Where computing hits the wall: 3 things holding us back
- By Mark Pomerleau
- Nov 05, 2015
While advancements in big data analytics are becoming increasingly important in helping humans understand vast quantities of data, there are fundamental challenges that could limit future development to only incremental advancements.
Jason Matheny, director of the Intelligence Advanced Research Projects Activity, outlined three areas that need further development before they can make a difference in big data analytics: high-performance computing, machine learning and causality.
The ability to analyze increasingly large datasets is currently restricted by computational limits. The usual way to boost computing power is to string machines together, but beyond a certain point, that solution will not scale. “To power a machine that runs at exaflops, which is a national goal – that’s about a billion billion calculations per second – would currently require a machine the size of a football field that consumes hundreds of megawatts of power,” Matheny told the audience Nov. 4 at a Washington event hosted by Defense One. "So it would need its own mid-sized nuclear reactor just to power the machine."
“In contrast," he added, "the human brain operates somewhere between 10 and 100 petaflops and works off of about 20 watts – so one-one millionth of the energy requirements of our fastest computers.”
One step on the road to exaflop-scale computing, Matheny said, is the recent executive order signed by President Barack Obama that established the National Strategic Computing Initiative -- “a whole-of-government effort” to maximize the benefits of high performance computing.
And even if a computational model capable of achieving exaflop performance is achieved, users still face the problem of winnowing usable information from the datasets that these massive computers analyze, which is where machine learning comes into play.
Given the number of large datasets that have been made available within the last decade, machine learning progress has accelerated rapidly. Furthermore, Matheny said
Examples of the strides made in machine learning include:
- Visual recognition services that have come within two percentage points of human accuracy – facial recognition has reached parity with human judgments in relatively good lighting, as has the recognition technology used in robotics and unmanned aerial systems.
- Automated spam filtering and recommendation systems that often function behind the scenes, hiding their sophistication as well as the intelligence of advertising or promotion targeting technology.
- Robotic controls, such as those used in autonomous cars.
- Natural language processing, machine translations and question answering.
Lastly, Matheny called attention to causality -- which he described as one of the most neglected aspects of big data analytics. Despite achievements in identifying correlations, “we’ve made relatively little progress in understanding how events are caused and affected,” he said. We rarely know how the world would be affected if one choice is made over another. Typically, this is evaluated with human judgment or randomized experiments in the medical field, but you can’t run randomized experiments in geopolitics, he explained.
“A richer statistical understanding of causality is probably the next major obstacle for machine learning to tackle,” Matheny said. Researchers must develop causation technologies that are broadly useful in medicine, chemistry, mathematics and geopolitics. “If we don’t solve that problem," he said, "then all of those exaflop-scale machines and all of those learning algorithms may be of limited value.”
Mark Pomerleau is a former editorial fellow with GCN and Defense Systems.