mentioning some of the future directions that you guys are excited about and maybe other challenges that people could get involved with as well?

• 39:56Speaker 02

  Yeah, so we're really excited about getting the app out to people, to farmers, to use, to test, to give us feedback on what works and what doesn't work for them. Building an app like this is a really iterative process. We don't really expect to get it right the first time, but we really would appreciate as much feedback. The earlier we get it, the better. And we're also going to be using this technology to diagnose diseases and pass in other crops. Crops like wheat, like corn, we're moving towards, potato is another crop that we're interested in. And we're always excited to hear from people about what problem they think that this technology would be perfect for. We're also continuing to publish

• 40:37Speaker 02

  our research. We're going to have a paper out soon with these results from testing the app in the field and providing much more metrics than what's conventionally provided to show people where the app does well and where it doesn't. Again, going more towards the nuance of performance in a real world setting. And yeah, that's about it.

• 40:59Daniel Whitenack?

  Awesome. Well, we'll post some links to some of those things in our in our show notes. Definitely, you know, keep up with with what this team is doing. It's it's awesome. Awesome to hear and just really appreciate you guys taking time to join us and it's just so great to talk to you. Thank you so much. You've really inspired us.

• 41:20Speaker 02

  Thank you. I'm always excited to talk about technology and agriculture and when people are interested in food and farming, that makes us super happy.

• 41:29Daniel Whitenack?

  Awesome. Well, thank you guys again and hope to hope to keep up with what you're doing.

• 41:35Speaker 02

  Thank you.

• 41:35Speaker 00

  All right. Thank you so much.

• 41:38Speaker 00

  All right. Thank you for tuning into this episode of Practical AI. If you enjoyed the show, do us a favor, go on iTunes and give us a rating. Go in your podcast app and favorite it. If you are on Twitter or a social network, share a link with a friend, whatever you gotta do. Share the show with a friend if you enjoyed it. And Bandwidth for Change Log is provided by Fastly. Learn more at fastly.com. And when we catch our errors before our users do here at Change Law because of Rollbar, check them out at rollbar.com/changelog. And we're hosted on Linode cloud servers. Head to linode.com/changelog. Check them out. Support this show. This episode is hosted by Daniel Whitenack and Chris Benson. Editing is done by Tim Smith. The music is by Breakmaster Cylinder. And you can find more shows just like this at changelog.com. When you go there, pop in your email address. Get our weekly

• 42:23Speaker 00

  email keeping you up to date with the news and podcasts for developers in your inbox every single week. Thanks for tuning in. We'll see you next week.