- AIM interviews Dr Ganapathi Pulipaka on gallium nitride processors for future space exploration, deep quest of AI, deep learning algorithms, reinforcement learning and high-performance computing.
Gallium nitride (GaN) processors are poised to power the next-generation semiconductor industry to revolutionize space exploration. Gallium nitride compound has been used in building light-emitting diodes (LEDs) and is highly efficient in conducting the electrons 1000 way better than silicon material. These are currently researched to explore in building high-temperature microprocessors for space applications.
The silicon powered integrated circuits start malfunctioning once they reach 300 degrees celsius. However, GaN material is highly stable chemically in high-radiation and temperature environments. There is still a lot of research going on in developing metal-oxide-semiconductor transistors. The GaN transistors and GaN chip architecture are way more complex than silicon chips. As a matter of fact, NASA is researching to produce a GaN processor that can operate efficiently above 500 degrees celsius. GaN processors can withstand various space missions in high-temperature planet worlds such as Venus, Mercury, Jupiter, Saturn, Uranus, or Neptune.
Gallium nitride processors for landing on 500+ degrees high-temperature planets.
Analytics India Magazine has interviewed Dr Ganapathi Pulipaka to understand several aspects of machine learning, deep learning, and mathematics needed for practitioners. An adept data engineer and data architect turned data scientist, Dr Pulipaka prolifically posts about data science trends, keeping the audience up-to-date on the latest information. He covered several topics in-depth on OpenAI, deep neural networks and fractal geometry, to name a few. Dr Pulipaka has written two books so far, and there is an upcoming book about to be published in October.
“Until we have infinite processing power on machines, we cannot produce the outcomes we need” – GP
Edited excerpt —
AIM: In several interviews and magazines, you mentioned that you assembled your first computer on your own at 16 years old and cited that as your entry point for your passion for technology. Did you always plan to do that to be able to program and build hardware and software applications? Also, why did you choose to build your computer?
GP: Well, it was not something I planned to build earlier. It was something my brain told me to do, and I explored the prospect of building a machine. The outcome and results were amazing.
A lot of people had laptops at the age of 16, but I didn’t. Laptops back in the day were a little bit expensive to own. The easier way to go was to build your own computer. If you can assemble all the components, you can build a computer, which was why I actually started building it.
In today’s world, it’s even harder to upgrade RAM on Macbook Air, Macbook Pro, or iMac as RAM is soldered to the motherboard. Though some electronic engineers successfully did such upgrades, it would be too risky and void the warranty, as forcefully removing the original soldered hardware from the motherboard may cause compatibility issues with the operating system. Getting another battery or SSD would be a much easier option to deal with.
AIM: Were there any freelance projects that you implemented before entering into corporate enterprise-wide rollouts?
GP: Right after I built my computer, I immediately pursued a few projects I wanted to implement as a freelancer while still in school. Back in the day, the Air Force Academy wanted to implement a traffic system project, handling the takeoff, landing, and predicting the weather patterns by creating an application or detouring a plane. So I approached the Wing Commander, explained my passion, and was kind enough to offer it to me. To facilitate that, I built an SQL database, primitive in nature, and had written a C++ front end. As my first project, I even received an appreciation letter from the Wing Commander for it.
I was also very interested in mathematics and statistics, which is important for solving data science equations such as Ordinary Differential Equations and Partial Differential Equations, linear algebra, and forecast problems. For example, many companies want to build a trajectory of performing for different claims or different manufacturing locations. There’s a lot of historical data that they can gather and build those time series forecasting models. I have implemented a few more projects for those companies, which I will be covering in another interview.
In 1996, I worked for another company freelancing, where I built a complete time series forecasting system and went live with it. Creating this system got me more interested in mathematical models and statistical models. In addition, I was interested in understanding how to go about the whole gamut of landscapes and infrastructure and architectures and how all of these systems integrate.
AIM: Can you share some light on other projects that you have worked on so far?
I’ve been researching for more than ten years. That’s when I got more interested in deep learning and machine learning. Before that, I was focused on data engineering, programming, data architecture on databases of SQL, NoSQL, data science, mathematics etc.
Reverse-mode automatic differentiation of ODE
Computation graph of the latent ODE model:
I also worked for many clients, like for an aerospace IoT integration, where it is important to know or forecast when something will break down, like a bridge inspection, to prevent a collapse. In a plane, you don’t know when a panel is going to break down. Here, there is an acoustic emission, and if you measure the acoustic emission with fractal geometry, that can tell us exactly when there is a change between the previous state of this panel versus the current state. There are different factors that you can calculate and provide different parameters. That was another interesting project to work on.
AIM: Let’s talk a little bit about your research projects.
On the research side, I’ve done implementations of pretty much all the algorithms. And there’s a book that’s coming out with my research — A Greater Foundation for Machine Engineering: The hallmarks of the Great Beyond in TensorFlow, R, PyTorch, and Python, where I’m covering whatever that I picked up and what I thought would be useful for other people in this field. It includes Python, TensorFlow, all the machine learning algorithms in Python, and a lot of statistical forecasting and other stuff built using Python.
GP’s Upcoming book, reviewed by Gloria Pullman of Reader’s Digest
Reinforcement Learning is also a very interesting area because many people are into Artificial General Intelligence (AGI). Reinforcement learning primarily shifts the paradigm of having supervised, unsupervised and semi-supervised machine learning/deep learning landscapes. It takes us into a completely different type of approach from Bellman back in the day.
I typically include a lot of natural language programming, leveraging NLTK, Gensim, and a number of other Python libraries. I think that NLP will have significant advancements in the future, that there’s going to be GPT-4 perhaps. The whole GPT-4 singularity depends on the recent AGI research based on 999 AI researchers, where they all express that it is possible to achieve singularity by 2060.
AIM: Can you highlight some key aspects mentioned in your upcoming book?
The natural language toolkit (NLTK) is primarily a collection of several libraries for text processing and natural language processing within Python. The University of Pennsylvania developed this library that allows the user to perform several NLP operations such as splitting texts and sentences from multiple paragraphs, identifying whether the statement is a question or an answer, detecting the part of speech of the words, and splitting up multiple words. You can find plenty of such NLP implementations in my upcoming book on Python.
I start with fundamentally very simple algorithms like linear and logistic regression so that the readers can understand the concepts. Then they can proceed with advanced concepts like reinforcement learning.
I got this idea when I tried to do some video training classes on reinforcement learning with the book publishing company Packt and QuickStart boot camp programs. Although I was delivered educational programs on machine learning, cloud computing, mathematics, and statistics, some people were having issues because they went right into reinforcement learning without knowing what machine learning was. So, that’s when I started introducing the fundamentals before moving on to advanced concepts. In the book, I explore some parts of history, like how it began with all the elements, how it advanced and where we are now.
The book coming out is 80% nothing but my dissertation on deep learning in high-performance computing environments, along with some compiled other work I’ve done for many years.
AIM: Please share a bit about your next dissertation on high-performance computing and why did you choose this topic?
My next dissertation will be on high-performance computing (HPC) because it will be completely different from my previous dissertations on data engineering, big data analytics and in-memory computing. For example, we have multiprocessing for MPI (Message Passing Interface –an open-source library standard for the distributed memory parallelization) available in C. We also need to develop the parallel processing capabilities for HPC. Python and Java have their own unofficial bindings for MPI, but most of the scientific engineering labs leverage it in C. MPI technique faces latency when handling millions of tasks through collective calls plagued with operating system crashes and errors due to scalability memory overloads. The MPIs are implemented primarily as libraries through distributed memory computing for high-performance computing.
That’s where you need to chunk the blocks — just like how we’re using Apache Hadoop with map, shuffle, and reduce technique — use cloud computing clusters to break down the data and process heaps of math equations, assemble the data back before you get it back into your project. MapReduce is a technique adopted by Apache Hadoop for processing a variety of high-performance computing applications, where the big data in petabyte size reside on the nodes of the computing machines. Each dedicated storage unit of the computing node cluster performs I/O-intensive, CPU-intensive, GPU-intensive Hadoop, YARN.
Research on HPC in nuclear plasma turbulence is already talking about nuclear weapons transportation. As a matter of fact, Sandia National Laboratories is developing landmark HPC algorithms for it. These are all parallel processing algorithms and will be intensive, but that’s where I would see myself in future.
“The brain is orchestrating all at the same time, so it’s a different type of mechanism altogether. If we can understand it, we’ll come up with more innovative models.”
Chen, R. T., Rubanova, Y., Bettencourt, J., & Duvenaud, D. (2019). Neural Ordinary Differential Equations. Retrieved from https://arxiv.org/pdf/1806.07366.pdf
Collins M.D. PhD, F. S., & Fink, L. (1995). The Human Genome Project. NCBI – US National Library of Medicine National Institutes of Health National Institutes of Health, 19(3)(), 190 – 195. Retrieved from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6875757/
Dilmegani, C. (2017). Will AI reach singularity by 2060? 995 experts? opinions on AGI. Retrieved from https://research.aimultiple.com/artificial-general-intelligence-singularity-timing/
DK, P. (2015). High-Performance Deep Learning (HiDL) – The Ohio State University. Retrieved from http://hidl.cse.ohio-state.edu
Du, Q., & Li, X. (2013, October 13). Li, Z., Shen, H., Denton, J., & Ligon, W. (2016, December 8th, 2016). Comparing application performance on HPC-based Hadoop platforms with local storage and dedicated storage. IEEE International Conference on Big Data (Big Data). http://dx.doi.org/10.1109/BigData.2016.7840609. IEEE Proceedings of 2013 3rd International Conference on Computer Science and Network Technology. http://dx.doi.org/10.1109/ICCSNT.2013.6967143
Hao, K. (2019). The computing power needed to train AI is now rising seven times faster than ever before. Retrieved from https://www.technologyreview.com/2019/11/11/132004/the-computing-power-needed-to-train-ai-is-now-rising-seven-times-faster-than-ever-before/
Li, Z., Shen, H., Denton, J., & Ligon, W. (2016, December 8th, 2016). Comparing application performance on HPC-based Hadoop platforms with local storage and dedicated storage. IEEE International Conference on Big Data (Big Data). http://dx.doi.org/10.1109/BigData.2016.7840609
Nowakowski, T. (2017). Gallium nitride processor—next-generation technology for space exploration. Phys Org. Retrieved from https://phys.org/news/2017-12-gallium-nitride-processornext-generation-technology-space.html
Pullman, G. (2021). The AI Bible: A Gigantic Masonry on a Greater Foundation of Machine Learning Engineering. Reader’s Digest. Retrieved from https://www.readersdigest.co.uk/inspire/down-to-business/the-ai-bible-a-gigantic-masonry-on-a-greater-foundation-of-machine-learning-engineering
Join Our Discord Server. Be part of an engaging online community. Join Here.
Subscribe to our NewsletterGet the latest updates and relevant offers by sharing your email.
Dr Ganapathi Pulipaka is Chief AI HPC Scientist and bestselling author of books covering AI infrastructure, supercomputing, high-performance computing for HPC, parallel computing, neural network architecture, data science, machine learning, and deep learning in C, C++, Java, Python, R, TensorFlow, and PyTorch on Linux, macOS, and Windows.