Ganes Kesari on "Saving Our Planet's Biodiversity with AI"
Mohamed: Welcome. This is Expert Open Radio. I'm Mohammed El-Sayed, one of the people involved with organizing TEDx. Today we are here with Ganes Kesari who is a speaker scheduled at this year's TEDxAsburyPark conference, which is scheduled for May 18th in Asbury Park. Well, welcome Ganes.
Ganes Kesari: Hey. Hi Mohamed. Thanks for having me. It's a pleasure to join you.
Mohamed: To open things up, we would love to hear a little bit about your background and the topic that you're going to be speaking about this year at TEDxAsburyPark.
Ganes Kesari: So I have got 17 years of experience in technology, half of which has been in the data science space, working on machine learning, AI (Artificial Intelligence). My journey with data analytics started after co-founding Gramener about eight years ago, and we help enterprises and NGOs (Non-governmental Organizations) tell stories from data. Broadly I'm passionate about data and how AI can be beneficial for us. And the title of my talk at TEDxAsburyPark this year is "Saving Our Planet's Biodiversity with AI."
Mohamed: How did you come up with the idea, and what got you really interested in focusing on biodiversity and the application of AI to biodiversity?
Ganes Kesari: Pretty whenever I talk to people, there are just two views about AI. It's either seen as the Satan or the Savior. So, unfortunately, there is no middle ground. Often, in my conversations, I try demystifying the concept of AI and how it can be beneficial and also how it is just another technology. Just like any new technology, people are anxious about it but it can be beneficial just like a calculator can help us make calculations faster, AI can help us in a variety of ways.
As part of our work, we have been partnering with Microsoft and about two years back we started working with a team in Microsoft called AI for Earth that tackles environmental problems using technology. They award grants to NGOs and universities who solve some of the toughest environmental problems facing our planet.
Under this partnership, we've been working with several NGOs around the world for a variety of conservation initiatives, from identifying salmon, detecting a variety of a species, about 5,000 plus species around the world, to spotting African elephants or content penguins in Antarctica. All of these are fairly interesting and fascinating problems and I've been really drawn into this work based on the impact it has on our environment and the potential it holds for the future.
So that was a background in terms of how I got into conservation of biodiversity and the application of AI into it. And when I came across TEDxAsburyPark, the event for this year, I thought that this would be a perfect one talking about the chaos our world is in now and how technology and AI can be a very useful aid in solving this problem.
Mohamed: In your experience, when you talk to people about this topic and the work you're doing, what exactly you think has attracted their attention and resonated well with them? Is it the biodiversity part of it, or is it the AI or is it something else?
Ganes Kesari: It's actually a combination. Firstly, when we talk about AI and actually simplify it and give some examples, people actually get it. And once they have more familiarity with the concept, with the technology, they can see its potential usage. So the usage of AI and its part is something which fascinates people.
And number two is in terms of biodiversity conservation. We hear about it I think for several decades now, we've been talking about it and perhaps for a few centuries. However, most of the initiatives that have been taken up so far have not given us great results or have not scaled. When we see how AI can really solve this problem, and it can be point and click and readily scalable, that attracts people, and they say, "Okay, why not apply this to some other area which is important to them, something very close to them." Both of those aspects interest people.
Mohamed: So to make it a little more concrete, can you give us a specific example on how AI has helped with conservation efforts?
Ganes Kesari: I think of one example from two years back when we started working with an NGO based in Washington state. It was called the Nisqually Foundation, so they monitor the movement of salmon in the waters in the Nisqually River. They have underwater camera traps, which take short bursts of video worth anywhere from 15 seconds to a minute. And these videos are reviewed by trained biologists to identify the species of salmon.
Firstly, what is the fish which has passed through the camera trap and what species of salmon it belongs to. It's a heavy manual effort because there are false triggers. At times there's no fish in the video. At times it’s just a minute long and you have a couple of seconds of sighting. And it is not scalable, and when we tried applying AI to it, we showed 1000 videos of the salmon to this AI model. And with that learning, it was able to detect and accurately identify the 12 species of salmon, and it drastically cut down the effort into about 80 percentage of savings in effort. And it was a scalable model which could be readily replicated around.
This is one example of how a using computer vision and application of AI in that area, we can detect species and be able to automatically identify not just the species, but also there are other examples we can identify the individual animal aspect.
Mohamed: But maybe let me step back a little bit and get a little more technical here if you don't mind. What's the difference between taking those video images or video clips of salmon and doing some pattern recognition on the video image and the use of AI. Is AI any different than just maybe doing some pattern recognition which has been in use for awhile?
Ganes Kesari: There is an overlap. Let me explain. So firstly, AI, in simple terms it is technology, which can see, hear, think or write or interpret basically getting some intelligence, mimicking the intelligence of humans in some aspect. So narrow AI is something which can do one or a few of those things. Be able to see something, be able to recognize, like pattern recognition that you mentioned. All of these can fall into that purview.
And some aspects of machine learning have been around for awhile, and at the risk of getting more technical, the real change that we're seeing now is that this branch of machine learning called deep learning. It's in the last five, six years. It's becoming very popular where you just pass examples. If we take the case of say face recognition, so we have human faces, how do you recognize that? For us, once we look at people a couple of times we are able to identify the patterns and recognize them the next time.
So how, how can machines recognize people? So a couple of decades back the effort was to painstakingly identify the features of the face and put that together and train machines so that whenever you see a similar combination you can say that this is the closest match. That's how intelligence agencies have been working for probably a few decades.
The recent change with advances in AI is that you don't get into that painstaking definition of features. Instead, you just show say 100 or 1000 such faces of a person, different snapshots, and you let the model figure out the pattern underneath. You don't encode or you don't give any specific instructions. You just show the example and you say, "This is Mohamed", and you show a series of pictures. And then the machine figures out how to identify Mohamed when a new picture is shown. That intelligence is what the machine gets after this initial pass through.
Similarly we can apply this to a variety of features just like human faces. We can apply this to objects and extending that to animals. So we will be able to spot, detect and even recognize any of those objects or people. So that's how this AI works.
Mohamed: Well that's very exciting. I can see the huge potential here given this concept of deep learning. But then the amount of data is ... I assume it's going to grow exponentially. The potential is unlimited here. But I assume that the scaling effort will have to even get better and better in the future to deal with that, with the potential opportunities and the huge amount of data. So what do you think is ... what are the advancements in applying AI to biodiversity that will allow you to scale even better in the future?
Ganes Kesari: Yeah, so like you rightly mentioned, the potential of data is huge, and we are seeing this explosion of data in various formats. So in social media we have people posting pictures. So tourists when they visit different parts of the world or different forests or they go boating, they have pictures, snapped or videos, which they post to YouTube.
So all of this is data which can be used by the models. And there's one organization called "Whale Book," which uses this to to conserve Whale Sharks as an example. Whale Sharks have this unique fingerprint of spots on their bodies. With this pattern, AI can uniquely identify two Whale Sharks. So when AI is trained on say hundreds of thousands of such pictures, it gets that ability to spot Whale Sharks and also uniquely identify an animal within the species.
So now what this organization has done is they are letting the AI scour social media, watch millions of YouTube videos and identify which are the relevant sightings and automatically catalog and say that, okay, this is the Whale Shark I last saw off the coast of China and then now I'm seeing this in this other ocean.
So that way they are able to match the fish across different sightings, across disparate videos over time, and they're able to build this catalog of, you can say almost like a census of Whale Sharks around the world. So to date, they have uniquely identified about 10,000 Whale Sharks purely from such sightings and social media pictures.
With this, for the first time, they have got almost a good count of the number of animals left in the species based on the sightings we've got so far and they continue to build. You can imagine how this can be scaled across different types of animals. They've done this for giraffes and for other animals, and very soon we'll have a reliable count of the populations of different animals, and also if we're talking about a unique animal and what's the lifespan. Is it spotted again? All of that as possible using this technology and you can imagine how this would help us in terms of conservation and monitoring the health of species.
Mohamed: This is great. How can we all contribute to this idea to make a difference?
Ganes Kesari: There are a couple of ways we can all be a part of this conservation initiative. One is as simple as sharing of pictures and videos. In this early example we saw how AI is able to pick up videos from social media. So sharing more of those sightings and videos is one way to contribute.
Another is in terms of contributing and volunteering time to train the model. So the model initially has this training phase where it needs enough pictures, say about a thousand or 2000 pictures where you need to identify an animal, let's say a penguin. We need to teach the model what a penguin looks like, so that we either draw boxes or draw labels around the animals in pictures, and that's a painstaking effort which is done initially, after which the model picks up and scales over time.
So that is an area where there are what they call citizen scientists, the general public volunteering time whenever they can spend 15 minutes or 30 minutes, they go onto the site and then just label the animals to help train the models. So these are ways in which people can be a part of this effort.
Mohamed: That's an easy thing to do. Posting pictures in social media, that's a very good way of helping out. Are there any specific requirements on those pictures or specific animals that you think are more useful than others?
Ganes Kesari: There's a list of endangered species and pretty much a lot of animals have moved on to that list unfortunately. Getting onto that list and sharing pictures of those sightings. So most of the Whale Sharks, or it could be lions, giraffes from the safari pictures. So all of those could be potentially useful, so any animal which is at the risk of extinction is an ideal candidate.
Mohamed: And do they have to be of high definition quality or ...?
Ganes Kesari: No, not at all. So it could be just pictures from our phone. Any low resolution snapshots, the model is able to identify and pick up from that, and if it's something that is really not at the level of quality, the model is able to differentiate as well. So there's no minimum requirement. Anything can be contributed.
Mohamed: Are there any books or references or authors you would recommend to our listeners if they want to get some additional perspective on this topic?
Ganes Kesari: One place to start, there aren't many books. There are good papers on journals on conservation using AI. Google searches will bring that up. And another place I would suggest is Microsoft AI for Earth. Like I mentioned they offer grants and they have funded 200 NGOs and universities so far. So I usually check the page, and I find a lot of interesting work. Many of the examples I've quoted are also mentioned there, and they also have these public models which can be used by organizations. So assuming we want to identify an animal from our own pictures, there's a place where we can upload pictures and identify what the model says as the animal species. So the Microsoft AI for Earth is a good starting point, which links to multiple such organizations working in this area.
Mohamed: In closing, any other thoughts about your talk that you'd like to mention to our listeners here?
Ganes Kesari: I'm looking at structuring this talk in the three aspects that broadly we spoke about it. First, part is in terms of giving examples in terms of why this is a big problem today and it terms of the chaos our world is getting into.
And second part would be in terms of “what is AI, how does it fit into this picture?”. And third would be real world examples, like the one we spoke about here, and showing what is the way forward and how all of us can be a part of it. So that's what I'd be talking about. I'm fairly excited to be doing this and yeah, I look forward.
Mohamed: Well, Ganes, thank you so much for being with us today. You have been listening to Expert Open Radio, and here's a reminder to get your tickets for the largest, highest rated TEDx conference on the East Coast. It's TEDxAsburyPark on Saturday, May 18th, 2019 and you'll have an opportunity to hear Ganes Kesari's talk about “Saving Our Planet's Biodiversity Using AI” and go into more detail. Thank you again.
Ganes Kesari: Yeah, thank you, Mohammed.