Our AI strategy is pivoted on NLP, automation and trust

IBM veteran, Gargi B. Dasgupta, took charge of IBM Research in India more than two years ago. As Director for IBM Research in India and CTO for IBM India/South Asia, she serves as a strategist for IBM Research, working closely with IBM’s Hybrid Cloud business and global services businesses. Prior to being director, she was a senior manager at IBM Research-India, leading the work in AI for Automation. Among Gargi’s accomplishments, she was recognized as an IBM Distinguished Engineer for her exemplary technological achievements in 2019.

In an interview with CXOToday, we asked her about her leadership style, her view on cutting edge technologies like AI, NLP and automation, and her thoughts on how to have more women in tech leadership positions. Edited excerpts:

You describe your focus areas as “realizing the vision of Future of Computing through the infusion of AI and Blockchain into the enterprise ecosystem” and helping “IBM India lead with innovations in Cloud and Cognitive”. What’s the progress you’ve made in the last two years in the context of these focus areas?

IBM is focused on two key things–helping our clients transform their businesses, which is via hybrid cloud. This involves enabling them to access the power of cloud. Then this cloud becomes a vehicle for everything that the company wants to do, which includes AI. Our prediction is that every company becomes an AI company, or has AI fed into their deep business processes. To do that at scale, you need a stable vehicle. And that’s our hybrid cloud platform. That’s exactly the mission of IBM Research India labs.

Through the last couple of years, we’ve been spearheading innovations and technology in both these domains–one in helping land the hybrid cloud journey with clients but with also innovation, which means you also intersperse AI as clients move to hybrid cloud.
What does that mean? It means I’m locked up in data centres, I have massive application. Some of my developer teams must have left. They might be accessing data from all remote locations. My data is, thus, fragmented.

Hence, I need a good solution to move to more modern technologies, but I cannot afford business downtime. I cannot take down my entire retail store or my banking system to do this move. So how do you do it? In bits and pieces or in an optimised way? How do you refactor code? How do you re-platform applications? Here’s where AI plays a very big role.

Please give us some examples…

The whole consumer world is building language models–very large language models. For instance, language models like BERT and GPT are taking the world by storm. And it should, because there is an enough data with consumer companies, and we should start building large language models. IBM also released a huge data set called Project CodeNet, which is a large dataset aimed at teaching AI to code. It consists of some 14 million code samples and about 500 million lines of code in more than 55 different programming languages.

Just like I can use BERT and GPT to complete my sentence, write or generate a sentence, I could use CodeNet to really create these language models for code, and then use that for my application modernization.

How are companies responding to Project CodeNet?

The hope is that with this project, we just launched it, that a lot of researchers get excited about these problems. I mean, we use it internally, in a lot of our client engagements, right. But the hope is that we foster more we know we have students look at it, faculty look at it, other users look at it. So, we’ll see. I think I can answer that question maybe a couple of months from now.

What’s your AI strategy?

Our AI strategy overall is really pivoted on three things. Its natural language processing (NLP), automation, and trust. This is being done over the years ever since, we played Jeopardy and launched Project Debater (an AI system that can debate humans on complex topics). And it has culminated into some significant accomplishments and got us recognised by Gartner for our AI leadership. I would like to break it down into two personas.

One is having AI tools and technologies that help developers write: How do I build a model faster? How do I test a model faster? The next is having machine learning platforms–in order to build a model, I need data pre-processing, I need to deploy, I need to continually monitor that. So, we moved up to the leadership quadrant. IBM’s Weather Channel platform, Watson Assistant and all the AI we do is also part of our AI application story.

When you look at a country like India, there’s a lot of talk on AI. where do you think India figures in the AI leadership ladder?

There’s a lot of traction in certain industries. But there’s also been a good study that’s been done on AI uptake in the last couple of years for India. They reveal that almost 50% of Indian companies absolutely think that AI and automation are going to be big game changers for them.

But those are surveys and studies. My sense is that most of what companies are doing is basic machine learning statistics rather than AI. I would love your perspective…

You’re asking the boundary question of where this statistical machine learning gets into AI. There is a roadmap to this entire landing with AI as an information architecture. This mandates that you get some basic things right, which is that you have your data, your data sources for your AI algorithms to work, you have trust embedded, or you have a mechanism to infuse trust in the data sources and the models you’re building. I think those are much important questions than really being pedantic about whether I am using the latest and greatest auto encoders and deep learning or simply statistical machine learning.

Basically, you’re saying that AI needs to solve a business problem…

Yeah, exactly. It’s most important. To solve the business problem and to solve it in a way, like I said, with data. Sometimes, as data scientists might assume, there’s a data lake and we just make a few calls. But the data could be fragmented or sparse and have bias in it. And the much of data generated might not be even movable. There are a lot of hard problems 80% of the time. I would say, let’s solve these problems first. Let’s help humans do things via AI and automation better, rather than getting too hung up on whether I am using the latest and the greatest deep learning tool. In research, we always have our eyes on what’s next on the horizon, but businesses might still be okay with simple regression-based algorithms, pattern matching — something that work for them.

Agree. Businesses also must look at the return on investment (ROI). That said, typically, large companies and medium-sized companies and large companies would have implemented AI in some form or the other. What about small companies? Can they afford AI tools? What would advise them?

I would say that a newer technology has been out in the open and being pushed by the open community. For example, if you want to start with simple classification, class clustering or recommendation tools, there’s a lot of support out there in open source–a lot of support from IBM releasing open source, and so does Microsoft and Google.

Hence, if you want to just want to get your feet wet, there’s enough out there in open source, and open source doesn’t mean that you would get something that’s not supported or not trusted. We’ve released our AI fairness toolkits, AI trust toolkits and Explainability toolkit in the open. So, you can get the end-to-end pipeline very easily ready by just using open source tools–both for the machine learning for the data pre-processing for trust, and explainability.

And then you see what that solution looked like at that point, you know, when you then decide that, how do I integrate it into my business process? There will be many questions on these open source tools. And that’s when you start thinking that what do you need to protect and license and grow and monitor? And where do you want to spend the bucks (money).

But small companies sometimes argue that they do not generate enough data to justify the use of AI. How would you respond?

Every company, whether they know it or not, generates data. Consider this case. My friend runs a gym. She trains young kids and adults. We typically associate gyms with a physical space and machines. But in her case, ever since they moved online, they’ve been capturing demographics–what kind of customers age, weight, target, etc. She uses tools that can take pictures, and she’s been applying AI to all that data that has been generated by roughly about 10 customers in a week. The numbers may seem small but each of those customers have multiple parameters and create samples and sub-samples of data that can be used to train machine learning algorithms for image recognition to understand if her clients are training right or not. For instance, if the posture isn’t right, let’s correct it.

Effectively, you can’t run a small, medium, or big business without data. The Quantum of data is where the confusion lies–do I need a terabyte of data to train my machine learning? Absolutely not. We can train logistic regression, SVM models, and decision trees with even a few hundred samples. The hope is that as you collect more data, you can get the machine learning algorithms to learn the features automatically. That’s what we want. We don’t want to do feature engineering. With small amounts of data, we remove the bottleneck of feature engineering as we get more data.

But the quality of data also will matter…

Excellent, I was coming to that. It’s in the context of automation that we speak of data and data quality. There are two ways to understand automation. The first is using AI for automation. Consider a loan-lending process. I have a recommendation process, I have a hiring process, I have a procuring process. How do I infuse AI in all these stages?

I can use historical data to understand patterns and apply them so that they really work. IBM infuses AI in automation with what we call ‘Watson Orchestrate’. It’s really an AI-based orchestrator that understands the skills I need to do my job. For instance, I may need integration with a ticketing system to do my job. Or I may need integration with a calendar system. There’s a bouquet of skills, and I express these in natural language. Effectively, Watson Orchestrate is an AI-based planner, and it’s leading with how you start infusing AI into all kinds of business processes.

The other part in automation is automating AI itself. What does that mean? It means that AI models are not built in silos or are not going to be successful in silos. Instead, you have this whole data cycle, data science cycle and the entire lifecycle. The data cycle really means what data do I have? Where is the data? Where are the data lakes? Can I get them together? Can I move them in most places for applications? How do I get the data ready so that it is consumable for the AI? It also includes tasks like removing biases and understanding if the data is appropriate for training.
Then there’s the whole data science automation–do I really need people to sit and do data engineering and feature engineering? Or can I have optimization that really does all this automatically for me–automate the process of data science and then deploy. For instance, let’s assume I built (a model) on three months of data but when I deployed it, COVID happened. Hence, all bets are off since the entire workflow and workloads have changed. I now have a completely different demographics of people requests. Who monitors and ensures that the data and the model we put out are in sync? Who decides whether it needs re-training or just a few tweaks in the model? Hence, whole lifecycle automation is especially important, all of which can again be done with AI tools. That’s really our automation.

Speaking about AI and automation, how do you perceive their impact on the workforce in terms of re-skilling or redeployment where necessary?

That’s a hard battle. It’s often not about technology. It’s often about people getting to know how to use these tools and then re-skilling and reorienting their thinking, and the processes. We faced this in the early years when we tried, for instance, to infuse AI in the account payables process. The problem was not about the accuracy of the AI tool. It was about the shift in mindset and the business process that had to be reinvented. We have handheld many clients to help understand how their business processes need to be new if they’re going to infuse AI.

I’ve seen this in operations–when we try to introduce an AI-based tool to help manage operations, companies had to change a lot of their management processes with New Age thinking before accepting algorithms. You must get them (managements) involved early on in the design process and tell them that this is going to be useful and have them as part of your journey as you design the UI (user interface).

In this context, what is your stance on AI and ethics in terms of integrating these into the design process itself?

Absolutely. Ethics must be the underpinning of everything that we do build, launch, and ask others to use. That’s important for us. That’s really our third pillar–ethics must be woven, like you said, into the entire lifecycle. It cannot be an afterthought after you build the model.

On a different note, you were the first woman to head IBM Research in India. How has the journey been so far, and how would you define your leadership style, especially in these trying times when the Covid pandemic has engulfed the nation?

Of course, it was a big honour to lead IBM Research India. I absolutely cherish the role because I had the best minds to work with and (my role involved) channelizing them towards something that means value for IBM, for IBM clients and for our community. It’s been a wonderful experience. I think we got stuck with something (the Covid pandemic) that we had not anticipated, and there was no playbook to know how to lead a team and how to keep your team safe. We had to invent and innovate. It has been challenging. But I’m very hopeful that we will come out of it.
And technology has always as has played a major role to help accelerate (digital) transformation that otherwise would have taken 6-7 years. I think technology has been kind of keeping our lights on–we all switch to online working to deliver value to our clients with the bat of an eyelid. Even in the second wave, I think we’re continuing to see how technology can help supply demand how technology can help citizens and IBM.

As for my leadership style, it is really to be more hands-on to understand a problem area. It is also to always ask the question: How does it benefit my end customer? Have you created an algorithm to act faster? and What are the use cases? And then of course, I think empathy is at the core of it. That’s how I’ll describe myself.

I agree but the fact is that a CTO’s job has traditionally been a male bastion. How did you manage to break the chain?

I think that the problem is not that many women were not getting the right opportunity. Even though India has decent numbers for women in STEM, they need to do a PhD. It’s not mandatory, but you know, we need deep thinkers. We must fix it–that’s one of the challenges (and) we have (people like) Anita Borg doing so much good work in this area. I know, the gender gap report is depressing. But I think that we’re at least making a dent.


Leave a Response