[-empyre-] Welcome to Week 3 on -empyre-: Internet/Online Representation

[B. Bogart]

Hello all,

I've been lurking around this week, but inspired to jump in now. I
wanted to highlight this comment from Abram in terms of the relevance to
a thread in a previous month:

On 15-04-18 07:58 PM, abram stern (aphid) wrote:
> in speaking with the artists, I learned that it records gender on a 30
> point scale, -15 being the most masculine and +15 being the most
> feminine.  Is this non-binary gender coding (by the API designers [or
> perhaps by whatever researchers they based the API on], not the artists)
> an assertion of progressivity or addressing a technical limitation of a
> system that can only guess its subject based on similarities between two
> defined poles produced by compositing lots of 'male' and 'female' faces? 

Does this -15 to +15 scale imply a gender continuum, or is it the
algorithms sense of confidence? i.e. -15 is a 'male' with 100%
confidence, while 0 could be 'male' with 50% confidence?

Even if there is this illusion of continuity based on confidence, the
algorithm may internally only classify in two groups, just with enough
inaccuracy to appear like a continuum.

There is also the possibility that this is not actually a gender
detector, but a female detector where 'maleness' is simply the inverse
of femaleness. A female is a person detected as male to a low degree of
confidence. This plays interestingly in the notion of maleness as gender
neutral, and all faces assumed to be male unless they lack the male
features.

I wonder what degree these gender detection systems reinforce or
problematize gender norms. Do they help viewers reflect on the gender
binary?

What biases are hidden in cases where the actual algorithm is veiled
(due to the artists' and viewers' inability to access and/or understand
it). Even more interesting is the training set used allow the algorithm
to detect gender, usually a pool of 'male' faces and a pool of 'female'
faces. How diverse are these pools of faces? How are these people
selected? Who determined they are male or female?

In some cases the algorithms are distributed freely, but these training
sets tightly controlled.

Ben Bogart, Ph.D.
wwww.ekran.org