architecture flexibility are vital if the electronics industry is to deliver
useful, accessible, and affordable computer-vision applications in the coming
years, according to a panel at 2014 DAC.
The June 5 panel, titled
"Hardware-Software Codesign for Computer Vision: Can We Make a Computer See?", saw
panelists spar over certain elements of implementation, but they generally
agreed that issues with transistor and power scaling and software
implementation demand new approaches.
"How do we
cope with the computational intensity challenge and the creative intensity of
the space?" Cadence Fellow Chris Rowen (pictured, right) asked during the June 5 afternoon
session. "That's going to influence everything that lands in silicon."
is a space he dubbed the "Wild West" because of its lack of standards, and
panelists agreed that a variety of approaches is required since the
applications will vary widely.
Computer vision learning
Professor Andrew Ng described
the challenge by relating the Google Brain project on
which he worked several years ago. The team used 16,000 CPUs to train the
system to identify faces by watching YouTube videos, a process called
supervised data analysis. The system eventually identified images of cats.
thing about this was it had discovered the concept of a cat by itself," Ng
noted. "These large deep learning algorithms are driving substantial economic
Ng said the
"hunger for bigger systems continues" as engineers tackle larger
computer-vision problems. But he cautioned against pursuing computing solutions
for the sake of computing, even where it might be put up against thornier
challenges like unsupervised data analysis (in other words not training a
system at first to look for certain clues to piece together a larger image of,
say, a human face or a cat).
Ng, who is also
co-chief scientist at Baidu, said:
"Hardware groups are building systems that
can simulate a trillion connections. Those are good supercomputing results but
the relevance to Baidu, Google, Facebook, Microsoft, or Apple is non-existent."
vision, Ng sees a major shift in the coming years away from supervised learning
approaches to learning from untagged data (unsupervised), and this will require
different approached to electronics systems design.
"The shift to
unsupervised algorithms has discovered the concept of a cat," he said. "Hardware
will need to flexible and programmable because we really don't know what the
algorithm needs to be. As a society, we have access to more unlabeled data than
Panel moderator Yankin Tanurhan of Synopsys, however, expressed some skepticism, noting that RAF analysts in World War II identified V-1 rockets from grainy pictures without, at least at first, knowing what a V-1 was. Would a computer vision system be as successful in the same situation? "I have my doubts about this so-called learning effect," he said.
Michael B. Taylor, a professor
with the University of San Diego's computer science and engineering department,
said another challenge is the scaling of transistors and energy efficiency—the problem
of so-called dark silicon as a function of the breakdown of Dennard scaling.
"We have exponentially
more transistors (but) we can't switch them and can't use them for computation,"
Taylor said. "But we can use them for memory."
"There's a pretty
interesting result that's coming along, which is that we're getting to the point
where we're not going to have to store all of our video off chip," Taylor said,
noting that those otherwise "dark silicon" transistors used as memory instead
of logic will enable that. "We're actually going to be able to fit a lot of it
on chip. And that's going to really help us with efficiency."
pixel-processing costs and early system success in the marketplace have created
big expectations, and this puts all the more pressure on engineering teams,
lack for problems to solve. What we lack are the combination of architectures
which have the kind of efficiency that allow them to be deployed in mass
quantities. It's all good to say I have 1000 servers and 16,000 processors,
but I don't want to do that on my wristwatch."
He urged the
audience to think about the problem along three axes:
the arc of the computation nodes? Is it hardwired data path? A general-purpose
processor? Something in-between?
How do we
interconnect blocks with imaging and vision DSPs and how do they all talk to
algorithms and programming models should be used to get things done. Open VX,
he noted, is a "big step" in the right direction.
EDA to the rescue?
Since it was an
EDA panel, the question of what should be design automation's role arose.
Jason Clemons, research
scientist with NVidia, said:
"We need a way to explore the design space. Give
us utilities to allow us to play around with basic components and performance
metrics that allow us to evaluated beyond ... area and power."
Intel Fellow Doug
Carmean urged the audience to think broadly about solving computer vision
problems because it's not so much about computer vision, per se, as it is about
"One of the things you've heard as a theme
is people wondering ‘are there fixed-function units or ... general-purpose
units?' It's not an ‘or;' it's an ‘and.' The chips of the future that we will
be designing will have general-purpose functionality, will have special-purpose
functionality... they'll have filters, DSPs, they'll be programmable, they'll
be configurable. That leads us to computers that can actually understand."
- DAC 2014 Keynote: EDA Can Tap Into New Revenue
- DAC 2014 Dual Keynote: How Automobiles Are Getting
- Embedded Vision Summit: Focus on Autonomy and Recognition