🚀 go-pugleaf

[Followups to comp.ai.neural-nets please.  -- Rick]

rick@cs.arizona.edu (Rick Schlichting) writes:

> Prof. S. Tamura at Osaka University Medical School, and Asst. Prof.  H.
> Kawai ... have demonstrated that the back propagation neural network can
> identify with high accuracy the sex of persons from their faces.

That is a significant claim.
 
> The test indicated that people have an identifying ability poorer than
> the neural network [sic].  How the neural network differentiates between
> the sexes has to be elucidated.  The neural network may extract facial
> features that people cannot detect.

The fact that a vanilla bp performs better than humans on a complex task 
at which humans normally excel makes it probable that the nn's success is 
based on 'artefact' information. Such artefacts could be easily intro-
duced into the training set by e.g. taking the photos of the men & women
in separate sessions with e.g. different background lighting. A very 
simple averaging calculation could then be performed by the nn to 
distinguish between 'male' (== background A) & 'female' (== background B).
 
> ... it answered 93% correctly even for given
> 8x8-array mosaic images, which were not recognizable as human faces to
> people.

That is a picture considerably smaller than a workstation icon, and the 
latter already require lots of clever pixel editing to resemble something
like a mailbox (even if graylevels are allowed). I find it hard to believe 
that the sex of a face can be reliably inferred from a 8x8 picture _at all_.
This performance claim makes an 'artefact'-based explanation very likely.

Cheers, henrik@robots.ox.ac.uk


PS: The Kahaner reports are typically very useful and worth reading, I just
    had to comment on this one. 
PPS:If bp really does the job, send me the code - they might give me a PhD
    for it here :-) !