Faster. Cheaper. Smaller. IBM FlashSystem 9100. – Get the Jeskell eBook.    |    1-877-JESKELL     |

How to Get Fast Results in Artificial Intelligence.

Dec 11, 2018 10:44:53 AM | Infrastructure Optimization, Analytics

Artificial intelligence has moved front and center as a consideration for federal agencies as they seek to improve customer experience, improve mission delivery and modernize.

Use cases for artificial intelligence are everywhere. A few examples:

  • Natural language processing. Federal agencies serve people with a wide range of cultures and languages. AI can enable spoken word applications to train for variations in syntax or accent. Combined with rapid processing of multiple data sets this has the potential to greatly improve customer support from federal call centers. Similarly, visitors to federal web sites can receive enhanced customer experiences when interactions are AI-powered.
  • Computer vision. Video and still imagery are exploding in volume and – thanks to high definition – detail. AI-enabled surveillance can look for specific individuals, activities or objects in both law enforcement and military applications.
  • Recommendation engines and predictive analytics. While AI cannot replace the human responsibility for decision-making, it can ingest large amounts of history and combine them with complex rules sets to produce credible and auditable decision recommendations in the many federal adjudication applications.

AI starts with the application

Just what is artificial intelligence? More than fast pattern recognition or gaming, AI is a technique for making computers learn and improve how they respond to specific desired outcomes as the AI algorithms are exposed to more, and more carefully curated, data. Success comes from a clear understanding of what you want the system to do. It’s also important to choose a software system with an algorithm that can “show its work, giving and auditable and transparent view into how it made decisions or recommendations. This is a key selection criteria in federal settings.

Advances in machine learning and artificial intelligence have reached a tipping point that puts new possibilities in reach for agencies with a variety of problem sets. Moreover, vertically integrated companies like IBM are producing hardware with the power and affordability to put AI within reach of nearly every agency.

But actually implementing AI – that’s another matter. Because AI is more than an application for, say, robotic process automation or enterprise resource planning, it takes a different kind of planning. It requires a nearly total re-imagination of processes designed to utilize the self-learning and self-adapting nature of true AI systems. Those initial steps – the business proposition for AI – are where agency expertise is best applied.

For example, many agencies adjudicate cases of one form for another. AI can speed the decision-making process not by making the decision – that’s reserved for human beings – but rather by ingesting data from multiple sources and quickly producing a range of compliance-based recommendations.

For the many agencies that pay weekly or monthly benefits or reimburse claims, AI-powered predictive analytics can point to potential cases of fraud much faster than existing techniques – before excessive improper payments go out.

Success in artificial intelligence also requires an IT infrastructure optimized for AI’s compute and data intensity, and specialized development tools. To achieve those step-function mission improvements in areas like citizen experience, military advantage, an agency’s systems must be up to the task.

Thus as agencies think about the interrelated policy objectives of modernization, improved customer experience, and better mission performance, they’ll need to think about how their infrastructures will support AI.

Building blocks

As we’ve stated, before the first server is installed, agencies must fully understand where and how they will apply artificial intelligence. In this stage, program, IT and agency management do their analysis of priorities and pain points in deciding where application of AI can produce the most results, the most quickly.

After that, it’s important to understand that ingestion of large data sets typically associated with AI and the training of the AI algorithms are highly compute-intensive activities. Like weather prediction or space flight calculations, AI is ideally done on a designated hardware platform that won’t hang up the project. To meet user expectations, IT staff will need a platform that will train models fast, keep iteration cycles short, and increase productivity. But that doesn’t mean you have to use agency staff to design and build such a system.

The major cloud providers offer AI capabilities. But they can’t always match the performance of an on-premises, dedicated AI hardware stack.

AI systems start with the basic processing capacity. Once confined to gaming and similar applications, powerful – and affordable – graphics processing units (GPUs) are powering a growing number of federal AI applications. GPUs, it so happens, are optimized not only for rendering but also for operations requiring highly repetitive processes. For example, repeating the same mathematics operations rapidly over large structured data sets.

An effective architecture for maximizing GPU efficiency is to tie the GPU to a fast CPU designed to “feed” the GPU by handling data fetching, caching and I/O. A great choice is IBM’s Power Systems, which include include GPU accelerator servers that use IBM’s incomparable POWER9 chip. In one experiment, a POWER9 accelerator coupled with an Nvidia GPU resolved in about 90 seconds a logistic regression classifier that took 70 minutes on a popular cloud platform using the open source TensorFlow library of tools.

IBM’s PowerAI platform gives users multi-node, multi-GPU power. IBM’s Distributed Deep Learning modules provide an efficient, distributed framework for running open source AI libraries. As a primary IBM partner, Jeskell’s implementation services save agencies the trouble of installing and configuring a cluster. That means faster onboarding of users. Plus, Jeskell’s engineers will work with the agency customer to make sure the resulting cluster addresses all the customer’s unique operational requirements and constraints.

A well-designed cluster accelerates training and avoid processing bottlenecks using host-based NVNe cards.

Of course, the purpose of hardware is to run software. The burgeoning market for AI frameworks and programming tools can make for a bewildering set of choices. Few federal agencies have the internal programming capability to build an AI system. Nor do they have the budgets – in time or dollars – to roll their own. If speed to results is a priority – as it is for nearly every agency – it’s wise to consider a pre-assembled software bundle that takes the expense and guesswork out of choosing your own software stack.

Case point: IBM’s PowerAI Vision suite. It’s designed to get organizations going fast, so subject matter experts can spend time exploring answers. It abstracts many of the routine but large tasks that occupy so much time for data scientists. For example, it applies deep learning models to automate data labeling and thereby squeezing time out of AI projects.

IBM’s Power AI Vision is a software library built from open source components, enhanced with IBM experience, and optimized for IBM’s Power AI hardware systems. Power AI Vision is a deep machine learning suite covering all the steps in building AI applications from installing the learning environment to retraining the initial model when new data becomes available.

Too often, barriers to entry into artificial intelligence are higher than they need to be. To be sure, a ready made system will require an initial outlay, but the time and labor costs it will avoid should also be part of the calculus. The turnkey approach to infrastructure lets an agency put AI to work faster, without worrying about the technical minutia of data science, workflows and hyperparameters.

Find a great partner

Having an integration partner well versed in the particulars of the government market can further speed AI results. Jeskell is such a partner to many federal agencies, having served the federal market exclusively since its founding in 1991.

Jeskell, a long term partner of IBM, has deep expertise in design, installation and configuration of Power9 systems and the AI Vision software they’re designed to run. Power9 is an AI “secret sauce” – it simply can’t be matched by x86 systems for speed and availability.

The company is bringing its experience to artificial intelligence projects to both the Defense Department and to civilian agencies such as the National Institute of Standards and Technology.  Many clients are looking at AI to help with analyzing exploding volumes of video surveillance data. As they learn, the systems can more accurately and consistently pinpoint objects of interest or the presence of a specific threats.

Jeskell has designed and built Power9 clusters that can also accelerate a variety of tasks to reduce the dependency on human operators to evaluate and categorize items in audio and signal intelligence. These systems continually improve the fidelity of the algorithms used to identify objects or patterns of interest.

In short, agencies have an AI stack available along with the expertise to help them get up and running. It starts with IBM Power hardware, together with the required supporting libraries, topped by deep learning frameworks. All delivered, installed and configured by Jeskell, an experienced small business that’s IBM’s leading AI partner.

For more information on AI and what problems it can solve for your organization, reach out to me any time.

Contact Steve

- Steve Koppenhafer, Systems Architect

< Back to All
Steve Koppenhafer

Steve Koppenhafer

For more information on this topic, contact me below.

Contact Us

Subscribe now to get our latest news