The second point I wanted to add was to make the computations power-aware – in the future world of sustainability, green computing/green AI is a must. Signal processing engineers are inherently trained to make their algorithms work on low power, low latency constrained embedded devices. The same principles need to be applied to designing power-efficient AI systems. It is the need of the day in this age of over-parameterized ultra-large and power-hungry AI models. In this context neuromorphic computing architectures that mimic human brain neuronal structures will need to be explored in place of Von-Neumann architectures of CPUs, Harvard Architectures of DSPs or massively parallel architectures of GPUs.
ndustry Leaders in Signal Processing and Machine Learning: Dr. Arpan Pal
Industry Leaders in Signal Processing and Machine Learning
By:
Dr. Abhishek Appaji
Industry Leader in Signal Processing and Machine Learning
Dr. Arpan Pal
Distinguished Chief Scientist and Research Area Head,
Embedded Devices and Intelligent Systems, TCS Research, Tata Consultancy Services, India
Dr. Arpan Pal - Bio: I have more than 30 years of experience in the area of Intelligent Sensing, Signal Processing &AI, Edge Computing and Affective Computing. Currently, as Distinguished Chief Scientist and Research Area Head, Embedded Devices and Intelligent Systems, TCS Research, I am working in the areas of Connected Health, Smart Manufacturing, Smart Retail and Remote Sensing.
I have been on the editorial board of notable journals like ACM Transactions on Embedded Systems, and Springer Nature Journal on Computer Science. Additionally, I am on the TPC of notable conferences like IEEE Sensors, ICASSP, and EUSIPCO. I have filed 180+ patents (out of which 95+ were granted in different geographies) and have published 160+ papers and book chapters in reputed conferences and journals. I have also written three complete books on IoT, Digital Twins in Manufacturing, and Application AI in Cardiac screening. I am also on the governing/review/advisory board of some Indian Government organizations like CSIR, and MeitY, as well as of educational Institutions like IIT, IIIT, and Technology Innovation Hubs. I am two times winner of the Tata Group top Innovation award in Tata InnoVista under Piloted technology category.
Prior to joining Tata Consultancy Services (TCS), I had worked for DRDO, India as Scientist for Missile Seeker Systems and in Rebeca Technologies as their Head of Real-time Systems. I have a B.Tech and M. Tech degree from IIT, Kharagpur, India and PhD. from Aalborg University, Denmark.
We approached Dr. Pal to learn more:
1. Why did you choose to become a faculty in the field of signal processing?
My BTech project at IIT Kharagpur was in Digital Signal Processing (DSP). It was in the late 1980s when I was introduced to the fascinating world of adaptive signal processing, which I would say is a form of AI for linear embedded systems. During my MTech, Radar Signal Processing piqued my interest and this interest persisted as I worked as a Research Scientist at DRDO, the R&D wing of the Indian Government's Ministry of Defence. . These two opportunities in the early years of my career led me to my decision to focus on Industrial Applications of Signal Processing and that’s what I have been doing for more than 30 years of my career. In TCS Research, I pioneered research work in the DSP field, and for the past 20 years, I have been working on various signal processing applications around radar, biomedical, wireless, and inertial sensors building systems with applications in healthcare, industry 4.0, smart retail, communication, and smart city.
2. How does your work affect society?
One of the core works we did in signal processing was to model cardiac disease conditions like Ischemic Heart Blocks (also called Coronary Artery Disease or CAD) using indirectly measured physiological signals like Photoplethysmogram (PPG), Phonocardiogram (PCG) or heart sound, and Electrocardiogram (ECG).
Coronary Artery Disease (CAD) or heart block is the leading cause of deaths not only in the developing countries but worldwide. In 2015, India reported that out of 800 million adults, 61.8 million had CAD which is almost 8% of the total population. This was a whooping 32% increase from 2010 reports. More importantly, there were approximately 8 million new cases of CAD among people aged less than 40 years between 2010 and 2015. A lot of the deaths can be prevented if this is detected early, but only definitive test for this is coronary angiogram which is invasive and is not available everywhere, especially in rural areas and Tier 2-3 cities. We have created a signal processing AI based system that can detect CAD from PPG, PCG and ECG without needing the coronary angiogram. The efficiency of this system is already proven on a small 200-member patient trial at a large hospital in Kolkata, India. As we speak, we are embarking on a countrywide-wide trial of 10K patients with Cardiological Society of India. If the trial is successful, this can impact a large number of people in India and in similar developing countries. Earlier we had done a 100-home trial at Singapore for elderly people living alone where we had sensorized their homes and used signal processing on the sensor data to find out the activities of daily living, creating early alert systems including fall detection.
3. What challenges have you had to face to get to where you are today?
The main challenge is how to marry the technological novelty to a visible and useful impact in the application. When I worked in DRDO, this challenge manifested in designing novel signal processing algorithms and keeping the radar working in multipath fading scenarios in millimetric wave band. In our cardiac health work in TCS, this meant designing signal processing algorithms that can work with non-medical grade wearable sensors that are inherently noisy. In our Industry 4.0 work, this amounted designing signal processing systems that can seamlessly fuse multiple sensor signals at a signal level.
The other challenge is obviously at the platform level where we have come a long way from tiny microcontrollers to DSP processors to AI chipset accelerators. But what has not changed is the fact that initially your algorithm will always be more time/memory/power than what is available in your target hardware and optimizing it to fit the target hardware is always a non-trivial and involves engineering task.
Our brain computes take only 20 Watts while a typical GPU cluster may use tens of kilowatts of power – how do we design AI systems that consumes power in the order of our brain?
4. What advice would you give to scientists/engineers in signal processing?
First advice will be to fall in love with signals, which has so much variety and unpredictability in its morphology as compared to more structured data like text and images - there is a whole lot of unexplored territory in the world of signals, as far as AI is concerned.
Second advice will be to keep an open mind and ready to adapt new technologies / techniques as they come – in today’s world signal processing is a much part of machine learning (ML) as machine learning is part of signal processing. We can do signal processing-based feature engineering for ML and we can also use ML to do adaptive signal processing.
Third advice is to look for application-level impact and then use technology to solve it, rather than going bottom-up to build a novel technology system first and then look for a suitable application.
5. Why did you choose Industry over academic research in signal processing?
The answer lies in the third point of my previous answer. I always wanted to create impact through my inventions. Industrial research has allowed me to work on real-world problems that comes directly from the end users or customers. TCS Research gives us access to problem statements of a wide spectrum of customers across multitude of industry verticals. As a solution to any of these problems ultimately demonstrates the potential impact of the invention, I am always driven by such opportunities in industrial research.
6. Is there anything else you would like to add?
The world is undergoing a disruptive change via AI technologies like generative AI. Today it is demonstrated in the areas of text processing via technologies like chatGPT. However, we need to understand that what is being disruptively possible in these applications today was made possible via significant development of natural language processing that allowed computers to understand the syntactic, semantic, and grammatical aspects. This disruptive change was aided by the fact that language has formal grammar -same is true to some extend for image and videos. But, what is the language and grammar of signals? It is sure that the language is only understood by signal processing and the grammar must be built on top of it - domain-by-domain and sensor-by-sensor. It poses a tremendous opportunity for signal processing practitioners - without this, a disruptive generative AI will not be possible in the world of signals.
The second point I wanted to add was to make the computations power-aware – in the future world of sustainability, green computing/green AI is a must. Signal processing engineers are inherently trained to make their algorithms work on low power, low latency constrained embedded devices. The same principles need to be applied to designing power-efficient AI systems. It is the need of the day in this age of over-parameterized ultra-large and power-hungry AI models. In this context neuromorphic computing architectures that mimic human brain neuronal structures will need to be explored in place of Von-Neumann architectures of CPUs, Harvard Architectures of DSPs or massively parallel architectures of GPUs.
,
on the way down and also o the way up.
! Image solution for the future.
