Boosting HPC Solutions with Deep Learning

In my last post, I explored the opportunities created by running AI applications on HPC infrastructure (AI-on-HPC). Now for the flip side: the potential to augment HPC with the addition of AI techniques (HPC-on-AI). Adding AI capabilities derived from tried-and-true Deep Learning (DL) practices can lead to substantial efficiencies for HPC.

The surprisingly-effective practices of Deep Learning that matured in the last 5 years are an excellent match to the type of problems being addressed within HPC. DL capabilities can be described, albeit with oversimplification, as identifying patterns in a multi-dimensional dataset. Several kinds of tasks are especially suited for DL, in particular classification of patterns (such as recognition of images), clustering patterns (e.g. identifying increased risks from life signs monitors), and anomaly detection (as in Fraud Detection).

When brought to scale, the same basic capabilities can be applied to some of the most complex HPC domains.

HPC is typically applied to challenges requiring very large datasets and massive amounts of computing, storage and networking. Typical HPC domains such as Bio-Sciences, Geosciences, and Chemical Engineering are addressed with a set of tools and capabilities that evolved over time in the areas of Modeling & Simulation as well as Advanced Analytics. AI can provide a very powerful additional set of capabilities and techniques, especially when it comes to High Performance Data Analytics.

Here are a few examples of these DL applications where data analytics play a significant role:

• In Meteorology, they enable early identification of forming storm conditions.
• In Astronomy, the technologies helped address some of cosmology's hardest challenges.
• In medicine, they are being used for genomic sequence analysis (variant detection, mRNA differential splicing and expression prediction), to identify abnormal tissue and potential tumor sites in scans and tomography, and to map the mind in real time.

A key characteristic of DL is that it does not need equations or a mathematical model to define the patterns being analyzed; it learns to track the patterns by processing large amounts of data, and by being assisted, with a varying degree of supervision, on pinpointing results. The DL models allow for data-driven discovery at scale and inherently open to new learning as new data is ingested.

Those characteristics can also be applied to HPC areas where the space is not easily modeled and is continuously evolving. FinTech is a great example, since it evaluates trends and events which are dynamically changing at high speed in an environment where one cannot have all the equations to define the complete model at any given point in time.

Security threat detection is another area where DL techniques allow systems to stay in pace with rapidly evolving patterns and signatures patterns while still effectively identifying likely anomalies and faint signals within very noisy data deluge.

Lastly, the ability to apply localized learning allows for creation of solutions that combine global datasets with those that are specific to a particular individual or setting. For example, DL is expected to bring about breakthroughs in precision medicine, with disease treatment and prevention that takes into account individual variability in genes, environment, and lifestyle.

The adoption of AI techniques in HPC large-scale solutions is underway, though still at an early stage. Given the underlying strengths of Deep Learning and other Machine Learning technologies, the successes demonstrated so far indicate that combining AI with HPC holds great promise for the future.

The next post will venture into the future of HPC and AI, and will offer some projections on possible or likely trends and significant advancements of the coming 3-5 years.

Thanks for your interest,

 Gadi

Below is a video which illustrates the application of AI techniques to track mind states at HPC scale, in a collaboration between Intel and Princeton University: