Design, deliver, and run enterprise blockchain workloads quickly and easily.
How to prepare your business for AI
Taming data is one of humankind's biggest challenges, and it has exceeded our ability to efficiently use reasoning or intuition to make sense of patterns within big data.
"The data flood is becoming a universal problem," says Eng Lim Goh, vice president and chief technical officer for SGI, Hewlett Packard Enterprise. "But with artificial intelligence (AI), a wild guess becomes an intelligent guess."
If you're considering AI for a key part of your IT infrastructure, pause here. A few of HPE's AI experts share their thoughts on where AI is headed, how it can transform a business, and the steps to get started.
AI grows up
AI has been showing potential for years, and companies have noticed. IDC predicts that global spending on cognitive and AI solutions will increase at a compound annual growth rate of 54.4 percent over the next few years, exceeding $46 billion by 2020.
AI has attracted this level of corporate interest because it can discover patterns in big data. Celebrated victories of IBM's Deep Blue over Garry Kasparov in chess in 1997 and Alphabet's AlphaGo over Lee Sedol in Go in 2015 highlight the potential of AI to navigate a landscape of vast choices and consequences.
This year, a new type of AI emerged. A program called Libratus, which runs on a supercomputer called Bridges, located at the Pittsburgh Supercomputing Center, defeated four professional poker players in the game of no-limit Texas hold 'em.
Chess and Go are complete information games. Each player can see the opponent's entire board position. But with poker, no player is in possession of every piece of information in the game. (In no-limit Texas hold 'em, two cards in every player's hand are known only to the player in question.) As a result, the approaches that bested Kasparov and Sedol would not work against the poker pros.
The Go problem, for example, relied on deep learning, the application of deep artificial neural networks to learning tasks. Artificial neural networks are modeled on the human brain and consist of layers of computation—input, multiple layers of processing, and output. Deep learning allows a system to automatically discover the hidden semantics and data representations needed to analyze complex raw data consisting of images, audio, and text. These representations help distinguish what you're looking for ("cat") from all the background noise ("not cat").
Libratus doesn't do that. Instead, it applies strategic reasoning based on the Nash Equilibrium, a game theory scenario in which no player has an incentive to deviate from their chosen strategy after considering an opponent's choice. This approach does not require training the AI by feeding it a large data set of, say, photos labeled "cat" and even more photos labeled "not cat." Instead, the AI applies a set of rules and an elementary strategy to select the best move at each point in the game.
Libratus was designed to find solutions from incomplete data. Unlike traditional machine learning technologies that recognize data patterns, it suggests the proper response for each discrete situation. Unlike other types of AI, Libratus does not simply provide "advice" or predictive analysis, says Daniel Wu, chief technologist for data systems and data science at HPE. Instead, it adds a "prescriptive" level to the processing. In other words, it is capable of taking an action based on its determination that a given action is likely to be the most beneficial.
Whether you are talking about deep learning, strategic thinking, or another species of AI, the foundation of its use is in situations that require lightning-fast responses, faster than humans are capable of responding. In applications like autonomous driving and network monitoring, where delay can result in either physical or digital harm, "you just have to let the AI take it over and take it away," says Wu.
AI promises to save time and money in applications such as customer service, where a robot answering in a second is cheaper than a human answering in minutes. The problem is that "every AI is a custom application," according to Pankaj Goyal, vice president of HPE's AI Group. In other words, it takes time, money, and skills to set up new AI in a way that makes business sense for your company.
Step by step
"No business can avoid the AI reality," says Wu. "Remember those diehards who swore they weren't going to use the Internet? Now they're running ads on Google or Facebook. To run a business, you need information and insight. The Internet gives you information but not insight. AI gives you both."
But AI is a "next big thing" characterized by hype and fear. The idea that AI is going to solve every problem is almost as common as the idea that we're headed for a robot apocalypse. How can we keep our heads while we come to terms with the practical reality of AI?
"As individuals, we interact with AI on a daily basis," says Goyal. "Think Alexa, for instance. But for enterprise customers, it is not real. They're salivating at this idea, but they don't know how to do it."
Goyal encourages anyone interested in applying AI to their business to take these three steps in order:
- Ask. Where can you apply AI techniques to get real improvements? Understand the available use cases for AI, and note those places where you can apply AI techniques to get real benefits—for instance, saving employee hours, decreasing response times, and saving expense on material repair and replacement.
- Experiment. You can't apply AI at a large scale without experimenting first. It took Google and Amazon "a long time to get AI in their DNA," Goyal says. "What is the right data infrastructure? What kind of human capital do you need?" Do you have the required data—volume and quality? Do you have the right people? Do you know how to set up the basic infrastructure? Do you know how to tune your model? It takes time to get AI right. Make sure you have it.
- Scale up. At some point, AI needs to be woven into your business strategy, product thinking, customer support, logistics, and marketing. After you achieve initial success using it, you'll need to scale up. This scaling up is across multiple dimensions: infrastructure, data, and business processes. "Ultimately," says Goyal, "you would like to reach a nirvana state where you think daily about how AI can help you in each element of your organization."
Science and art
As businesses figure out how to employ AI, it's important for them to get comfortable with uncertainty. "AI is not just a science; it's an art," says Goyal.
AI, though built on mathematics, is not as tidy as one might think. "It's pretty much asking for trouble to make predictions about AI," says independent data scientist Miranda Mowbray. "The record of such predictions is hilarious, with both extreme over-optimism and extreme over-pessimism exhibited, and with the expert consensus at times being outrageously wrong."
One key challenge is that today's AI applications are extreme data hogs. AI programs typically need an enormous amount of data to learn to do a single task, such as drive down the road, play chess, or even recognize a hot dog.
AI is supposed to help you navigate the data flood, not exacerbate it. For example, for an autonomous car to be "totally safe," says Wu, "you'll need 1 billion miles of driving data, and each mile is 15 gigabytes."
Ultimately, researchers hope that voracious demand for big data insights will be met by artificial general intelligence, or AGI. This is the Holy Grail: an AI that can do more than the one extremely specific thing.
But there's a catch. "If you train a current AI system to do one thing and then another, in addition to being no quicker at learning the second thing, it can get worse at doing the first thing," says Mowbray.
Google is experimenting with "multimodal systems," which have had some luck applying machine learning in one area to problems in another. This is a step toward but still far from AGI.
AI needs massively efficient and powerful compute; increasingly, today's AI workloads run on supercomputers. But due to the impending demise of Moore's Law, today's computing architecture is inadequate to handle the massive processing demands of AI. One solution is HPE's Memory-Driven Computing project, The Machine. Given the data needs, as well as power and memory demands, AI will require an extremely efficient and yet powerful system to run on going forward. The Machine, which makes memory central to the computing architecture and allows users to employ any compute node to work that pool of memory, is one tool that may make AI more practical.
In the meantime, Libratus is helping us move from needing to complete information to successfully reason with an incomplete data set.
At the end of the day, you should neither wait until AI is perfect before examining it for your business needs nor jump forward without considering shortcomings. Like anything else, AI is a tool that can help or hinder a business. Which one it winds up doing for your company depends on your needs, how you implement technology and processes, and your patience.
AI for business: Lessons for leaders
- Embrace discomfort. AI is a game changer, but it requires a measured approach and a certain level of unease.
- Identify where to apply AI techniques to get real benefits—saving employee hours, decreasing response times, etc.
- Plan to experiment. It takes time to get AI right.
This article/content was written by the individual writer identified and does not necessarily reflect the view of Hewlett Packard Enterprise Company.