Why Machine Learning needs HPC
By   |  November 10, 2015

Steve Conway, IDC Research Vice President, HPC

It seems to me that a reasonable definition of technology is artifice that augments our native human abilities. With it, we can lift better (lever), move better (vehicles), see better (microscope, telescope), hear better (radio, hearing aids), feel more sensitively (haptics), and more. I’m not aware of much progress in improving our interdependent senses of taste and smell.

Calculators and computers extended this premise from our physical beings to our brains, helping us to think better, at first by relieving us of mathematical grunt work. But the dream of true thinking machines, devices that can be taught to reason and act on their own, has long loomed in the human psyche. The dream dates back at least to ancient Greece and phenomena such as the tripods that the god Hephaestus purportedly crafted to walk on their own to Mount Olympus and back: early driverless vehicles.

Not surprisingly, in each historic era, the proposed mechanism for the machine «brain» closely mirrored the presumed mechanism for the human brain — from divinely created, to mechanical in the Machine Age, to electro-chemical-digital in our time (cf., the sub-headline of a 2013 published paper: «physicists have developed a technique that can tell which parts of the brain rely on analog signals and which rely on digital signals»). Once created, none of these powerful metaphors entirely disappears from our beliefs. Our imaginations today live with an amalgam of prior and contemporary brain metaphors.

As computer capabilities have advanced, these machines — notably supercomputers — have been able to beat the best humans at increasingly complex games, from checkers to chess to television’s Jeopardy! quiz show. (In the context of this discussion, we loosely define a game as a non-lethal activity associated with a goal and rules for reaching it.) One of the next major «games» will be driverless vehicles. In the short run, they may not outperform the best human drivers, but they will certainly outperform the worst, in part because driverless vehicles don’t get drunk or distracted by tweets. Another important game will be personalized (precision) medicine, where computers are already starting to be used for decision support for difficult, near-real time diagnoses and treatment plans.

If it seems odd to use the term «game» for such serious pursuits, consider the source. Human thinking is heavily goal-oriented, teleological. Maybe that’s in our DNA. If we hadn’t learned the game and the rules, Darwin says, we would not have survived as a species. It’s not surprising that we have created our protothinking machines in our own image, to follow the same fundamental behavior pattern.

Today, the field of machine learning can be defined within the framework of game-playing.

At one end of the spectrum is mainstream machine learning, where the computer is given the goal and then painstakingly trained to play the game reasonably well. This falls under the general headings of supervised learning and its less explicit relative, reinforcement learning.

At the other end today is unsupervised learning, in which the computer is presented with data and must ferret out the goal and the rules on its own.

The best of contemporary unsupervised learning is often called cognitive computing, because it’s thought to mimic the cognitive processes of the human brain. Mind you, we humans don’t exactly know yet how human cognition works. Figuring out how we acquire and use knowledge is a question that continues to tantalize psychologists, philosophers and brain researchers. Presumably, we understand more about machine cognition, since we are the ones who’ve been teaching the machines how to think.

The gap between human and machine brains will continue to narrow. Farther, probably much farther, in the future there will be artificial intelligence (AI) that passes the Turing test: an evaluator of a text exchange can’t tell which participant in the conversation is the human and which is the machine. A more cynical perspective sees AI as a capability that will always lie in the future, because threatened humans will always move the goal post when a machine reaches the one-yard line.

Sci-fi visions to the contrary, we humans needn’t worry about machine domination any time soon. It’s trite but true that human thinking is far more intricate and nuanced than any realistic roadmap for machine cognition. No one seriously expects a thinking machine to replicate the fullness of Shakespeare, Einstein or Louis Armstrong. At most, we will get Rain Man, a machine (or collection of machines) with exceptional niche intelligence.

This intelligence likely will be geographically distributed, via the evolving Internet of Things (IoT), with concentrations of networked brainpower varying by location as needed. It may be akin to the view from an airplane traversing a nighttime landscape: bright fusions of light in large cities, more moderate brilliance in smaller towns, and isolated points of light in the countryside, with illuminated strands of roads and highways forming the interconnect fabric.

China, for one, is counting on an HPC architecture called REST 2.0 to manage the country’s future IoT. As Dr. Zhiwei Wu, Chinese Academy of Sciences (CAS), outlined at the ISC’13 Big Data conference in Heidelberg, CAS is developing a new server they hope will scale to 1 billion threads, with a dramatically simplified architecture and hardware/software stack — along with storage and an «elastic processor» with a functional instruction set architecture (FISC) design. Closer to home, international ecommerce giant PayPal has been investigating the use of HPC to manage a global graph infrastructure that will embody increasing intelligence about the behaviors of retail shoppers.

Whether the objects of interest are human shoppers, driverless vehicles in traffic or smart electrical grids, it’s clear that the IoT will be far more than just an «Internet of stupid things.» HPC will be needed to move massive data, to process it as quickly and as locally as feasible, and to help maintain the wellness of this network with its distributed intelligence.

One of HPC’s main contributions to the future of machine learning, including its use in the nascent IoT, will be the ability to exploit parallelism. As recent IDC studies have confirmed, few algorithms used by the mainstream machine learning industry have been parallelized. That community got its start in about the year 2000, in an era when generational jumps in the single-threaded performance of x86 processors often improved performance without the need to parallelize. Failure to parallelize algorithms can have important consequences, such as limiting applications to «good enough» speed and resolution, or forcing users to run problems at sub-optimal sizes or, in certain sectors, increasing reliance on costlier physical experimentation.

But if history is any guide, the mainstream market for machine learning will soon enough evolve to the point where good enough is no longer good enough to compete and survive. This is where the HPC community will come to the rescue.

© HPC Today 2019 - All rights reserved.

Thank you for reading HPC Today.

Express poll

Do you use multi-screen
visualization technologies?

Industry news

Brands / Products index