Read the transcript of my podcast with Bharath Sudarsan of SomaDetect. In this inteview, we discuss the internet of cows, how AI and machine learning are used on a dairy farm, the type of data that comes from every cow and how that data is used to drive decisions. We also discuss the black box of AI, and what's ahead for Artificial Intelligence and the future of mankind. Give it a read below:
Peter Schooff: Hello, this is Peter Schooff, editor of Data Decisioning and today I am very pleased to be joined by Bharath Sudarsan, the founder and director of artificial intelligence for SomaDetect. And that's pretty much what we're going to focus on in this podcast: exactly how AI and Internet of Things is contributing to companies at their operational level. In fact, I think I just recently saw that you guys referred to the IoT as the Internet of Cows, and I definitely want to hear about that. So first of all Bharath, thank you so much for joining me today.
Bharath Sudarsan: Thank you for having me, Peter.
PS: So can you tell me a little bit about what does SomaDetect do? And can you tell me about this internet of cows?
BS: Well, so we're trying to get as many animals online into our system and on the internet, so we just dubbed it, jovially, as the internet of cows. What SomaDetect essentially does is we try to make the best milk possible by supporting the farmers. So the idea is to connect and empower the entire dairy system from every single cow on a farm all the way to the consumer who buys the milk in a grocery store.
PS: That's pretty cool. Now you first showed up on my radar with a quote, your model accuracy is directly proportional to how good your annotation is. How would you boil this into the real world? What does this mean for you day to day?
BS: You know, interestingly, when I was back in grad school, my professor always used this acronym GIGO. Garbage in garbage out. So essentially, when it comes to artificial intelligence, the computer doesn't necessarily know what I am asking it to see. It's going to say, all right, this pattern of data that you're giving me, whatever you want to call it, is it matching that label or not? In other words, I can have a bunch of images of let's say, a cat, and a bunch of images of like, a dog. The word DOG doesn't mean anything to the computer. It's just that all these pictures are similar. And since you call it a dog, I'm going to say that picture looks like the word that you're calling it, right.
So the quality of annotation that goes into your dataset is going to determine how biased your system is. There was one real world example of Google building a classification algorithm that helps differentiate between different types of shoes, dress shoes, and casual shoes, right? So they ended up crowdsourcing the information from a lot of people uploading pictures. And interestingly, most people think of Converse as a sneaker as a casual shoe. There are other versions of casual shoes, but since most people in North America seem to have that type of casual shoes, now you're data set is biased towards Converses or sneakers as casual as it doesn't really consider anything else as a casual issue.
So how good is your model? It completely depends on how good is your data set? And how well does it represent the real work? So that's what I meant by saying, it is directly proportional to the quality of your annotation.
PS: Yeah, so that's really important. Now, you kind of touched on this already, but I guess were pretty cow focused on this podcast. So how does your data go to a single cow? What data do you gather from a cow to help drive quality?
BS: Right. So when we get down to the nitty gritty of how SomaDetect works. It is a small sensor that has laser. And this laser light is going to hit milk as soon as it comes out of a cow. So our sensor is strategically placed right under a milking station in every farm who are our customers. So when a cow is being milked, our milk will go through a vacuum tube all the way into this huge tank called the bulk tank. And in the middle, we have one of our sensors, which is going to be in line in the farm. And as the milk leaves the cow, it goes through our sensor and then reaches this bulk tank. So when it passes through our sensor, laser light hits the milk, scatters, and the camera captures this image of how the light scatters. From that we derive information about the quality of that milk. In other words, the health status of that cow. We analyze this on the cloud. And for every farm that is our client, we know how every single cow is doing every single time she gets milked. So every cow at every milking, SomaDetect help keep tabs of her health. That is how we have data for the cows.
PS: Cool. So now you're talking to Data Decisioning. So what are some of the decisions you make based on this data?
BS: Oh, there are plenty of decisions that you can make. So firstly, farmers are paid based on how excellent is their milk quality, right? But how do you define quality—the quality is directly proportional to the amount of fat or the richness, because from milk they make butter and all sorts of cream and other dairy products. So if your milk is mostly water, it's not exactly a rich milk now is it. And then, from milk, you could also get proteins, like whey protein, etc. So how much fat is there? How much protein is there? Are there somatic cells in this milk? Which is a corollary to how sick the cows are. It's the immune response of the cow.
So if somatic cells are high, then the farmers are fined, or the pay extra to dilute that somatic cell down. So bonuses versus fines, these decisions could be made by farmers by figuring out as simple as, "Okay, my cow, Betty, has extremely high somatic cells. If I mix that milk with the rest of the milk from the healthy cows, I will be fined. So today, I'm going to just pour her milk on the floor, because it's going to poison the rest of the rich milk."
That's a decision, right? And then which cows are high performers, and you want to take care of your assets that give you the best return on investment. So which cows are producing rich milk, feed them better take care of them better make sure they are kept. And which food are you feeding them—what type of feed gives the best optimal result. So it's cost benefit ratio, right? You eat tons and tons of food very less rich milk, very little amount of food, and hi rich milk. So which feed is giving you the best return on investment? That's a decision you can make. And pregnancy.
Last but not least, if a mammal is not pregnant, she doesn't give milk. So you don't want to miss the pregnancy. Can you detect heat in a cow efficiently and make sure she doesn't lose her babies? Can you protect the calves? These are extremely important decisions that a farmer can make to make sure his farming operation is running smoothly. So pregnancy detection. These are all the things that SomaDetect provides the farmers to make sure they can run a smooth operation.
PS: Well, that definitely sounds like the Internet of Cows to me if there ever was one. Now, you're pretty much the first director of AI we've had on Data Decisioning. What do you see in the future of AI? And even what do you see today in AI?
BS: That's a great question. Um, the word artificial intelligence, I've heard it from the days of the Terminator just, you know, inspires fear, World domination. Oh, they're gonna take over, etc, yada yada. I would say, right now, as artificial intelligence stands, there's literature that says that it's as smart as an eight month old baby. Give it a chance. We are the engineers who are the forerunners for building it based on our own image. So if we take the responsibility of the tasks that we have at hand, I think the future for AI is going to be a bright future. And I also believe that the human potential is meant for far greater things than just waving a flag or flipping a burger.
With time, we can invest our intelligence in order to make the planet better, or traverse the stars—whatever it is, the sky is the limit, if not space and infinity itself. If we are distracted with mundane tasks, like reading plans, or flipping burgers, which any machine can do, then we're not going to have the time or the freedom to work on these things. I think AI is going to support mankind itself as we move forward.
PS: I agree with you completely about that. Now, a lot of people recently have been discussing the black box of AI, you know, you put something in and you don't necessarily know why it does what it does. So how important do you see the transparency of AI?
BS: I would go back to the human condition, where the human condition is the need to understand everything to the minutest of details. I get that. I myself am a research scientist, and I want to know why something ticks. But at the same time, the way I perceive, let's just say, the taste of an apple might not be the same way you would perceive the taste of an apple. And I have no way to fully understand how you actually perceive it. All I can say is that it's sweet? And that's what you're going to say.
So if we take the perspective of image recognition: How do I see a dog? And how do I know this dog is different from a wolf? From childhood, I could have learned that gray wolves grey colored wolf like animals are mostly wolfy, maybe they are a little bigger than dogs. So that's how I make the association and the difference. Someone else with different life circumstances might learn something different.
Similarly, a convolutional neural network, for instance, that recognizes a dog based on its features and filters, probabilisticly, is going to figure out an algorithm where it says—you know, I'm going to simplify this significantly—but if it is this type of color, and the nose is yay far away from the bridge of the ears, and the ears are floppy, it's probably a dog. If the ears are perky and sharp and it has a nose like this it's going to be a wolf. We may not understand how it works, but as long as it has been tested and it works in a real world condition where your data set this representative of the variation that you see in real life. Does it truly matter understanding this black box completely?
Testing is what we need to make sure that the black box is robust. I think we should focus more on how well can we test this black box and how well we can rely on it, the reliability of it. When you think about single pixel attacks, where when a strategic one pixel is changed, the AI algorithm is not able to recognize something that it was recognizing with 99% accuracy—if that one pixel in that image is not changed, well, now that is scary for a self driving car. However, understanding how it perceives a human. I don't think that is as important as putting our energies and efforts into testing procedures and making sure these algorithms are robust and reliable.
PS: That's an excellent answer. That's a really good way to summarize it. Now, this is a huge topic, and we're really right at the cusp. But if if you want to our listeners to come away with one takeaway from all of this what would you want that takeaway to be?
BS: Artificial intelligence is coming. There's no stopping it. We should embrace it. We are the first generation of developers, programmers, scientists and users of artificial intelligence. Honestly, I commend and respect anyone for their courage to take on a task to bring an AI into the real world today, right now. Being the forerunners, I sympathize with them for the lack of support when it comes to the tech stack, etc. Because everything is very nascent, and everything is new, untested. A lot of this is r&d technology that just breaks. I would say, let's all be aware of the responsibility that we have, and build a world and show people how it is done when such a technologies brought to the real world by responsible capable people. Let's do that.
PS: I agree completely. You know, I'm in the news side of things, and we tend to sensationalize stuff, and we get clicks by talking about Terminator. But we can be extremely hopeful too.
This is Peter Schooff of Data Decisioning speaking with Bharath Sudarsan of SomaDetect. Excellent, excellent points. Really exciting time right now. So thank you so much for joining me today.