[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Spooking of neural nets and image recognition...




>4) placing a magnet on the side of the camera (does this work?)

Not with CCD cameras.


All the other methods you mentioned require special physical access--
it won't work if one gets photographed while placing post-its on the
camera window...

The high error rates of image recognition make this whole scenario
a future issue.  One FAA funded experiment used mice to detect 
excessive adeneline (the mice go nuts or their heart rates increase
just by smelling the excitement); the idea was to catch hijackers
who would generally be a bit excited.  This sounds obvious boneheaded
because of all the people who fear flying...but the stated reason
for abandoning the research was that mice don't rate too well as
anonymous tipsters!  Anyway, the error rate was very high too.

However, the whole "profile" thing used by the WoD is essentially a
conscious application of generalization that neural networks do.  This
may expand if "suspicion detection" is socially acceptable--my guess
is that it would not be accepted given the speeding ticket automation
systems that have been widely rejected (they probably could have gotten
it accepted if the reduced a speeding ticket cost by a magnitude and 
considered it like a parking ticket--but these are legal changes, not
technological changes, so they are much more difficult to do).

The OCR of cash serial numbers would be highly probabilistic--that is,
too many transactions would not be tracked so the knowledge of the flow
would be partial.  More likely would be that all large cash deposits
would be scanned for general analysis just as large cash transactions 
require that a bank fill out a special form and send to the State.  


Paul E. Baclace
[email protected]