Our bet on deep tech and frugal innovation will deliver a huge societal impact: Sriram Rajamani, Corporate Vice President, Microsoft Research India – The Indian Express

Sriram Rajamani is passionate about the societal impact of Microsoft Research Labs and the way it has helped change lives on the ground.
For Rajamani, Corporate Vice President and Managing Director, Microsoft Research (MSR) India, apart from the societal impact-centric projects that MSR India takes up, it is the core computing problems that the lab is working on that could be a gamechanger in the long run.
In an interview to indianexpress.com, Rajamani talks about two such deep tech projects — Extreme Classification and Causal Reasoning for LLMs (large language models). He also shares his vision for the organisation, his projects with a focus on societal impact, the deep tech work at the labs which could be a game changer, and the impact on the ground with his work. Excerpts:
Sriram Rajamani: Started in 2005, it is part of a team of Microsoft Research Labs across the US, UK, Europe and China. We have a team of 50 core research scientists, and a team of 150 including engineers, research fellows, post-doctoral fellows, and interns.
Our focus in Bengaluru lab is on Algorithms, Systems, ML and AI, and Technology and Empowerment. We work on many problems, some which are immediately relevant to the Microsoft ecosystem and some of which are what we call ‘open ended’ problems.
We work on deep scientific work which might take years of work to materialise into research results, and our research finds its way to peer reviewed publications. Our researchers have also won many top tier awards such as the Shanthi Swarup Bhatnagar Prize, Infosys Prize, Knuth Prize, McArthur fellowship etc., both from India and abroad.
Sriram Rajamani: Let me take two examples — Extreme Classification and Causal Machine Learning.
Normally in machine learning, there is classification of data into a small number of classes. But now researchers at Microsoft are looking at extreme classification where they classify data into categories in the order of hundreds of millions, and even billions.
This is being picked up by the adtech and search industries, but could have long term societal impact when it could get used as recommendation engines by doctors, giving personalised advice to patients based on large amounts of data collected earlier on the patient and on the disease.
Sriram box
Most Machine Learning models study correlations, but not cause and effect relationships in a proper manner. We are looking at causal capabilities of LLMs and their implications. Our goal is to also make causal methods easier to adopt in the future.
This would have a huge impact, for instance, on interventions using health or mental health data. For example, with large sets of cardiac health data, we might be able to zero in on a few causal factors, which have an outsized impact on recommending the right course of action to improve cardiac health. There could be many such applications like these.
Sriram Rajamani: We now work on around 15 projects and about a fourth of these projects are exclusively focused on societal impact. However, all our projects would have a large societal impact going forward, only that it might not be very apparent in the short run.
One of the interesting projects that we run is Project VeLLM, which tries to bridge the digital divide particularly in a landscape that is being transformed by Large Language Models (LLMs).
In Project VeLLM, we are building a co-pilot to help teachers from government schools to query the platform for building additional teaching material or lesson plans.
We are now training VeLLM on various languages like Kannada and Hindi so that it can understand vernacular language requirements better and answer the user in his or her language. It is like a ChatGPT for Indian teachers.
We hope that the underlying technology can, going forward, help nurses, doctors and farmers querying the platform in their local language.
Sriram Rajamani: For societal impact projects and for those projects for which there is no potential fit within the Microsoft ecosystem, we first open source the IP and may also provide some seed grant so that they can become independent companies, and sustain themselves with funding from the ecosystem.
Some of these projects were incubated here and later found a new life and purpose of their own, outside our organisation. Some of our researchers do continue to collaborate with these entities and work with them to improve their performance.
99Dots is one such project which was incubated in Microsoft Research India. It began as a project to use low-cost technology intervention to help with Tuberculosis Medication Adherence. World over, many TB patients do not follow the medication regimen, and hence this was affecting their cure. The solution proposed by our researchers was effective, and hence we spun it out as an independent company, called EverWell Health. It is thriving now, and addressing many other problems in the healthcare sector.
Sriram Rajamani: Project ELLORA: Enabling Low Resource Languages, is where we design and build speech and NLP systems for low resource languages. Some of the languages we work on are Gondi, Mundari and Idu Mishmi.
We do that with innovative methodologies for data design and collection, and in some cases since there might not be a large number of speakers, we also crowdsource the data collection, and perhaps gamify the collection process to increase interest levels.
Apart from the low resource languages, we also work on data rich languages where we look at trying to reduce the inherent bias that could creep in language models. For example, we look at and try to rectify gender bias that could enter Machine Learning models by intentionally collecting data that does not have gender bias.
Sriram Rajamani: Karya’s vision is to provide dignified digital labour for marginalised communities in rural areas. Project Karya empowers local communities to do digital work and be paid for it.
With the advent of AI, and the desire to make AI work in local languages, there is a huge demand for data collection work in local languages. We are working with rural and poor communities to provide them with the opportunity to do this work and benefit financially from doing the work.
This exposure to digital literacy and platforms would also unlock other opportunities for these users, apart from being a supplemental income provider for rural communities.
It has more than 30,000 workers on the platform and has a presence in more than a hundred districts across the country.
Sriram Rajamani: It is a deep tech project where the core goal is road safety, and we do this by using technology to give feedback to drivers about how they are driving, and how they can drive safer.
In Project HAMS (Harnessing AutoMobiles for Safety) we use commodity smartphones to monitor the state of the driver and how the vehicle is being driven. The system can track if the driver is sleepy, or if they are not following basic safety norms or if the driver has a sharp braking event etc.
This technology is being adopted in the country in different ways. For instance, seven RTOs are now using HAMS for driver license testing, and there is minimal manpower involved. The system uses just a standard mobile phone to test the applicant’s ability on the test track. There is now greater transparency in the issuance of driving licenses. More than one lakh driving licenses have been issued using this technology, with a pass percentage of 60 per cent.
We wondered if there would be resistance from users, but driver’s license applicants appreciate the transparency in the exercise, and the reception has been positive.
Comparable technology is available in the US, but the devices are expensive. In HAMS, we achieve this using a commodity smartphone. We believe in frugal innovation in such efforts, and it has helped with broader adoption.
Sriram Rajamani: We collaborated with OpenNyai, AI4Bharat and the Bhashini project of the Indian Government to help them build an AI-powered chatbot Jugalbandi. The chatbot can understand questions in multiple languages, retrieve information about government programmes, then answer the user in their own language. This would be helpful for citizens to query government programmes and projects.
For example, a farmer could either type in the question about an agricultural benefits programme, or even just ask the question in their own language, and the AI bot converses with them to help them get an accurate answer. Jugalbandi’s vision is to make such chatbots available in over 50 Indian languages.
Venkatesh KannaiahVenkatesh Kannaiah is indianexpress.com's consulting editor for South … read more


Leave a Comment