As part of the new year’s initiatives on ML India, we promised you that we will be talking to leading researchers and academics in the field of machine learning to understand their work better. We start off this series by putting out the first interview we had! We have a couple of more very interesting interviews of academics from IIT-Delhi, IISc, IIT Kharagpur and more. Stay tuned!
We interviewed today Dr. Manish Gupta. He is Vice President at Xerox Corporation and Director of Xerox Research Centre in India. Previously, Manish has served as Managing Director, Technology Division at Goldman Sachs India, and has held various leadership positions with IBM, including that of Director, IBM Research - India. Manish holds a Ph.D. in Computer Science from the University of Illinois at Urbana Champaign and is a distinguished IIT-Delhi alumnus.
Manish talks about four areas which Xerox is innovating in - Education, Healthcare, Transportation and the services vertical. He describes how challenges in each of these domains can be posed as hard computer science and specifically, machine learning problems in areas as diverse as vision, NLP and speech. He talks of the problem in of automatically preparing the table of contents of a video, parsing equations described in lecture videos, tweaking deep neural networks to create intelligent chat bots and more.
Manish on ..
- The flavor of problems being tackled at XRCI
- Innovation in active healthcare
- Education technology and smart chat bots
- Startups space in India
- Outreach programs by XRCI
ML-India: Hi Manish. Thanks a lot for agreeing to talk to us on such a short notice. We understand you must be having a busy schedule.
Manish: Oh, no worries. In fact, Xerox is organizing a winter school for machine learning and today was its first day. I may have to head back to it in sometime. Hope that’s fine.
ML-India: That’s great to hear. We’ll be keen to know more about it and will chat about it in a while. Sure, we’ll keep a tab on the time. We wanted to start off with understanding what problems are of interest to Xerox research.
Manish: Sure. Today, a vast section of our population is underserved in the most fundamental services like health care, transportation and education. There is a room for improvement in the general area of personalization. A case in point is education. We are beginning to see a shift from traditional human intensive education systems to massive online courses where lectures given by eminent professors are utilized at a massive scale. The aspect of personalization is however still missing there. There’re no systems currently which learn what content suits you best, what style and pace of lecture delivery helps you learn best etc. There has to be a two way interaction between the users and providers of technology driven services. The unique needs and preferences of the end users need to be an inherent part of the technology. This is the key focus area of Xerox Research Centre India.
Xerox Research is fairly new – we set up shop in 2010 in India. We have a strong focus on targeting the top tier conferences to showcase the quality of our work; some of our recent work includes a paper at ACM multimedia, the top conference for video (multimedia) where we addressed the problem of automatically creating table of content for online lecture videos. We had two papers in Intelligent User Interface (IUI), one of the best conferences for showcasing working in human computer interfaces; and a couple in ICASSP or Interspeech, the best conferences for speech processing related work, where we published our work done in education-related products.
So to specifically list out XRCI’s key areas of focus, they’d be -
- Health care
- Transportation
- Technologies for education
- Services across different industry verticals. Example of this is customer care contact centre.
I could probably talk about each of them in some detail, which will help give your readers a good grasp of the kind of specific problems we’re pursuing.
ML-India: Sounds good!
Manish: I’ll start with healthcare.We are targeting two goals there. First, personalization and second, making it proactive. Today’s health care system is very reactive in nature. A multitier system exists today where the ICU is the most expensive tier. The aim of our healthcare systems should be to minimize the escalations into the ICU. There is a lot of activity these days in healthcare analytics but we all have a long way to go. Technologies which predict complications in ICUs, some of which even Xerox has built, do not work very well because they are still based on a bunch of clinical rules formulated by doctors. These are good preliminary interventions to have, but these systems are wrought with false alarms. There is a need to apply modern techniques of machine learning and data science to improve the accuracy of these systems.
In this regard, one of the areas we’ve worked on is in predicting which patients get admitted to ICUs. The state of the art method of predicting this is based on a scoring system called Modified Early Warning Scores (MEWS), which is based on the measurement of temperature, blood pressure, heart beat and many such body vitals. The alert generated by these methods are generally ignored by doctors and nurses given the high false alarm rates the present systems have. Our team has worked on one specific complication called the acute hypertension episode. Our recent results show we’ve managed to do better than the best known results by a significant margin. Encouraged by these results, our business groups have given us the challenge that of having 99% specificity so that there are at least 3-4 true positives for every false positive reported by the system. We’re pleased to say that we’re taking this challenge head on and pushing the state of the art.
Another area we’re tackling is in breast cancer screening. We are trying to resurrect an old idea which the health community gave up on sometimes ago. Today, breast cancers are screened for using mammography, which involves pumping X-rays into the body. It’s a very painful procedure and additionally, it’s not effective at all in younger woman. The medical community has known for a while that one can detect a tumour using thermal cameras. When a malignant tumour grows fast, a distinct thermal signature around that tumour can be captured through a thermal camera. In the early 80s, there were too many false positives in analysing these images. By applying modern machine learning and image processing techniques, we have been able to dramatically decrease the false positives for this detection. We are also working with a cancer hospital in Bangalore to collect more data and run our algorithms for a larger datasets and a larger class of cancer detection problems. This work has been submitted to a top-tier peer reviewed conference.
ML-India: That’s great to hear.
Manish: In the education space, we have built just the product to provide what I spoke of in the introduction. It’s called TutorSpace which is a comprehensive suite for a learner to learn rich and relevant content from the web. It automatically stitches the most relevant content for you by skimming through scores of lecture videos on the internet and also summarizes the key concept and the equations/content that’s delivered.
Customer care is the third area where we are doing a lot of work on automation. This is the first year where our group has two papers accepted at AAAI, one of which is our work in customer care. We have a product in place currently where virtual agents, as against humans, try to automatically answer customer queries. These agents could be designed to answer queries over web chats or over phone calls. We started with the easier problem of designing agents for web chats. Our team at XRCI has built a watchdog which constantly monitors the interaction between the customers and these virtual agents and identifies when a human representative ought to intervene and take over the process. Such a system could also be applied to how one might automatically analyze everyday web chats. In essence, what we have demonstrated is that it is now possible to analyse these chats in real time, unlike the current mechanism in which a post-hoc analysis of such logs take place. With the successful demonstration over web chats, we are now moving towards replicating similar systems for voice chats. Our speech research group at XRCI is coming up with its own enhancements of how deep learning is applied to speech, surpassing Andrew Ng’s recently published results. We will be submitting our work to some of the best conferences in speech research, where we shall get a chance to compare our work with the state of the art.
In transportation, the focus is more on algorithms and optimization. We work on solving integer linear programming problems, coming up with greedy algorithms to get close to optimal solutions. All this is done in the backdrop of scheduling and routing problems.
ML India: That’s great. So from what we’ve heard, is it the business team pushing XRCI with challenging problems? How does it start?
Manish: It’s a combination. We try to have a balanced portfolio. Roughly one third of our work is exploratory in nature. For that one third unit you are not driven by a business unit telling you what to do. That is more bottom up where our researchers identify a relevant problem. They don’t need a close business case but they at least need a hypothesis. Our work on education and breast cancer screening are examples of exploratory research. In both these cases, we are working with startups to commercialize our technology. The problem of predicting complications in ICUs was given to us by a business group. Half of the two third work being informed by our business groups is relatively short term and other half is relatively medium term.
ML India: You mention working with lots of startups. What is your take on startups innovating in the machine learning space?
Manish: In our case specifically, we try to connect with those startups which do not have much depth in machine learning. They bring to the table agility, market access and products. We combine it with our research depth to bring something unique and technically sound into the market.
Having said that, I think there’s a very positive vibe in the startup community regarding machine learning. I think today’s entrepreneur realizes machine learning helps in giving an edge and making services more competitive. I would say lot of the current capability of many of these startups are still not great, but I am very optimistic that it will get there. I would really think of your company as again a shining example of guys who have been actually publishing papers at KDD and the likes. In due course of time I would imagine that others would follow suit.
ML India: What steps are Xerox taking? I see machine learning school being one such initiative. Is there any other initiative that you are trying to push forward so that you engage with the community at large?
Manish: We are organizing a winter school in machine learning where we have ten teams spending a fully paid two week internship at XRCI, where they will be working on problem of complication prediction in ICUs. We also give out faculty research grants which is a global process. Our lab is championing 5 faculty research grants of which at least three are in machine learning. We are funding Indrajeet Dhillon at UT Austin, Jennifer at MIT and Chiranjib Bhattacharyya at IISc. It’s a grant of thirty thousand USD per annum for three years. What we are finding is that current state of art is not anywhere close to being good enough for the practical problems that we are dealing with and we’re collaborating with the best in the industry to push the frontiers of both, theory and practice of machine learning.
ML-India: Do you have any take on talent pool that India is producing through university?
Manish: I am very optimistic. I was positively surprised by the kind of response that we got for the research challenge. Two out of the top ten entries were very neat peace of work which were totally new ideas on how to solve that problem. We have a special program called budding scientist program where we give undergraduates a two year appointment and very challenging problems to work on and then encourage them to pursue a PhD after that. The rationale is that many of these undergraduates have no idea what is out there in the industry and are easily attracted by high paying corporates, many of which are not exciting technically. Such a program helps retain those academically inclined within academia. One of the students we recently offered already had one journal paper and submitting a paper at NIPS. He had applied deep neural networks to automatically differentiate between Beethoven and Bach with over 95% accuracy. I find it very encouraging that you have such amazing talent coming out of our schools.