Podcast Host – Madhura Gaikwad, Synerzip

Podcast Guest – Vinayak Joglekar, CTO at Synerzip | Gaurav Gupta, AI Expert and Senior Engineering Manager at Synerzip

Summary

Artificial Intelligence and Machine Learning are becoming mainstream in the post-pandemic age. Businesses are leveraging them to resolve unimaginable life and business challenges.

In this episode, we bring in a seasoned AI professional as a guest along with our CTO, Vinayak Joglekar, to unravel the advancements in the implementation of AI and ML that include :

• AI democritization
• Availability of Structured Data
• Active learning
• Reinforcement learning
• A lot more…

Transcript

Madhura Gaikwad (00:08):

Hello, and welcome to zip radio podcast powered by Synerzip. I’m your host Madhura Gaikwad. And I am joined today by Vinayak Joglekar CTO at Synerzip and Gaurav Gupta Senior Engineering manager at Synerzip. Gaurav is an AI/ML expert and has agreed to join us today to discuss the top technology trends and predictions for 2021. In the last episode we discussed cloud is one of the top technology trends in predictions for 2021. Vinayak shared some great insights on the acceleration of cloud adoption hybrid cloud edge computing, and more today, we will deep dive into the second most trending technology that will take center stage in 2021, which is artificial intelligence and machine learning. And this, we have identified based on our survey of 170 plus CXOs and technology leaders, as well as our observations from client interviews and research documents for listeners wanting more information on these trending technology. We have also curated an eBook that covers the nine technology trends and predictions for 2021. In detail, you can find the link to download this eBook in the description, so welcome onboard Vinayak and Gaurav.

Madhura Gaikwad (01:15):

So, my first question to both of you is that there is a lot of progress in artificial intelligence and machine learning algorithms. These algorithms are easily available on public platforms, such TensorFlow and AWS. So previously very few had access to such ready algorithms, but there has been a very big shift in 2020, and AI democratization is one of the top contenders in 2021 predictions. So how will all of this pan out in?

Vinayak Joglekar (01:44):

Yeah, so this is we going to see more of it. If you see the number of people who were not actually data scientists, they were just developers who have picked up, you know, by doing south courses or Udemy and coursera and stuff like that. They are pretty much up to speed on how to use algorithms that are available in the public domain as you likely pointed out. So AWS has now Sage Maker. So another thing that has got democratized is availability to computing. So there are some of these algorithms that it’s not only the algorithm that was to find or hard to get at, but it was also the computing power that is needed to support it, that wasn’t so easily available. But today you have on the cloud, all the computing power and very quickly you can put together the needed computing as well as the algorithms.

Vinayak Joglekar (02:43):

So that is to move to the next level. So you have a lot of GPUs available on the cloud, and that is going to change the landscape where it’s not something which is resort for a few people sitting in the ivory towers. This is going to be completely democratized. So it is just like going to be a part of the standard development and programming, just like what you have on the UI or database machine learning, artificial intelligence is going. So the bigger problem, and I’ll come to that problem later is that data, I, I mean the processing and learning and algorithms, and that part is pretty much available. What is the problem that we are going to have is the availability of data and secondly, experience in putting these things in production and putting them to actual use and not just having these as laboratory experiment, you are just proving that these things work. And I let Gaurav tell a little bit as to what he sees this, where he sees these things going, but in my opinion, it’s not going to be the algorithms or the computing that is the bottleneck. It would be the data and ability and experience of people in these models in product that’s where we are going to see the bottleneck. But I’ll let Gaurav speak what his opinion is. So, Gaurav, what do you think?

Gaurav Gupta (04:07):

Yeah, Vinayak. As you said, rightly this is going to be majorly on the data side and also on the, how we productionize these models and monitor those models. So I think 2021, the major focus is going to be the MLOps area where we are going to manage the models, build the models, and monitor the models, how to build that infrastructure, because that is the major pinpoint in any machine learning work, which has to be done. And as these algorithms are getting democratized one major benefit, which is going to come down the line, that domain specific understanding, which is required for machine learning slowly, that may move out where you may use any domain data and algorithms are good enough to find out the features for your domain instead of you defining the features. So that kind of democratization for domain agnostic machine learning algorithms is going to happen. And MLOps is going to be a big deal in the 2021.

Vinayak Joglekar (05:07):

Yeah. So talk later domain agnostic is something which kind of makes it very, very interesting because particularly for people who are in a country, such as India, and we are trying to support our clients who are all over, mainly in the Western countries like United States and Europe. So we don’t have the context or the domain knowledge that is required to support modeling. So for building that model domain knowledge is needed today. What Gaurav you are saying is even that may not be required going forward. So about that a little, but Madhura, you had conducted some sort of survey and what are people saying.

Madhura Gaikwad (05:49):

Right. Yes, Vinayak. So based on the survey, one thing that we’ve observed is that there is so much progress that is expected AI/ML, and that is very much you know, visible through the survey. And the surveys also observed that 37% of the respondents, which were mostly technology leaders and CXOs said that AI/ML were among the top scale gaps that they foresee in 2021 that will affect their product go back.

Vinayak Joglekar (06:18):

Oh, that is interesting. Yeah. Skills, availability of skills is also another issue that we will have to overcome. So, you know, in my opinion, it’s not rocket science, so it’s just the way it happens with every skill that is on the horizon. You know, initially to start with is always this skill gap and for which people have to scramble to hire. But very soon you’ll find that lot of new talent is attracted to this hot field, which is very promising, which is very remunerative, and people will pick up. So there will be a lot of availability very quickly. So, by that time we start productionizing and doing MLOps and, you know, meeting the, overcoming the data famine in various ways. I think talent availability is also by will keep pace with that is going to improve. But yes, for now I think people have to the buckle up and start learning these skills. So that is what I think. So, Gaurav, what do you think, how are we going to deal with the current shortage of talent and how long do you think it takes for someone to pick up these skills?

Gaurav Gupta (07:33):

I think there are two things means the people who have the experience in software development, or coming from some other backgrounds, generally learning one time. And just applying that knowledge becomes a straightforward thing for people, but in machine learning is somewhat different that you have to, whenever a new problem comes, you have to think through from the scratch, you have to understand the domain, you have to understand the business and outcomes. So I think that kind of understanding few people are not very much capable of doing that sometimes, but yeah, I understand. So, if people are interested, I think it may take three to six months to acquire this skill at a good scale from where they can start working on some new projects or something, and maybe a one to two years of long-term roadmap where they can become the expert in the area.

Vinayak Joglekar (08:25):

Yeah. So the problem, you know, model that, at least in my experience with software engineers is that they think that every problem has a solution. And in machine learning, there are several problems which don’t have a single so, and some of them don’t have a solution at all. So, it’s a, a journey as go rightly pointed out and you have to apply several different approaches and you can never reach a 100%. You’re always, you know, if you measure in terms of accuracy or in terms of recall, you are somewhere near 80% 90, you can, and it becomes hard and harder as you approach 99%, right? I mean, and you can never reach 100%. That is one change that software engineers who are used to the precise way to define things and get the right outcome and the output using the right coding that I think that it’s a change of experience for them. So as Gaurav rightly pointed out, it’s not the same, it’s slightly different.

Gaurav Gupta (09:31):

And also, one more thing to add over here, that experience of multiple fields and domains like math, physics, chemistry, maybe images, 3D words, those interests. If you have multiple domains, then it becomes easier to do the things. For example, a software engineer may not know very well how audio works, but if you want to do machine learning for audio, like piece to text, you have to have the knowledge of how audio works, right? Some audio knowledge. So that kind of multidisciplinary knowledge is also required a good amount of times to become an expert in this kind of machine learning areas.

 

Vinayak Joglekar (10:10):

And I can’t agree more with you, you know, even somebody who is doing text processing needs to know English grammar, right? I mean, yeah

Gaurav Gupta (10:19):

You have to be a linguist kind of guy.

Vinayak Joglekar (10:21):

It’s not very straightforward. You need to have that domain in the context without which it’s not the same as, you know, picking up say a front-end skill such as Angular or React. And, you know and if you have a wire frame, you can just build it. So, Madhura you know, this data thing is takes us to the Natural Language Processing. I mean that is where I think plenty of availability of data is there. So structured data, which is you know, where a lot of the machine learning used to happen in the past, you know, that is hard to come by. One thing company who have lot of data, they are not completely digitally transformed and they’re sitting on the data, and they don’t have the wherewith thought to utilize it in the right fashion. And there are companies who are further advanced in machine learning, but they are not having the data.

Vinayak Joglekar (11:15):

I mean, there are a lot of the companies they’re like, all are dress up and nowhere to go because they have the algorithm, they have the NoHo, but they don’t have the data. Right? So to come out of this chicken and egg situation, I think Natural Language Processing seems to be the natural choice because there’s lot of data that is available on to be scraped from the web, like whether it is Twitter streams or social media streams and other social media, or even images and videos. And there’s plenty of available public domain where you can use it for Natural Language Processing in combination with machine learning. So, Gaurav, I think in 2021, do you think you’ll see a lot of this Natural Language Processing taking lead over pure play machine learning where you’re crunching, just numbers.

Gaurav Gupta (12:04):

True. I think National Language Processing, Robotics and IOTs, these kinds of areas are going to pick up like anything in 2021. And one more thing which I’ve observed, which is going to happen in 2021 people and companies are going to get educated. What data has to be captured, even if they want to do machine learning in future, because you need historical data sets. Many of the companies I worked with, we have found that they had the data, data was half big. It was not the complete data. And that data becomes useless as a historical data for the machine learning. So companies knowing even if they want to do the machine learning down the line a year or so, they need historical data. And what data has to be captured, that is very important. If you don’t have the proper data, then whole data capturing becomes you useless all that data.

Vinayak Joglekar (12:56):

Yeah. In old times you used to say garbage in, garbage out.

Gaurav Gupta (12:59):

Yeah. So one customer, they were capturing the data from the global scale, but they were not keeping from where requests are coming. They were having time stamp and everything. Now everything translated to GMT time stamp, but not knowing from where these requests are coming. Now, you can’t try any model, which can take the time as the major feature. Right. And their product required major feature is the time. Now that data is useless, right. Because we don’t know from where these requests were coming from.

Vinayak Joglekar (13:29):

Yeah. So, you know, who knew at that point of time that you would be using for machine learning. So another alluded to IOT and edge computing? So that is another way of overcoming data is you are going to have millions of these devices that are going to be spread all over and they will be sensor data that will be gathered from these devices. And already, I think we have more of these devices than cell phones, and it’s going to be like hundred times the number of sales phones that we have. So it’ll be close to a trillion of these devices very soon. We will, that those devices will be sending so much data. Now, one trend that I foresee that will happen is that instead of sending all the data to the cloud and then doing the decision making on the cloud, so lot of the decisions also need to be deployed at the edge because the business context exists at the edge and that’s where the decision has to be implemented.

Vinayak Joglekar (14:36):

So, it’s far easier when you reach the decision where it is to be implemented at the place where it is to be brought into action. So that is what is going to, you know, bring the new paradigm, which is, you know, on the cloud, you’ll do the learning and creation of model, which would be a periodic, if not one time exercise. And then once this model is ready, that model will be pushed to the edge. And on the edge, it could be very, very low computing, maybe your cell phone, maybe a raspberry pie, or maybe an Arduino board in which the model will reside, and the inference will be drawn at the edge. And outcome of the inference would be displayed to the human expert or the execution, the person who’s actually going to take action on that decision, that he would be able to reach based on the inference. So this type of working is going to bring in what we call more of active learning wherein you know, the human at the end may say that, okay, his particular inference is right or wrong, or it needs to be improved. And that same would in turn, go back into training the model further. So with every use, the model will become smarter and smarter. So, I, let Gaurav talk a little bit more about this phenomenon that we are going to see of active learning and what he sees as the future in going forward in 2021?

Gaurav Gupta (15:59):

I feel that active learning is going to be a major player in the market because most of the behaviors which we are seeing on mobiles and mobile applications, many things can be done by artificial intelligence from the cloud and models can be deployed on the device. For example, piece two text kind of thing, which it requires the web. Now Google is pushing the models on the device itself. So many Android devices, new version are having the piece two text function, which is on the device, not on the cloud. So I think that is go going to change the way companies are using machine learning for many scenarios.

Vinayak Joglekar (16:37):

So, you mean like they can be dictating notes to my Android phone, even though it is not connected to the internet.

Gaurav Gupta (16:44):

Yes. You can do today. Also means today they have the application for a speech to text and Android on Google application that doesn’t use the web that does it to offline up to my knowledge. They’ve done it a few months back.

Vinayak Joglekar (16:57):

Yeah. And you know, I had recently worked on one project where this is specific to the pandemic era where, you know, you want the camera to be smart enough to know whether someone is wearing a mask or not. So that needs to happen in a very quick way. That is no way to process that the inference by sending the image on the cloud. So, person who’s waiting at the gate can wait, can’t wait for that long just to check temperature and whether he is wearing a mask or not such image data can be processed right there at the where the camera is very, very, very low computing power. So, there are many, many applications like that we are going to see in 2021. And, you know, active learning is one thing, but you know, this data famine, if you don’t have enough data then have heard about this and this, I think Gaurav is your area of expertise that you can actually borrow an existing proven model or, you know, and use that and improve upon that. So can you talk a little bit, I think what I’m talking about is transfer learning. If I’m not wrong Gaurav, can you tell me?

Gaurav Gupta (18:05):

Yeah. What you are talking about is exactly transfer learning. Is so idea is very similar that if some kid knows one language and you want to teach them another language, they can pick up very fast because they already know the concepts of what the language is. So, it’s coming from the same thought process that how humans learn, transfer learning that if they have a skill similar kind of skill learning is not a big deal. It can be done very fast, not putting that much effort. So in transfer learning, for example means you take the images and you train a model which can reduce the dimension of the image to few variables and recreate that similar image from those variables. So, we know we name as latent variables. So, this is like a encoder decoder architecture, where you encode the whole image to few variables from these variables, you decode the image properly, you train this kind of model a lot.

Gaurav Gupta (19:04):

Now the problem, and based on this latent variable, you remove the decoder and then you do the prediction, whether what is the category of the image, let’s say, now the thing is your domain images may be very less. For example, I want to classify between two basic images, let’s say, suppose a cycle and a bike. I want to segregate two things. I may have maybe 50, 100, 200 images, let’s say, but it’s very difficult to train a model only with 100-200 images you need millions of images. For example, now here we can use the transfer learning approach. What we can do we can take free hand images from the web. A lot of images, let’s say millions of images, train a model with end quarter decor, which doesn’t learn anything except to encode the image and decode the image. Take that once that model is stream, take that model, take your hundred, 200 images of cycles and bikes and tune the mirror model with the very low learning rate so that it can add up to new data because already it has understood the concepts, how image has to be encoded and decode it. Once this model learns your bikes and cycles with a very few learning rate. And maybe in a very small time after that this latent variables of the end quarter can be direct used for the classification. So, you can apply from one domain to another domain. Similar thing now, cutting edges, going into the NLP side also where people are doing similar thing on NLP, they trade the Wikipedia data, train it, train a language model, which is not understand the whole English. Yeah.

Vinayak Joglekar (20:43):

I think you have done it, right. I mean

Gaurav Gupta (20:45):

Yeah. We have done both things for different customers. We have used the transfer learning for images also for a customer for segregating their images between multiple class, finding out the similar images, basically that if I give image, can you find the similar image from the dataset? We have used the similar technology transfer learning for that customer. And for the NLP also, we have used a similar technology where customer had data from the API data. And we are using that API data in the language model, which was trained on the Wikipedia to retrain with our API data and finding out, doing the classification of the API data sequences, whether they are in this direction, customers are going, whether they’re going to application more or less or what. So yeah, transfer learning is a very powerful concept. And I think in other areas like video processing, robotics, robotics and car driving, I think people are using it from long time in these areas.

Vinayak Joglekar (21:47):

Go, this start driving, for example, this is great. Totally. I mean, you know, I’m just concluding the discussion that we had on transfer learning. Like the necessity is the mother of invention. And when we have, we are faced with data famine, we overcome that even with less data we can make do with, by using this and variables encode and decode model that’s, this is something which is totally very innovative. But you mentioned this in passing just in the last minute about driving, right? I mean, that is a very, very different process, right? I mean, you start from somewhere, you reach the goal, but the actions you take, and let’s say, if you have fitted a camera on a self-driving car and you try to see what all steps happen, it differs from time to time. And, you know, every time a driver doesn’t reach the destination in the same way, I mean, there could be hundred different events and actions that happen along the way. So it’s not a straightforward. And then how do you train such models? I mean, this is something which is totally mind boggling.

Gaurav Gupta (22:55):

Yeah. What we’re saying is too, that this is mind boggling and even understanding how it really works is quite complicated. So in a nutshell, if I want to say the techniques which we use here is not the transfer learning, mainly view the enforcement learning, where we have a simulated environment, for example, a real life car also, but that will make so many accidents, how many cars we can use. And if speed is going to be slow, robot is learning how to change the gear, how to go up and down, or how to accelerate, how to deaccelerate. It may take maybe months on a real car for a model to train, even to deaccelerate and accelerate, maybe learning. So usually what people are doing in this situation, we, we use reinforcement learning in a simulated environment. Like you can take the game of like NFS game engine and you have different perspectives. So use every perspective camera as the..

Vinayak Joglekar (23:51):

Need for speed game, right? I mean, it’s a video game.

Gaurav Gupta (23:55):

Yes. And need for speed. It’s a game. It supports all the features which cars have. And it already, it also has different camera angles, which you can use. It has the night mode. It has all the weathers. Your car will is scared, mud roads. Everything can happen, which happens in real world. So you can create the simulated environment, train the model on a reinforcement learning techniques, where you are working as a real reward and punishment mechanism that your target is to reach this destination in between. If your car gets damaged, then it’s a punishment for you. If you go properly, it’s a reward. If you reach, if you take too much time again, it’s the punishment. So we set up a punishment and reward mechanism over there and it’s learning continuously so that it can achieve the target properly. We do generally simulated environments like training the robots, how to pick up the objects, how to paint a car, how to assemble a car, everything is done in a simulated environment. And once models are trained, then again, it’s not exactly transfer learning, but quite a kind of transfer learning that we take that model put on the real-world robot and put on the real-world car. And then we ask this model to be detuned on the real-world data, which is coming from the cars and the robots. So here we use the reinforcement learning for training the cars and all that.

Vinayak Joglekar (25:25):

So, I think this is something that you know, it’s very easy for us to say that self-driven autonomous cars are going to come, but you know, we don’t know what goes behind. So, there’s something which what we are going to see maybe 2021, or maybe towards the end, or maybe next year, we are very likely to see extensive use of reinforcement learning. So, right. This is exciting, Gaurav. And I can’t wait for these things to pan out very, very quickly so that we can start using these things for our benefit. So, Gaurav thanks a lot for your time. And thanks for joining us. So, Madhura, is there something that you wanted to cover today that we didn’t?

Madhura Gaikwad (26:10):

No. I think Vinayak, we’ve covered almost everything that we had discussed, and this was a great session, and I would also like to thank Gaurav for joining us. And thank you, Gaurav, and thanks, Vinayak.

Vinayak Joglekar (26:23):

You most welcome Madhura it’s always a pleasure for us.

Madhura Gaikwad (26:26):

Thanks. And that was a brilliant session and we will continue to discuss these tech trends and predictions for 2021 in our upcoming episodes. Thank you everyone for tuning in. If you are looking to accelerate your product roadmap, visit our website, www.synerzip.com for more information, stay tuned to future zip radio episodes for more insights on technology and agile trends.

For more insights on technology trends and predictions, download  – 9 Technology & IT Trends and Predictions 2021

ZipRadio is available on these platforms