I also wonder if just one skillfully performed twitch with the left leg
could trip a gait detection algorithm? There are many holes to poke
into. Having access to the interpreting system, as those researchers
did, makes it obviously much easier to find the right "markers" to
tweak. But considering that economies of scale will most likely give to
rise to a few default classification networks, accessible for $£€ over
an API, some of there inner workings might be discovered over time.
Isn't the prying open of a black box peoples favorite pastime?
Regarding the rise of the "AI". Totally agree, it "became" something
like climate change. An inevitable wicked problem, of which the
involved's right hand demands careful consideration of the consequences
while the rest of the body is pushing for its implementation at full
speed. I very much like to stress that at the moment it is just machine
intelligence, not sentient, or as Zuckerberg said: it's just math.
Nevertheless, it remains a very powerful tool, and it is in the hands of
a very few (and their software engineer/programmer management layer).
On 01/11/2017 21:33, Morlock Elloi wrote:
And this just in:
https://arxiv.org/pdf/1707.07397
We introduce the first method for constructing real-world 3D objects
that consis-
tently fool a neural network across a wide distribution of angles and
viewpoints.
We present a general-purpose algorithm for generating adversarial
examples that
are robust across any chosen distribution of transformations.
Video of a rather impressive demo (turtle gets classified as a rifle) at:
https://www.labsix.org/media/2017/10/31/video.mp4
https://www.labsix.org/physical-objects-that-fool-neural-nets/
The point of all these attacks appears to be that "AI" is just plain
old primitive classifiers, rebranded by the marketing, all extremely
brittle, working under naive assumptions (but good enough for demos
and PR.) "AI" sounds more scary and induces defeatism, resignation,
and deference to technology, which is its sole purpose.
..
# distributed via <nettime>: no commercial use without permission
# <nettime> is a moderated mailing list for net criticism,
# collaborative text filtering and cultural politics of the nets
# more info: http://mx.kein.org/mailman/listinfo/nettime-l
# archive: http://www.nettime.org contact: nett...@kein.org
# @nettime_bot tweets mail w/ sender unless #ANON is in Subject:
--
DE: +49 (0)160 9549 5269
UK: +44 (0)75 0655 0520
http://vincentvanuffelen.com
http://transmit-interfere.com
http://deepmediaresearch.org
# distributed via <nettime>: no commercial use without permission
# <nettime> is a moderated mailing list for net criticism,
# collaborative text filtering and cultural politics of the nets
# more info: http://mx.kein.org/mailman/listinfo/nettime-l
# archive: http://www.nettime.org contact: nett...@kein.org
# @nettime_bot tweets mail w/ sender unless #ANON is in Subject: