Deep Learning for Science: A Q&A with NERSC’s Prabhat

By Kathy Kincade

November 7, 2017

Deep learning is enjoying unprecedented success in a variety of commercial applications, but it is also beginning to find its footing in science. Just a decade ago, few practitioners could have predicted that deep learning-powered systems would surpass human-level performance in computer vision and speech recognition tasks.

These tools are now poised to help scientists contend with some of the most challenging data analytics problems in a number of domains. For example, extreme weather events pose great potential risk on ecosystem, infrastructure and human health. Analyzing extreme weather data from satellites and weather stations and characterizing changes in extremes in simulations is an important task. Similarly, upcoming astronomical sky surveys will obtain measurements of tens of billions of galaxies, enabling precision measurements of the parameters that describe the nature of dark energy. But in each case, analyzing the mountains of resulting data poses a daunting challenge.

Prabhat, NERSC

A growing number of scientists are already employing HPC systems for data analytics, and many are now beginning to apply deep learning and other types of machine learning to their large datasets. Toward this end, in 2016 the U.S. Department of Energy’s National Energy Research Scientific Computing Center (NERSC) expanded its support for deep learning and began forming hands-on collaborations with scientists and industry. NERSC users from science domains such as geosciences, high energy physics, earth systems modeling, fusion and astrophysics are now working with NERSC staff, software tools and services to explore how deep learning can improve their ability to solve challenging science problems.

In this Q&A with Prabhat, who leads the Data and Analytics Services Group at NERSC, he talks about the history of deep learning and machine learning and the unique challenges of applying these data analytics tools to science. Prabhat is also an author on two related technical papers being presented at SC17, “Deep Learning at 15PF: Supervised and Semi-Supervised Classification for Scientific Data” and “Galactos: Computing the 3-pt Anisotropic Correlation for 2 Billion Galaxies,” and is conducting two deep learning roundtables in the DOE Booth (#613) at SC17. He is also giving a plenary talk on deep learning for science on Sunday, November 12 at the Intel HPC Developer Conference held in conjunction with SC17.

How do you define deep learning, and how does it differ from machine learning?

At the Department of Energy, we tackle inference problems across numerous domains. Given a noisy observation, you would like to infer properties of the object of interest. The discipline of statistics is ideally suited to solve inference problems. The discipline of Machine Learning lies at the intersection of statistics and computer science, wherein core statistical methods were employed by computer scientists to solve applied problems in computer vision and speech recognition. Machine learning has been around for more than 40 years, and there have been a number of different techniques that have fallen in and out of favor: linear regression, k-means, support vector machines and random forests. Neural networks have always been part of machine learning – they were developed at MIT starting in the 1960s – there was the major development of the back-propagation algorithm in the mid-1980s, but they never really picked up until 2012. That is when the new flavor of neural networks – that is, deep learning – really gained prominence and finally started working. So the way I think of deep learning is as a subset of machine learning, which in turn is closely related to the field of statistics, and all of them have to do with solving inference problems of one kind or another.

What technological changes occurred that enabled deep learning to finally start working?

Three important trends have happened over the last 20 years or so. First, thanks to the internet, “big Data,” or large archives of labeled and unlabeled datasets, has become readily accessible. Second, thanks to Moore’s Law, computers have become extremely powerful. A laptop featuring a GPU and a CPU is more capable than supercomputers from previous decades. These two trends were prerequisites for enabling the third wave of modern neural nets, deep learning, to take off. The basic machinery and algorithms have been in existence for three decades, but it is only the unique confluence of large datasets and massive computational horsepower that enabled us to explore the expressive capabilities of Deep Networks.

What are some of the leading types of deep learning methods used today for scientific applications?

As we’ve gone about systematically exploring the application of deep learning to scientific problems over the last four years, what we have found is that there are two dominant architectures that are relevant to science problems. The first is called the convolutional network. This architecture is widely applicable because a lot of the data that we obtain from experimental and observational sources (telescopes and microscopes) and simulations – tend to be in the form of a grid or an image. Similar to commodity cameras, we have 2D images, but we also typically deal with 3D, 4D and multi-channel images. Supervised pattern classification is a common task shared across commercial and scientific use cases; applications include face detection, face recognition, object detection and object classification.

The second approach is more sophisticated and has to do with the recurrent neural network: the long short-term memory (LSTM) architecture. In commercial applications, LSTMs are used for translating speech by learning the sequence-to-sequence mapping between one language and another. In our science cases, we also have sequence-to-sequence mapping problems, such as gene sequencing, for example, or in earth systems modeling, where you are tracking storms in space and time. There are also problems in neuroscience that take recordings from the brain and use LSTM to predict speech. So broadly those two flavors of architectures – convolutional networks and LSTMs – are the dominant deep learning methodologies for science today.

In recent years, we have also explored auto-encoder architectures, which can be used for unsupervised clustering of datasets. We have had some success in applying such methods for analysis of galaxy images in astronomy, and Data Bay sensor data for neutrino discovery. The latest trend in deep learning is the generative adversarial network (GAN). This architecture can be used for creating synthetic data. You can feed in examples from a certain domain, say cosmology images or Large Hadron Collider (LHC) images, and the network will essentially learn a process that can explain these images. Then you can ask that same network to produce more synthetic data that is consistent with other images it has seen. We have empirical evidence that you can use GANs to produce synthetic cosmology or synthetic LHC data without resorting to expensive computational simulations.

What is driving NERSC’s growing deep learning efforts, and how did you come to lead these efforts?

I have a long-standing interest in image processing and computer vision. During my undergrad at IIT Delhi, and grad studies at Brown, I was intrigued by object recognition problems, which seemed to be fairly hard to solve. There was incremental progress in the field through the 1990s and 2000s, and then suddenly in 2012 and 2013 you see this breakthrough performance in solving real problems on real datasets. At that point, the MANTISSA collaboration – a research project originally begun when I was part of Berkeley Lab’s Computational Research Division – was exploring similar pattern detection problems, and it was natural for us to explore whether deep learning could be applied to science problems. We spent the next three to four years exploring applications in earth systems modeling, neuroscience, astronomy and high energy physics.

When a new method/technology comes along, one has to make a judgment call on how long you want to wait before investing time and energy in exploring the possibilities. I think the DAS group at NERSC was one of the early adopters. We recognized the importance of this technique and demonstrated that it could work for science. In the experimental and observational data community, there are a lot of examples of domain scientists who have been struggling with pattern recognition problems for a long time. And now the broader science community is waking up to the possibilities of machine learning to help them solve these problems.

What is NERSC’s current strategy for bringing deep learning capabilities to its users?

Since NERSC is a DOE Office of Science national user facility, we listen to our users, track their emerging requirements and respond to their needs. Our users are telling us that they would like to explore machine learning/deep learning and see what it can do for them. We currently have about 70 users who are actively using deep learning software at NERSC, and we want to make sure that our software, hardware, policies and documentation are all up to speed. Over the past two years, we have worked with the vendor community and identified a few popular deep learning frameworks (TensorFlow, Caffe, Theano and Torch) and have deployed them on Cori. In addition to making the software available, we have documentation and case studies in place. We also have in-depth collaborations in about a dozen areas where NERSC staff, mostly from the DAS group, have worked with scientists to help them explore the application of deep learning. And we are forming strategic relationships with commercial vendors and other research partners in the community to explore the frontier of deep learning for science.

Do certain areas of scientific research lend themselves more than others to applying deep learning?

Right now our success stories span research sponsored by several DOE Office of Science program offices, including BER, HEP and NP. In earth systems modeling, we have shown that convolutional architectures can extract extreme weather patterns in large simulations datasets. In cosmology, we have shown that CNNs can predict cosmological constants, and GANs can be potentially used to supplement existing cosmology simulations.  In astronomy, the Celeste project has effectively used auto-encoders for modeling galaxy shapes. In high energy physics, we are using convolutional architectures for discriminating between different models of particle physics, exploring LSTM architectures for particle tracking. We’ve also shown that deep learning can be used for clustering and classifying various event types at the Daya Bay experiment.

So the big takeaway here is that for the tasks involving pattern classification, regression and creating fast simulators, deep learning seems to do a good job – IF you can find training data. That’s the big catch – if you have labeled data, you can employ deep learning. But it can be a challenge to find training data in some domain sciences.

Looking ahead, what are some of the challenges in developing deep learning tools for science and applying them to research projects at NERSC and other scientific supercomputing facilities?

We can see a range of short-term and long-term challenges in deep learning for science. The short-term challenges are mostly pragmatic issues pertaining to development, enhancement and deployment of tools. These include handling complex data; scientific data tends to be very diverse (compared to images and speech), we are working with 2D, 3D, even 4D data and the datasets can be sparse or dense and defined over a regular, or irregular grid. Deep learning frameworks will need to account for this diversity going forward. Performance and scaling are also barriers. Our current networks can take several days to converge on O(10) GB datasets, but several scientific domains would like to apply deep learning to 10TB-100TB datasets. Thankfully, this problem is right up our alley at HPC centers.

Another important challenge faced by domain scientists is hyper-parameter tuning: Which network architecture do you start with? How do you choose an optimization algorithm? How do you get the network to converge? Unfortunately, only a few deep learning experts know how to address this problem; we need automated strategies/tools. Finally, once scientific communities realize that deep learning can work for them, and access to labeled datasets is the key barrier to entry, they will need to self-organize and conduct labeling campaigns.

The longer-term challenges for deep learning in science are harder, by definition, and include a lack of theory, interpretability, uncertainty quantification and the need for a formal protocol. I believe it’s very early days in the application of deep learning to scientific problems. There’s a lot of low-hanging fruit in publishing easy papers that demonstrate state-of-the-art accuracy for classification, regression and clustering problems. But in order to ensure that the domain science community truly embraces the power of deep learning methods, we have to keep the longer term, harder challenges in mind.

About the Author

Kathy Kincade is a science & technology writer and editor with the Berkeley Lab Computing Sciences Communications Group.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

Atos Outlines Plans to Get Acquired, and a Path Forward

May 21, 2024

Atos – via its subsidiary Eviden – is the second major supercomputer maker outside of HPE, while others have largely dropped out. The lack of integrators and Atos' financial turmoil have the HPC market worried. If Read more…

Core42 Is Building Its 172 Million-core AI Supercomputer in Texas

May 20, 2024

UAE-based Core42 is building an AI supercomputer with 172 million cores which will become operational later this year. The system, Condor Galaxy 3, was announced earlier this year and will have 192 nodes with Cerebras Read more…

Google Announces Sixth-generation AI Chip, a TPU Called Trillium

May 17, 2024

On Tuesday May 14th, Google announced its sixth-generation TPU (tensor processing unit) called Trillium.  The chip, essentially a TPU v6, is the company's latest weapon in the AI battle with GPU maker Nvidia and clou Read more…

ISC 2024 Student Cluster Competition

May 16, 2024

The 2024 ISC 2024 competition welcomed 19 virtual (remote) and eight in-person teams. The in-person teams participated in the conference venue and, while the virtual teams competed using the Bridges-2 supercomputers at t Read more…

Grace Hopper Gets Busy with Science 

May 16, 2024

Nvidia’s new Grace Hopper Superchip (GH200) processor has landed in nine new worldwide systems. The GH200 is a recently announced chip from Nvidia that eliminates the PCI bus from the CPU/GPU communications pathway.  Read more…

Europe’s Race towards Quantum-HPC Integration and Quantum Advantage

May 16, 2024

What an interesting panel, Quantum Advantage — Where are We and What is Needed? While the panelists looked slightly weary — their’s was, after all, one of the last panels at ISC 2024 — the discussion was fascinat Read more…

Atos Outlines Plans to Get Acquired, and a Path Forward

May 21, 2024

Atos – via its subsidiary Eviden – is the second major supercomputer maker outside of HPE, while others have largely dropped out. The lack of integrators Read more…

Google Announces Sixth-generation AI Chip, a TPU Called Trillium

May 17, 2024

On Tuesday May 14th, Google announced its sixth-generation TPU (tensor processing unit) called Trillium.  The chip, essentially a TPU v6, is the company's l Read more…

Europe’s Race towards Quantum-HPC Integration and Quantum Advantage

May 16, 2024

What an interesting panel, Quantum Advantage — Where are We and What is Needed? While the panelists looked slightly weary — their’s was, after all, one of Read more…

The Future of AI in Science

May 15, 2024

AI is one of the most transformative and valuable scientific tools ever developed. By harnessing vast amounts of data and computational power, AI systems can un Read more…

Some Reasons Why Aurora Didn’t Take First Place in the Top500 List

May 15, 2024

The makers of the Aurora supercomputer, which is housed at the Argonne National Laboratory, gave some reasons why the system didn't make the top spot on the Top Read more…

ISC 2024 Keynote: High-precision Computing Will Be a Foundation for AI Models

May 15, 2024

Some scientific computing applications cannot sacrifice accuracy and will always require high-precision computing. Therefore, conventional high-performance c Read more…

Shutterstock 493860193

Linux Foundation Announces the Launch of the High-Performance Software Foundation

May 14, 2024

The Linux Foundation, the nonprofit organization enabling mass innovation through open source, is excited to announce the launch of the High-Performance Softw Read more…

ISC 2024: Hyperion Research Predicts HPC Market Rebound after Flat 2023

May 13, 2024

First, the top line: the overall HPC market was flat in 2023 at roughly $37 billion, bogged down by supply chain issues and slowed acceptance of some larger sys Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Some Reasons Why Aurora Didn’t Take First Place in the Top500 List

May 15, 2024

The makers of the Aurora supercomputer, which is housed at the Argonne National Laboratory, gave some reasons why the system didn't make the top spot on the Top Read more…

Leading Solution Providers

Contributors

Eyes on the Quantum Prize – D-Wave Says its Time is Now

January 30, 2024

Early quantum computing pioneer D-Wave again asserted – that at least for D-Wave – the commercial quantum era has begun. Speaking at its first in-person Ana Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

The GenAI Datacenter Squeeze Is Here

February 1, 2024

The immediate effect of the GenAI GPU Squeeze was to reduce availability, either direct purchase or cloud access, increase cost, and push demand through the roof. A secondary issue has been developing over the last several years. Even though your organization secured several racks... Read more…

Intel Plans Falcon Shores 2 GPU Supercomputing Chip for 2026  

August 8, 2023

Intel is planning to onboard a new version of the Falcon Shores chip in 2026, which is code-named Falcon Shores 2. The new product was announced by CEO Pat Gel Read more…

The NASA Black Hole Plunge

May 7, 2024

We have all thought about it. No one has done it, but now, thanks to HPC, we see what it looks like. Hold on to your feet because NASA has released videos of wh Read more…

GenAI Having Major Impact on Data Culture, Survey Says

February 21, 2024

While 2023 was the year of GenAI, the adoption rates for GenAI did not match expectations. Most organizations are continuing to invest in GenAI but are yet to Read more…

How the Chip Industry is Helping a Battery Company

May 8, 2024

Chip companies, once seen as engineering pure plays, are now at the center of geopolitical intrigue. Chip manufacturing firms, especially TSMC and Intel, have b Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire