Intellyx Development Stories video for Zoho, featuring Ramprakash Ramamoorthy and Jason Bloomberg
In part 2 of the Zoho Creator – Intellyx Development Stories, Intellyx President Jason Bloomberg talks with Ramprakash Ramamoorthy, Director of AI Research, Zoho Corporation about the future of AI/ML as it will be operationalized for real business.
Click here to watch the video on YouTube: https://youtu.be/siBoHsOVWpo
The Future of AI & Machine Learning – Zoho Dev Stories #2 Transcript
Jason: Hello everybody and welcome to this video presentation by Zoho and Intellyx. My name is Jason Bloomberg. I’m president of Intellyx. And our guest today is Ramprakash Ramamurthy. So Ram, why don’t you go ahead and introduce yourself.
Ram: Yeah. Hi Jason. Hi audience. So, I’m Ramprakash (or Ram). I’ve been leading the AI efforts for Zoho.
We started off when AI wasn’t this hyped sometime in 2011. What started off as a very small experiment — today we have AI across our suite of products. So basically Zoho, we call it the operating system of business. Anything to run your business from CRM to help desk, to payroll, to books management, and all of the whole software stack.
We have, and I am responsible for shipping ai across this suite of apps. So broadly we have three divisions in Zoho where we do statistical machine learning, computer vision and natural language processing. So I have worked across the stack from statistical machine learning to natural language processing.
And at Zoho we are a very privacy-friendly company. So we have built and delivered a lot of AI modules with very good privacy levels and since we target a lot of small and medium businesses, we have also built models that are really small. Even with little amount of data the model will be as good as something trained on a huge amount of data.
So this is pretty much about me: I lead the AI stack with privacy attached and models that are really small. So even small and medium businesses can get started with AI.
Jason: So you mentioned statistical machine learning, natural language processing, computer vision.
These are all parts of the AI story. It’d be great if you could sort of clear up what is the difference among these. So I think a lot of people think that machine learning is. Some total of ai, but obviously it’s not. So I’d like your take on sort of the differences. Yeah. Well so statistical machine learning is more like you know what we call is.
Ram: It is like statistics with a fancy marketing team. So things like anomaly detection, forecasting, looking at patterns, and then trying to figure out the patterns or trying out, trying to figure out if it is a frequent pattern or a random pattern or an unseen pattern and things like that. So that is basically.
Statistical machine learning where you capture events, you try to correlate them, you try to forecast them, you try to find out root causes of events that have happened, in the current time. And then you look at the past to find out the root cause, things like that. So, statistical machine learning was the, one of the earliest ML things that came into existence because technically ML is all about looking at past data and then inferring something about the future or percent. So statistical machine learning is where even we started it. Today we have advanced it to things like multivariate anomaly detection, like your multiple variables, and we study how these variables interact with each other.
Then find out anomalies and then multi-way forecasting where you can do what- if analysis. So what if I price the product I increase the price of this product by five. Will I have the same revenue? Will I have the same number of customers coming in? So you can run a lot of simulations and then it goes on to things like digital twins and whatnot.
So this is about statistical machine learning and coming into computer vision. Basically since we do software for businesses, the place where computer vision generally they explain computer vision as cat and dog classifiers, but we don’t have space for cat and dog classifiers in enterprise software.
We are looking at it from a document intelligence perspective today, even though a lot of documents are generated digitally, but when they are trans transmitted across businesses, across business divisions they’re actually transmitted in the analog format, right? So you can have a PDF or you go buy something in a shop, and then they give you a receipt.
Even though that receipt is digitally generated, you have a hard copy of it. And then let’s say you’re expensing it to your company. You’re on a business trip and you’re expensing it to your company, you take a photo of it and then you upload it into an expense management tool. So I don’t think anybody’s a big fan of doing expenses, right?
So now we built a module where the receipts can be auto extracted by just taking a photo of the receipt, and then all the values can be populated. Now apply this to a variety of business use cases. It can be an invoice, it can be a purchase order, or let’s say you are running an insurance thing and you have ID cards coming in.
Now, all with privacy at forefront. This is one place -document Intelligence is one place where you put computer vision to best possible use, and of course other things like identifying objects. And then like let’s say when you search for an image, let’s say you go to a conference and you take photos of slides, and then you add it to your notes.
And now when you want to find out that slide where they talked about computer vision. So you just search for computer vision. That image turns. So these are places where we see the use of computer vision in enterprise software. And the last part, natural language processing. This is where it gets interesting.
So natural language processing. We have built a stack where we have grammar we have grammar correction, detection, we have translation. Now the challenge here is lot of times, lot of enterprise software user users use tools that are built for consumers to translate and to grammar check what they have written.
Now, there is a privacy angle. So you, you are using tools that are built for the consumer and you have sensitive enterprise information that is going out to these tools. So now when all of this is built within your enterprise software text editor, now, there’s a lot of value add to it. So we started off very humble in our NLP division with chat bots where, and now you ask a couple of questions and then it looks very frequently as the questions, the knowledge base you have in your organization.
Responses. And then when it is not able to understand a particular question, the the, the control is transferred to a human agent. We started off there we did some grammar correction deduction. We did invest on translation, and now we are working on things like voice to text. Know, let’s say we are on this meeting and a transcript of this meeting is generated from which you can you can find out to-dos.
You can find, you can actually give out minutes of the meeting, things like that. So this is how we are applying these three broad divisions to enterprise software. And we are seeing our AI usage grow up over the years. In fact, we’ve been around for 11 years doing ai, but in the last two, three years, there’s a.
AI adoption in the enterprise. Previously it was just a purchase checklist where, oh, this is an AI feature. Yes, go ahead. But now, especially post the onset of the pandemic, we are seeing our AI usage levels go.
Jason: Well, there’s a lot going on in AI, both at Zoho and in the industry at large. What, what do you feel as an expert in this space is the most interesting part of this AI story today?
I mean, there’s a lot going on. Machine learning may be not as interesting, but some of this natural language processing is, is more interesting than I thought. So what do you find the most interesting?
Ram: Yeah, Jason, so one thing I see is the hype has kind of died down. Right. So previously there were AI summers and then AI winters, but now I think AI has become one of the mainstream technologies.
And the other interesting thing I see, I think there has been a lot of noise around it. People are talking about dall-E, people are talking about GPT-3 which is technically generative AI, meaning some three, four years back, had you told me AI would become very strong in using the right brain aspect of things.
I mean, AI would get creative, but all we had looked at is things like robotic process automation where AI is trying to find out patterns and arrive at conclusions or automate redundant things so that your employee productivity is, is served to things where their attention is really needed. But interestingly, in the last two years, post this release of GPT-3 and Dall-E, We see AI getting more and more creative, right?
And, and in fact, the three divisions we talked about, computer vision and natural language processing being two of them. I see the gap between both of these blurring because we have a common transformer architecture where you know, both these CV and NLP, I mean computer vision and natural language processing models are built.
And now, for example, you take Dall-E, you give it a prompt, right? You give it a prompt in natural language, it is able to understand that and then it’s able to generate an image, which is again, a computer vision problem. So the gap between your NLP and CV is blurring, and AI is getting more and more creative, which is counterintuitive because computers are not creative, right?
So it’ll be interesting to see where it goes from here. How will we tap into this creativity offered by computers? What will it do to things? Arts and the, the creative field in general, it’ll be interesting to watch.
Jason: Yes. This notion of AI being creative, this is definitely point of controversy.
You get people on both sides of this argument. I mean, Dall-E, you know, where you put in a sentence, outcomes, some sort of original image, but it’s it’s original image based upon elements that were fed to it as part of its training regimen. The same with this AI that generates articles, right? We see, you know, articles or even novels.
Now AI can generate a whole novel. And then the question is, well, is that really being creative or is it assembling pieces from. It’s training material and there’s AI can generate computer code. It’s the same idea. Is it really just assembling pieces that may now be under copyright? And, there’s controversy on that point.
So the way I see it, the, the real question is when AI creates some sort of unique creation, so the a Dall-E image say, is it really being creative or is it just being trained to mimic creativity? So is it really something that humans exhibit or is it just. Doing better and better at pretending to do something that humans exhibit.
And I’m, I’m wondering what you think of that. Yeah, I think I mean the creative is just by looking at the results. But technically, if you look at it underneath, I mean, we have had gans for over the years, generative adversity networks where they look at past data and then. Try and mimic the past data.
Ram: Mm-hmm. now the evolution of things like diffusion models, which is what DA is based on. These diffusion models get so better at it that they look much better than humans. Right. So and like you said there has always been this notion of. Oh, AI is going to take over the world. AI is going to replace the humans.
But, but I would say it’s a natural evolution of software. For example 10, 15 years back when, when, not 15 years back, maybe even 30 years back when databases started evolving, it was superhuman to search through, let’s say millions of records in Subsecond. So even today it is a natural evolution of that thing.
We didn’t, we didn’t call it like, oh, this, this database is going to replace humans and all that. The same way today, software has evolved to, to understand text, to generate images. It is, it is only as good as your training data. So if you see, and then recently there was this lot of noise about AI becoming sentient and all that.
So whenever I read such articles, I always ask to myself, can mad multiplication ever become senti? No. And if you see things like Dall-E, the entry barrier to creating something like Dall-E is very high. You need tons of data, tons of comput to, to create something like Dall-E today, which is not possible for most individuals and most businesses right now because again, you need like terabytes of data, gigahertz of computing power stretched over months and so on.
What I would say is I see this as an natural evolution. Basically, these models have gotten better to mimic their training data and then combine these training data. For example, when I say when I say something like an astronaut, riding a unicorn, right? Mm-hmm. So it works. It has training data with images of astronauts and then with images of unicorns.
And finding the way to combine both of these is what this model’s uniqueness is. And coming to your second point, Copyrights and things like that. I say this to be commoditizing art, right? So for example, we had a, a photography photograph is an art, let’s say 50 years back. Photography was done only by professionals and then we had point and shoot cameras, and then we slowly, today we have iPhone cameras that are as good as your DSLRs and your professional camera occupants.
Now you see that that has been commodity test and just photos were taken and. We have an Instagram coming in where you can apply filters at, at, at, you know, at at, just by doing a touch you can apply a different filter. So again, photo editing was a professional art. Now, now that thing was commoditized.
So I see this moving towards commoditization of this art. So the entry barrier to art is reduced. So that is one way of looking at it when you see things like AI generating code. I see it to be, I don’t see it to replace programmers, but I see it to be augmenting programmers, helping them identify errors even before the code is committed to the repository.
Helping them proven mistakes by looking at their past data. Because any, any technology company has a wealth of. Code, right? Everybody has like tons of lines of code that is lying idle in the repository. So it only makes sense to put them into use and help the newer programmer who is coming in by generating stuff.
So I’m very bullish on how this technology can evolve and it can only help humans coming to the copyright angle. Yeah, that is important because we see things like deep. We see things like, you know, writing an article on behalf of somebody and then spreading it on social media. But actually that person would not have done it.
That business would not have done it. That country would not have done it. So again, we also see the evolving AI tools where we can identify piracy. We can identify that this person has not done this, so it is like a knife or a fire where on one side the problem is there. On the other side, it is helping a lot of things, but then there are techniques.
Put things in place and verify the authenticity of a particular image or a particular news material, or verify the authenticity of a source code before it is being deployed into production so that there are no corporate issues and then there are no other issues and all that. So overall, I’m bullish.
There are ways to stop all of these corporate infringement and all of that.
Jason: Yeah, so that’s the deep fake problem, right? Deep fake technology, it’s AI driven, it’s getting better and better. So now we can see politicians and other celebrities on screen looking perfectly real and natural, saying things they would never say.
And it’s causing a lot of controversy. We see it on social media all the time. And so we need AI to counteract that now to identify the deep fakes, but the deep fakes are getting better so that the, the technology to identify them has to get better. And we have. This, this sort of strange battle going on over these, the fake images and videos.
So just as a final question, since we’re running outta time, what is Zoho doing in this space? Is it doing anything with generative ai? Or, or, or where do you see it going if it’s, if it’s more of a future looking thing?
Ram: Yeah, we have a couple of projects going on in our labs. Nothing has been out yet, but this is an area we would like to be.
We would not want to miss the bus. So think of things like for example, there is a commerce where we enable very small and medium business sellers to use our e-commerce platform and then sell things. Now, professional photography for the products they sell might be very expensive. They will just have the photo of the, let’s say somebody selling some handcraft from India.
Now you have that handcraft and then you start generating a lot of backgrounds for it. And then you identify which pictures could get you much traction, which pictures could get you that conversion ratio. So things like that, insights like that in our commerce pipeline. And then we also have our CRM solution where we are trying to identify build something like a sales coach where.
How, how to better sell. Oh, last time you mentioned words like these the lead and convert. So maybe you should avoid using phrases like this. So things like that where it can help people do better at what they do. So that is where we see a lot of these generative use cases coming in and things like in, in our bi engines of analytics.
Now, Today we have Ask Zia, which is Zia as our conversational assistant, where you can go ask questions on your data, and then Zia can infer a lot of things. So in this use cases, now it can get more creative with, with your past data. So where else can I contribute? What will happen, like running? What if simulations on your data, what will happen?
Certain things might not be possible in real life. So running such experiments would be enabled with the help of such creative. Well, very good that I’m pretty excited about what’s coming down the road at Zoho.
Jason: So, so we’re outta time, so we should wrap up. Thank you very much to Ramprakash Ramamurthy, director of AI Research at Zoho Corporation.
I’m Jason Bloomberg, president of Intellyx, and thanks for tuning in.
©2022 Intellyx LLC. At the time of recording, Zoho Corporation is an Intellyx subscriber.
