Alter Everything Podcast

AlteryxMatt · ‎11-20-2024

In this episode of Alter Everything, host Megan Bowers converses with Naveen Krishnaraj, a senior AI scientist at Sabanto. They discuss Naveen’s journey from academia to the consulting industry and his current work in the agriculture field. Highlights include his development of machine learning models, the challenges posed by data cleaning, and the future potential of AI and quantum computing. Naveen also shares insights on his transition from long-term academic projects to rapid-delivery consulting roles, and his excitement for emerging technologies such as generative AI and quantum machine learning.

Panelists

Naveen Krishnaraj, Sr. AI Scientist @ Sabanto - LinkedIn

Megan Bowers, Sr. Content Manager @ Alteryx - @MeganBowers, LinkedIn

Topics

Ep 172 (YT thumb) (1).png

Transcript

Episode Transcription

Ep 172 AI Applications in Agriculture and Beyond

[00:00:00] Megan Bowers: Welcome to Alter Everything, a podcast about data science and analytics culture. I'm Megan Bowers, and today I am talking with Naveen Krishnaraj, AI scientist at Sabanto. In this episode, we chat about creating machine learning models in academia and consulting. His work with AI in the agriculture field and what makes him excited about future AI developments.

Let's get started.

Hey, Naveen, it's great to have you on our podcast today. Could you give a quick introduction to yourself, where you work, where you're located for our listeners?

[00:00:39] Naveen Krishnaraj: Thanks, Megan. I'm a big fan of the podcast - first of all just want to say that. About me. I'm currently a senior AI scientist at Sabanto it's a tractor automation company, so we primarily focus on building autonomy kit for tractors.

So that's my current role. I've been in the data space for close to three years now, and previously I worked at a firm called Cera. I worked as a consultant there, and before that I was in academia. I was a PhD student and I got my PhD in petroleum engineering from University of Houston. Also did my master's there as well.

And yeah, that's my whole short journey.

[00:01:20] Megan Bowers: Awesome. I'm really excited to chat today to hear more about that journey and all these different experiences that you've had with machine learning and ai. So I'd love to just start off with your background in academia and research. Could you tell us a little bit more about that?

[00:01:36] Naveen Krishnaraj: Yeah, absolutely. So I did my master's and PhD from University of Houston, like I said, in terms of my research background, right? I was working on signal processing behind, uh, NMR, nuclear Magnetic spectrum. So this is same physics as MRI. So. I think everybody who goes through MRI scan at some point in their life, right?

So this is the same exact physics, but we use it in a different setting, and that was a petroleum engineering department. So what we use it there for is in an ideal scenario, we use it to detect oil and water, how much oil we have and how much water we have, and like other types of things in the reservoir.

So that is a main reason or value of this work. I primarily focused on the processing set of things. So that was the overall work. And in terms of the motivation for the work, why we actually got it was actually initially a small part of what I was supposed to do, but then it turned out to be a whole thing, which I spent six years on.

So the reason for us was we were using two different processing software for this NMR, and both of them gave us different answers. So we were intrigued and we wanted to understand why it's giving. And we went and dwell deeper into the problem and understood what is doing on the backend and really un go dig deep into the math part of it.

And master's work was more about defining the overall problem, got an opportunity to work with really smart scientists and professors at University of Houston, and, uh, really was able to dig deep and understand the problem and frame the problem better. But we didn't really invent or develop anything novel there.

But then PhD work was more about developing a technique to solve the issues, which we found during my master's work. And we developed a technique for. Joint inversion and unsupervised learning. So that's two, uh, two technical words, right? I'm gonna explain what it is. Inverse problem, uh, is everywhere.

Inverse problem is throughout our life. GPS processing is an inverse problem. Basically, it's receiving sub signals and calculating a push, right? And, uh, there are various other, CT scanner is an inverse problem, so inverse problems are everywhere. And, uh, un voice learning is the, basically it's a clustering problem.

I, again, that's everywhere around us. What we did was we developed a technique which combined both these inverse problem and un voice learning and solid it in a single step. This way we save a lot of compute and we were able to generate a lot of practical use of it. We save a lot of compute and we also increased, uh, accuracy in various different cases.

So it's a fun, interesting problem, but this gave me the fundamentals to go from end to end, go into the details of mathematical, so. That, that was the whole fun part of it.

[00:04:34] Megan Bowers: That's really interesting. And how long did it take to develop that combination model that you're talking about for your PhD program

[00:04:42] Naveen Krishnaraj: To fully mature?

It took four years, but the actual grind of the work was like one, one and a half years. But there was a lot of things which didn't work before we were trying. It's always about trying and learning from them and pivoting. So it, it took some time, but I would say to fully match it, it took four years, the whole suite of the PhD.

And still a lot of things to do, but I graduated and I got up.

[00:05:05] Megan Bowers: Gotcha. Yeah. And so after academia, you moved into consulting, like you mentioned earlier, where you had to deliver these models much faster. So what did that shift look like for you and how did you navigate that challenge of having to deliver a lot faster?

[00:05:22] Naveen Krishnaraj: It's certainly challenging. I would definitely acknowledge that. It's a whole mindset shift. So in academia, like you already asked me, it took four years to dev, fully develop and mature something, and you are developing something novel and things don't get out until it's fully refined and you're pushing the knowledge edge a little bit.

You are not really like changing a lot of things, but you are definitely adding something small to the research community or the scientific community. So that's what PhD. Academia is about, but like moving into industry and moving into more applied machine learning or like consulting where I started out, it's a totally different mindset shift.

You don't have four years to solve something. You have only a few months and you need to generate value. So that is a whole mindset shift. There was, there were, the only way to learn is to by making mistake. Uh, but you need to learn from them. So, uh. Build a lot of mental models. How do you approach a problem?

What are your processes? You need to build that. And only way, only best way to build that is, is to be surrounded by smart people, which I had the fortunate of booking with some smart people at credit. And also be able to learn from your mistakes and be able to refine your process. So the, those are the initial struggles we But you, in terms of background wise, you already had that, I already learned a lot of math background and stuff like that.

From the academia, but scale is different. So cloud computing and those kind of things, I had to learn that. So those were the challenges and there were people to support us through the journey, but uh, you have to also invest some of your weekends and keep learning. So that's the only way to do it.

[00:07:06] Megan Bowers: And I'm curious, I think there was an article published a while ago now about data science and about how you spend 80% or so of the time just working on cleaning the data.

Did you find that to be true in your experience or do you feel like that's changed over time?

[00:07:23] Naveen Krishnaraj: Yeah, it depends on where you end up, but in any case, any role you have to still clean the data and do a lot of data engineering work. And in my case, yes, that was a big difference. There. The research was primarily focused on developing the algorithm side of things, so you don't really deal with a lot of messy data, and the real data is always messy, and that cleaning up takes a lot of effort.

But in an academic setting, we still tested it on real data, but we were able to get the data, which initially simulate some artificial data and kind of work with that. So. I wouldn't say 80%, but yeah, definitely close to that.

[00:08:00] Megan Bowers: Makes sense. Now that you've moved into your current role as a senior AI scientist in the agriculture field, I'd love to hear about what kind of projects you're working on with ai.

I think it's a super interesting application that probably a lot of our listeners haven't thought about before.

[00:08:17] Naveen Krishnaraj: Thanks for asking that. Yeah. In terms of my current role, the field as such is AG tech, but, but our, my primary focus is on automation, like routing problems. So basically what are the optimal parts the tractor needs to take?

So that, that is one of the problem which I work on. And also another problem which I work on is computer vision. So basically developing. State of art, computer vision models to be run on the tractor for, to detect and understand the scenes better, being able to detect optics so that, those are a couple of problems which I work on actively.

But also there are other data work like means putting in the pipeline and kind of ML ops work. So those things are part of it. But those are two high priority problems that we work on right now.

[00:09:03] Megan Bowers: For the computer vision problem, you're designing these algorithms to help the tractors navigate the fields or like what does it help the tractors do and what the value that people see out of that?

[00:09:17] Naveen Krishnaraj: Absolutely. Uh, in terms of computer vision, like you said, one aspect is safety or we already have a system. Which is based out of infrared arrays and it's able to use, it's called IFM, and we are able to detect objects in the field and we are able to stop. Pause on that. But what we are doing with the computer vision is we are adding to that.

We are adding one more layer of efficiency and safety because the IFM sensor, which we have cannot really differentiate different types of object right or computer vision algorithm can do. It can direct, there is an object there and can issue pauses and that is doing well and. We want to be able to differentiate those objects, right?

We want to be able to identify if a human is coming or there's an ML coming. So those, those are the things which we're working on, and the main value is improving the safety and also the efficiency of the system. That's the overall workflow. And we are, man, it's a continuously learning model. It's not like.

We put in something and then it's set there. There are constantly different objects we are seeing in the field and constantly the algorithm has to learn and develop more intelligence, I would say.

[00:10:30] Megan Bowers: That's really interesting. And so do, does that in turn like help farmers increase their yield or what's the value for people driving these?

Not really driving these tractors 'cause they're autonomous, but controlling the tractors, I guess.

[00:10:46] Naveen Krishnaraj: Yeah. For our customers, like I said, we, IFM sometimes has false pauses, so we are, we are trying to eliminate it. For example, it picks up a bird and it can pause on that. So we are trying to eliminate some of those.

Inefficiencies in the current system and also not that the current system is not doing very well. We have a lot of customer adoption, but we are trying to nudge the bar forward and we are trying to improve the efficiency. And also like we are trying to improve the safety. For example, if there is a new object, which is, which is getting directed, can we pick that up with all our perception system and not just computer vision, we are exporting other.

Sensors as well, like radar sensor. And there are other sensors which we are exploring, and computervision is a part of it. And the system is only gonna get better with more sensors and be able to handle different environments very well as well.

[00:11:41] Megan Bowers: So then more broadly, how have you seen AI change the field of agriculture while we've been working at this company?

[00:11:48] Naveen Krishnaraj: Before joining Sabanto I, I was familiar with like eel prediction and I was familiar with people using weed detection algorithms to elevate some of the weeds. And that's the computer vision side of things. And there are, I. There are more general, broader data science workflow just dedicated towards enhancing efficiency and also like the EO of everything boils down to that, right?

Generating more for the farmers. While at Ban, being focused on the tractor automation part, I've seen what other people are working on and uh, what kind of algorithms they are developing. A lot of the focus is on being able to do real time computer provision and doing it on edge devices and like. Being able to automize the missionary, which they have, be it any kind of missionary, not just tractor, which they use in agriculture and also like predictive analytics, like predict what's gonna happen in the future.

Those other things, which I feel like people are working on it actively. The main constraint is, is the edge side of things. If you had to do it real time, it has to be on a remote location, on edge device. So how do we handle that? So that is the main constraint.

[00:13:00] Megan Bowers: Can you define edge device?

[00:13:02] Naveen Krishnaraj: So think of you are pretty much in a remote location in somewhere in Midwest where you have very little connectivity and you cannot like upload the data to the cloud, get it back, and you have to react based on real time data which you collected there.

So what edge devices help you do is to be able to process a lot of information. On the location and be able to use that information. But it has to be resource constrained. Uh, it cannot consume a lot of power. It cannot have a lot of processing power as well, right? It cannot, it doesn't have the same scale as a cloud.

So it is a resource constrained environment. For example, we have a camera, it's an edge camera. It can process certain level of tops. Operations on the device. So you have to design a model based on the constraints which you have there. So based on how much data it can process in real time based on what or what frames per second, we need the data to be processed at.

So those are some of the constraints.

[00:14:03] Megan Bowers: That makes sense. It's interesting, I'm thinking of autonomous vehicles in a city area where there's great wifi connection and how you could do a lot real time with that versus this. It sounds like it's a lot more challenging to compute and model like real time when you are out in the agriculture fields.

[00:14:22] Naveen Krishnaraj: Yeah, and I think in, even in autonomous cars, they do it on the device. They do because it has to take addition and few microseconds, right? So the, and for example, it detects a person, it needs to stop immediately. So it doesn't have, I don't know, the going through the cloud and bring it back would be an efficient, I haven't really looked into the infrastructure, but I think everybody's working on it.

[00:14:47] Megan Bowers: Yeah. So it helps with the speed of decision and it makes sense with what you were saying on the safety piece earlier. If there's someone standing in the field or an animal in the field, it's gotta make that quick decision and be capitated super fast.

[00:15:01] Naveen Krishnaraj: Spontaneous. Yeah.

[00:15:02] Megan Bowers: Yeah. Okay. Super cool. So given your diverse experience with machine learning, academia, consulting, now, data science and agriculture, what makes you excited about the future of ai?

[00:15:16] Naveen Krishnaraj: That's a question I think about a lot. I can tell you that first Mostly. Mostly because this is very dynamic right now. Right? Uh, and it has changed quite a bit over the years. Currently the way the gold rush is right means generative ai, everybody is actually using generative ai, right? chat GPT, from image generation, multimodal models.

Those are, that's where the current focus is, and that is, I closely follow it. I closely keep learning all the tools and tricks and, and it's very exciting, I can tell you that, and it, it is transformative. To humanity in general. And you can see that from the Nobel Prizes, which people got for the deep learning has been recognized.

Uh, and people are looking at it as a beneficial to humanity and scientific society in general. So that is something which I'm tracking in terms of like short term, see how the field is evolving. We are still in discussion of using generated AI in our product as well. And, uh. That is the current buzz, but there are still a lot of noise around there as well, and I'm just waiting for the dust to settle down and see where this is heading.

It means is the L LMS the only route or not? But that's the, that's one thing. Short term, long term, my focus is on, I'm really excited about quantum machine learning, which is an intersection of quantum computing. Machine learning. And that is more a long term thing, which I'm keeping a track on. And I feel like that would significantly change the field and transform a lot of different industries like never before.

So,

[00:16:52] Megan Bowers: yeah, definitely. And what are the barriers that, that you're seeing right now to quantum machine learning? Like what is standing in the way of seeing that realized?

[00:17:02] Naveen Krishnaraj: Yeah, there, there are. Let me talk about quantum computing a little bit, right? It is. Which is, quantum computing in general is a little counterintuitive, so bits are the fundamental part of normal computing that is either zero or one, right?

That's the fundamental unit blocks of normal computers, right? Which are, which we currently use, but quantum computers are a little different or significantly different, where there is a super portion between zero and one. If you have a, a qubit, which is like. A super push in between a zero and one. So you can encode a lot of information in this qubit, and that is transformative in terms of using it for machine learning application.

The reason for that is you can process a lot of information. I. With a very little amount of qubits, but in terms of some, what you asked, right, in terms of some of the roadblocks there is the qubits itself. We don't have enough qubits working together in coherence to be able to generate value yet. I think roughly around a thousand qubits is, I think candor, IBM candor systems around a thousand qubits, but that's not enough.

The qubits have errors now and. People are actively working on error correction algorithms, which would make them more effective. But in order it to be practical, we need millions of qubits or a particular error collection algorithm, which for example, you can crack the RSA encryption with I think, 4,000 qubits.

The qubits shouldn't be noisy, but the current qubits are noisy, right? So we, we need millions of them to do error correction and be able to crack it. But yeah, that's, that. That whole research environment is exciting and means now it's not practical, it's all research. It's not applied. So just, there's something I passionately follow and just track it.

Uh, but I don't think it's anywhere close to being, like being able to productized. And there are certain aspects of quantum computing, which are being product productized. For example, there is a company called Dwe, which is working on a certain type of quantum system, which can be. Use for practical application.

I've seen people use it for particular application, but the real actual, full blown quantum computer is a little far. Uh, and hopefully some breakthrough happens and the AI people will start using it like more actively.

[00:19:23] Megan Bowers: It seems like that would be a huge breakthrough. Similar in scale to if we were to get to artificial general intelligence.

I don't know. I might put a poll on this episode to see which one our listeners think we'll get to first. But do you have a guess on, on what we would see first, whether it's like general intelligence from AI or quantum computing, or are they intertwined? Do we need one for the other?

[00:19:48] Naveen Krishnaraj: I think we need one for the other.

If we get something breakthrough in terms of quantum computing, it'll help us get to a GI faster. But in general, it's gonna have a. Both of them are gonna have a profound impact on humanity, but in particular, quantum computing, it'll help us a lot in the coming years in terms of firing climate change and being able to, for example, the habis process, which is a process of creating ammonia, which is one of the process which hasn't changed because we don't really understand the things at a deeper level.

How the. Interactions happens. So the quantum computer with we are able, we will be able to simulate like the quantum system much better. So we will be able to optimize the processes. So that is one use case how it can impact humanity. You might be able to predict the weather better or you might be able to control the traffic better.

So there are tons of other use cases other than just machine learning in general. So. That's something which is, which really excites me and hope somebody makes a breakthrough. I, I constantly keep a track on where the field is heading.

[00:20:56] Megan Bowers: Definitely super exciting for me as well. And thanks so much for joining us today, Naveen.

I know I learned some things from you and it was cool to hear about all of your experiences in the data science field. And if people want to stay in touch with you, can they find you on LinkedIn or what's the best place

[00:21:15] Naveen Krishnaraj: I'm really active on Linden, so feel free to shoot me a message or connect with me and thanks for having me, Megan.

I really enjoyed the conversation.

[00:21:25] Megan Bowers: Thanks for listening. To learn more about topics mentioned in this episode, head over to our show notes on alteryx.com/podcast. And if you like this episode, leave us a review. See you next time.

Ep 172 AI Applications in Agriculture and Beyond [00:00:00] Megan Bowers: Welcome to Alter Everything, a podcast about data science and analytics culture. I'm Megan Bowers, and today I am talking with Naveen Krishnaraj, AI scientist at Sabanto. In this episode, we chat about creating machine learning models in academia and consulting. His work with AI in the agriculture field and what makes him excited about future AI developments. Let's get started. Hey, Naveen, it's great to have you on our podcast today. Could you give a quick introduction to yourself, where you work, where you're located for our listeners? [00:00:39] Naveen Krishnaraj: Thanks, Megan. I'm a big fan of the podcast - first of all just want to say that. About me. I'm currently a senior AI scientist at Sabanto it's a tractor automation company, so we primarily focus on building autonomy kit for tractors. So that's my current role. I've been in the data space for close to three years now, and previously I worked at a firm called Cera. I worked as a consultant there, and before that I was in academia. I was a PhD student and I got my PhD in petroleum engineering from University of Houston. Also did my master's there as well. And yeah, that's my whole short journey. [00:01:20] Megan Bowers: Awesome. I'm really excited to chat today to hear more about that journey and all these different experiences that you've had with machine learning and ai. So I'd love to just start off with your background in academia and research. Could you tell us a little bit more about that? [00:01:36] Naveen Krishnaraj: Yeah, absolutely. So I did my master's and PhD from University of Houston, like I said, in terms of my research background, right? I was working on signal processing behind, uh, NMR, nuclear Magnetic spectrum. So this is same physics as MRI. So. I think everybody who goes through MRI scan at some point in their life, right? So this is the same exact physics, but we use it in a different setting, and that was a petroleum engineering department. So what we use it there for is in an ideal scenario, we use it to detect oil and water, how much oil we have and how much water we have, and like other types of things in the reservoir. So that is a main reason or value of this work. I primarily focused on the processing set of things. So that was the overall work. And in terms of the motivation for the work, why we actually got it was actually initially a small part of what I was supposed to do, but then it turned out to be a whole thing, which I spent six years on. So the reason for us was we were using two different processing software for this NMR, and both of them gave us different answers. So we were intrigued and we wanted to understand why it's giving. And we went and dwell deeper into the problem and understood what is doing on the backend and really un go dig deep into the math part of it. And master's work was more about defining the overall problem, got an opportunity to work with really smart scientists and professors at University of Houston, and, uh, really was able to dig deep and understand the problem and frame the problem better. But we didn't really invent or develop anything novel there. But then PhD work was more about developing a technique to solve the issues, which we found during my master's work. And we developed a technique for. Joint inversion and unsupervised learning. So that's two, uh, two technical words, right? I'm gonna explain what it is. Inverse problem, uh, is everywhere. Inverse problem is throughout our life. GPS processing is an inverse problem. Basically, it's receiving sub signals and calculating a push, right? And, uh, there are various other, CT scanner is an inverse problem, so inverse problems are everywhere. And, uh, un voice learning is the, basically it's a clustering problem. I, again, that's everywhere around us. What we did was we developed a technique which combined both these inverse problem and un voice learning and solid it in a single step. This way we save a lot of compute and we were able to generate a lot of practical use of it. We save a lot of compute and we also increased, uh, accuracy in various different cases. So it's a fun, interesting problem, but this gave me the fundamentals to go from end to end, go into the details of mathematical, so. That, that was the whole fun part of it. [00:04:34] Megan Bowers: That's really interesting. And how long did it take to develop that combination model that you're talking about for your PhD program [00:04:42] Naveen Krishnaraj: To fully mature? It took four years, but the actual grind of the work was like one, one and a half years. But there was a lot of things which didn't work before we were trying. It's always about trying and learning from them and pivoting. So it, it took some time, but I would say to fully match it, it took four years, the whole suite of the PhD. And still a lot of things to do, but I graduated and I got up. [00:05:05] Megan Bowers: Gotcha. Yeah. And so after academia, you moved into consulting, like you mentioned earlier, where you had to deliver these models much faster. So what did that shift look like for you and how did you navigate that challenge of having to deliver a lot faster? [00:05:22] Naveen Krishnaraj: It's certainly challenging. I would definitely acknowledge that. It's a whole mindset shift. So in academia, like you already asked me, it took four years to dev, fully develop and mature something, and you are developing something novel and things don't get out until it's fully refined and you're pushing the knowledge edge a little bit. You are not really like changing a lot of things, but you are definitely adding something small to the research community or the scientific community. So that's what PhD. Academia is about, but like moving into industry and moving into more applied machine learning or like consulting where I started out, it's a totally different mindset shift. You don't have four years to solve something. You have only a few months and you need to generate value. So that is a whole mindset shift. There was, there were, the only way to learn is to by making mistake. Uh, but you need to learn from them. So, uh. Build a lot of mental models. How do you approach a problem? What are your processes? You need to build that. And only way, only best way to build that is, is to be surrounded by smart people, which I had the fortunate of booking with some smart people at credit. And also be able to learn from your mistakes and be able to refine your process. So the, those are the initial struggles we But you, in terms of background wise, you already had that, I already learned a lot of math background and stuff like that. From the academia, but scale is different. So cloud computing and those kind of things, I had to learn that. So those were the challenges and there were people to support us through the journey, but uh, you have to also invest some of your weekends and keep learning. So that's the only way to do it. [00:07:06] Megan Bowers: And I'm curious, I think there was an article published a while ago now about data science and about how you spend 80% or so of the time just working on cleaning the data. Did you find that to be true in your experience or do you feel like that's changed over time? [00:07:23] Naveen Krishnaraj: Yeah, it depends on where you end up, but in any case, any role you have to still clean the data and do a lot of data engineering work. And in my case, yes, that was a big difference. There. The research was primarily focused on developing the algorithm side of things, so you don't really deal with a lot of messy data, and the real data is always messy, and that cleaning up takes a lot of effort. But in an academic setting, we still tested it on real data, but we were able to get the data, which initially simulate some artificial data and kind of work with that. So. I wouldn't say 80%, but yeah, definitely close to that. [00:08:00] Megan Bowers: Makes sense. Now that you've moved into your current role as a senior AI scientist in the agriculture field, I'd love to hear about what kind of projects you're working on with ai. I think it's a super interesting application that probably a lot of our listeners haven't thought about before. [00:08:17] Naveen Krishnaraj: Thanks for asking that. Yeah. In terms of my current role, the field as such is AG tech, but, but our, my primary focus is on automation, like routing problems. So basically what are the optimal parts the tractor needs to take? So that, that is one of the problem which I work on. And also another problem which I work on is computer vision. So basically developing. State of art, computer vision models to be run on the tractor for, to detect and understand the scenes better, being able to detect optics so that, those are a couple of problems which I work on actively. But also there are other data work like means putting in the pipeline and kind of ML ops work. So those things are part of it. But those are two high priority problems that we work on right now. [00:09:03] Megan Bowers: For the computer vision problem, you're designing these algorithms to help the tractors navigate the fields or like what does it help the tractors do and what the value that people see out of that? [00:09:17] Naveen Krishnaraj: Absolutely. Uh, in terms of computer vision, like you said, one aspect is safety or we already have a system. Which is based out of infrared arrays and it's able to use, it's called IFM, and we are able to detect objects in the field and we are able to stop. Pause on that. But what we are doing with the computer vision is we are adding to that. We are adding one more layer of efficiency and safety because the IFM sensor, which we have cannot really differentiate different types of object right or computer vision algorithm can do. It can direct, there is an object there and can issue pauses and that is doing well and. We want to be able to differentiate those objects, right? We want to be able to identify if a human is coming or there's an ML coming. So those, those are the things which we're working on, and the main value is improving the safety and also the efficiency of the system. That's the overall workflow. And we are, man, it's a continuously learning model. It's not like. We put in something and then it's set there. There are constantly different objects we are seeing in the field and constantly the algorithm has to learn and develop more intelligence, I would say. [00:10:30] Megan Bowers: That's really interesting. And so do, does that in turn like help farmers increase their yield or what's the value for people driving these? Not really driving these tractors 'cause they're autonomous, but controlling the tractors, I guess. [00:10:46] Naveen Krishnaraj: Yeah. For our customers, like I said, we, IFM sometimes has false pauses, so we are, we are trying to eliminate it. For example, it picks up a bird and it can pause on that. So we are trying to eliminate some of those. Inefficiencies in the current system and also not that the current system is not doing very well. We have a lot of customer adoption, but we are trying to nudge the bar forward and we are trying to improve the efficiency. And also like we are trying to improve the safety. For example, if there is a new object, which is, which is getting directed, can we pick that up with all our perception system and not just computer vision, we are exporting other. Sensors as well, like radar sensor. And there are other sensors which we are exploring, and computervision is a part of it. And the system is only gonna get better with more sensors and be able to handle different environments very well as well. [00:11:41] Megan Bowers: So then more broadly, how have you seen AI change the field of agriculture while we've been working at this company? [00:11:48] Naveen Krishnaraj: Before joining Sabanto I, I was familiar with like eel prediction and I was familiar with people using weed detection algorithms to elevate some of the weeds. And that's the computer vision side of things. And there are, I. There are more general, broader data science workflow just dedicated towards enhancing efficiency and also like the EO of everything boils down to that, right? Generating more for the farmers. While at Ban, being focused on the tractor automation part, I've seen what other people are working on and uh, what kind of algorithms they are developing. A lot of the focus is on being able to do real time computer provision and doing it on edge devices and like. Being able to automize the missionary, which they have, be it any kind of missionary, not just tractor, which they use in agriculture and also like predictive analytics, like predict what's gonna happen in the future. Those other things, which I feel like people are working on it actively. The main constraint is, is the edge side of things. If you had to do it real time, it has to be on a remote location, on edge device. So how do we handle that? So that is the main constraint. [00:13:00] Megan Bowers: Can you define edge device? [00:13:02] Naveen Krishnaraj: So think of you are pretty much in a remote location in somewhere in Midwest where you have very little connectivity and you cannot like upload the data to the cloud, get it back, and you have to react based on real time data which you collected there. So what edge devices help you do is to be able to process a lot of information. On the location and be able to use that information. But it has to be resource constrained. Uh, it cannot consume a lot of power. It cannot have a lot of processing power as well, right? It cannot, it doesn't have the same scale as a cloud. So it is a resource constrained environment. For example, we have a camera, it's an edge camera. It can process certain level of tops. Operations on the device. So you have to design a model based on the constraints which you have there. So based on how much data it can process in real time based on what or what frames per second, we need the data to be processed at. So those are some of the constraints. [00:14:03] Megan Bowers: That makes sense. It's interesting, I'm thinking of autonomous vehicles in a city area where there's great wifi connection and how you could do a lot real time with that versus this. It sounds like it's a lot more challenging to compute and model like real time when you are out in the agriculture fields. [00:14:22] Naveen Krishnaraj: Yeah, and I think in, even in autonomous cars, they do it on the device. They do because it has to take addition and few microseconds, right? So the, and for example, it detects a person, it needs to stop immediately. So it doesn't have, I don't know, the going through the cloud and bring it back would be an efficient, I haven't really looked into the infrastructure, but I think everybody's working on it. [00:14:47] Megan Bowers: Yeah. So it helps with the speed of decision and it makes sense with what you were saying on the safety piece earlier. If there's someone standing in the field or an animal in the field, it's gotta make that quick decision and be capitated super fast. [00:15:01] Naveen Krishnaraj: Spontaneous. Yeah. [00:15:02] Megan Bowers: Yeah. Okay. Super cool. So given your diverse experience with machine learning, academia, consulting, now, data science and agriculture, what makes you excited about the future of ai? [00:15:16] Naveen Krishnaraj: That's a question I think about a lot. I can tell you that first Mostly. Mostly because this is very dynamic right now. Right? Uh, and it has changed quite a bit over the years. Currently the way the gold rush is right means generative ai, everybody is actually using generative ai, right? chat GPT, from image generation, multimodal models. Those are, that's where the current focus is, and that is, I closely follow it. I closely keep learning all the tools and tricks and, and it's very exciting, I can tell you that, and it, it is transformative. To humanity in general. And you can see that from the Nobel Prizes, which people got for the deep learning has been recognized. Uh, and people are looking at it as a beneficial to humanity and scientific society in general. So that is something which I'm tracking in terms of like short term, see how the field is evolving. We are still in discussion of using generated AI in our product as well. And, uh. That is the current buzz, but there are still a lot of noise around there as well, and I'm just waiting for the dust to settle down and see where this is heading. It means is the L LMS the only route or not? But that's the, that's one thing. Short term, long term, my focus is on, I'm really excited about quantum machine learning, which is an intersection of quantum computing. Machine learning. And that is more a long term thing, which I'm keeping a track on. And I feel like that would significantly change the field and transform a lot of different industries like never before. So, [00:16:52] Megan Bowers: yeah, definitely. And what are the barriers that, that you're seeing right now to quantum machine learning? Like what is standing in the way of seeing that realized? [00:17:02] Naveen Krishnaraj: Yeah, there, there are. Let me talk about quantum computing a little bit, right? It is. Which is, quantum computing in general is a little counterintuitive, so bits are the fundamental part of normal computing that is either zero or one, right? That's the fundamental unit blocks of normal computers, right? Which are, which we currently use, but quantum computers are a little different or significantly different, where there is a super portion between zero and one. If you have a, a qubit, which is like. A super push in between a zero and one. So you can encode a lot of information in this qubit, and that is transformative in terms of using it for machine learning application. The reason for that is you can process a lot of information. I. With a very little amount of qubits, but in terms of some, what you asked, right, in terms of some of the roadblocks there is the qubits itself. We don't have enough qubits working together in coherence to be able to generate value yet. I think roughly around a thousand qubits is, I think candor, IBM candor systems around a thousand qubits, but that's not enough. The qubits have errors now and. People are actively working on error correction algorithms, which would make them more effective. But in order it to be practical, we need millions of qubits or a particular error collection algorithm, which for example, you can crack the RSA encryption with I think, 4,000 qubits. The qubits shouldn't be noisy, but the current qubits are noisy, right? So we, we need millions of them to do error correction and be able to crack it. But yeah, that's, that. That whole research environment is exciting and means now it's not practical, it's all research. It's not applied. So just, there's something I passionately follow and just track it. Uh, but I don't think it's anywhere close to being, like being able to productized. And there are certain aspects of quantum computing, which are being product productized. For example, there is a company called Dwe, which is working on a certain type of quantum system, which can be. Use for practical application. I've seen people use it for particular application, but the real actual, full blown quantum computer is a little far. Uh, and hopefully some breakthrough happens and the AI people will start using it like more actively. [00:19:23] Megan Bowers: It seems like that would be a huge breakthrough. Similar in scale to if we were to get to artificial general intelligence. I don't know. I might put a poll on this episode to see which one our listeners think we'll get to first. But do you have a guess on, on what we would see first, whether it's like general intelligence from AI or quantum computing, or are they intertwined? Do we need one for the other? [00:19:48] Naveen Krishnaraj: I think we need one for the other. If we get something breakthrough in terms of quantum computing, it'll help us get to a GI faster. But in general, it's gonna have a. Both of them are gonna have a profound impact on humanity, but in particular, quantum computing, it'll help us a lot in the coming years in terms of firing climate change and being able to, for example, the habis process, which is a process of creating ammonia, which is one of the process which hasn't changed because we don't really understand the things at a deeper level. How the. Interactions happens. So the quantum computer with we are able, we will be able to simulate like the quantum system much better. So we will be able to optimize the processes. So that is one use case how it can impact humanity. You might be able to predict the weather better or you might be able to control the traffic better. So there are tons of other use cases other than just machine learning in general. So. That's something which is, which really excites me and hope somebody makes a breakthrough. I, I constantly keep a track on where the field is heading. [00:20:56] Megan Bowers: Definitely super exciting for me as well. And thanks so much for joining us today, Naveen. I know I learned some things from you and it was cool to hear about all of your experiences in the data science field. And if people want to stay in touch with you, can they find you on LinkedIn or what's the best place [00:21:15] Naveen Krishnaraj: I'm really active on Linden, so feel free to shoot me a message or connect with me and thanks for having me, Megan. I really enjoyed the conversation. [00:21:25] Megan Bowers: Thanks for listening. To learn more about topics mentioned in this episode, head over to our show notes on alteryx.com/podcast. And if you like this episode, leave us a review. See you next time.

This episode was produced by Megan Bowers (@MeganBowers), Mike Cusic (@mikecusic), and Matt Rotundo (@AlteryxMatt). Special thanks to @andyuttley for the theme music track, and @mikecusic for our album artwork.

Alter Everything Podcast

Podcast Guide

172: AI-driven Agricultural Innovations

Panelists

Topics

This episode was produced by Megan Bowers (@MeganBowers), Mike Cusic (@mikecusic), and Matt Rotundo (@AlteryxMatt). Special thanks to @andyuttley for the theme music track, and @mikecusic for our album artwork.