DARPA AI will trawl petabytes of UAV vid for enemy cows

[Eugen Leitl]

http://www.theregister.co.uk/2009/04/15/darpa_deep_learning/print.html 

DARPA AI will trawl petabytes of UAV vid for enemy cows

Will see horses, deduce existence of evil cattle

By Lewis Page

Posted in Science, 15th April 2009 15:13 GMT

Renowned Pentagon tech-tomfoolery agency DARPA has announced a new plan to
create mighty artificial intelligences. The so-called "Deep Learning"
machines will be used to trawl through petabytes of video from robot aircraft
prowling the skies - initially, apparently, seeking out threatening horses
and cows.

According to DARPA boffinry chiefs, setting out the rationale for "Deep
Learning" technology, the US military and spook communities are hip-deep in
surveillance and intel data, and sinking fast. Hence the need for artificial
intelligence (ha ha):

    A rapidly increasing volume of intelligence, surveillance, and
reconnaissance (ISR) information is available to the Department of Defense
(DOD) as a result of the increasing numbers, sophistication, and resolution
of ISR resources and capabilities. The amount of video data produced annually
by Unmanned Aerial Vehicles (UAVs) alone is in the petabyte range, and
growing rapidly. Full exploitation of this information is a major challenge.
Human observation and analysis of ISR assets is essential, but the training
of humans is both expensive and time-consuming. Human performance also varies
due to individualsb capabilities and training, fatigue, boredom, and human
attentional capacity.

    One response to this situation is to employ machines ...

It seems there are already plenty of basic "shallow learning" AIs in use,
including such Stone Age expedients as "Support Vector Machines (SVMs),
two-layer Neural Networks (NNs), and Hidden Markov Models (HMMs)". But these
are scarcely better than a human with poor "attentional capacity"*. The
trouble with the shallow learners is that they can learn, erm, only at a
shallow level:

    Shallow methods may be effective in creating simple internal
representations ... A classification task such as recognizing a horse in an
image will use these simple representations in many different configurations
to recognize horses in various poses, orientations and sizes. Such a task
requires large amounts of labelled images of horses and non-horses. This
means that if the task were to change to recognizing cows, one would have to
start nearly from scratch with a new, large set of labelled data.

In essence, a specialised horse-spotter machine unable to recognise a cow
isn't much use for sorting the sheep from the goats. (We're plainly in the
War On Livestock here.) That's why DARPA want "deeply layered" learning
machines, able to apply horse sense to recognising cows, sheep and goats.

    Deeply layered methods should create richer representations that may
include furry, four-legged mammals at higher levels, resulting in a head
start for learning cows and thereby requiring much less labelled data when
compared to a shallow method. A Deep Learning system exposed to unlabelled
natural images will automatically create high-level concepts of four-legged
mammals on its own, even without labels.

The existence of horses logically implies that there must be other furry
four-legged mammals. And horseshit, of course.

Basically, then, the Deep Learning machine, seeing horses, would develop the
concept of sheep, goats and cows. In fact it will develop a lot of concepts,
according to DARPA:

    Before you have a label, you must have a concept to label.The Deep
Learning program assumes the ability to learn from unlabelled data in the
first phase of the envisioned program.

So the machine will be given access to the multi-petabyte robo-vid
surveillance archives of the Pentagon, and start developing unknown concepts
on its own using that data as a base. Here are a couple of samples, courtesy
of Multi National Forces Iraq HQ:

Perhaps the computer will deduce the existence of hostile two-legged mammals,
which must and will be exterminated using smart weapons. Or it may fall back
on early programming and start a savage global war against all cloven-footed
animals.

Pleasingly, there will be chances to see what the technology does with
different data sets - provided we aren't all wiped out at once - as DARPA
says it will release all its kit for others to play with.

    A further end objective of the Deep Learning program is to support
increased growth and development of the broader machine-learning community by
making publicly available many, if not all, Deep Learning software modules,
algorithmic approaches, evaluation criteria and datasets in several
application domains for use by researchers.

There's more from DARPA in pdf here
(https://www.fbo.gov/download/f45/f451dac7a3cbcd699ca5f4cdcba862a9/Deep_Learning_BAA_Apr_10.pdf).
B.

* Aiee!