[Imports] Facebook's AI-Assisted Road Tracing for OSM

[James]

Having the complete data set for analysis would be a must, as all imports
go through this to validate the quality of the data.

What I'm saying about the processing part: you feed it a bunch of drawing
of dirt roads/paved roads. It will start to see what is similar and what is
different (kind of like using KDIFF to analyse differences between texts,
but a little bit more complex). It will see patterns i.e. the colour green
is usually on both sides of this brown that is considered a road(over
simplification)

Neural Networks work much like our brains: look for patterns that we can
recognize. How do you identify a road from imagery Christoph?
Subconsciously there is a lot going on such as: width of feature, colour
similairity, resemblance to past examples of what you determined as a road,
is there a clear path/direction that this width is headed off into(long and
slim?),etc etc Neural networks work in the same fashion: find patterns from
what it has "learned" are roads.

Bye.

On Thu, Mar 16, 2017 at 7:36 AM, Christoph Hormann <osm at imagico.de> wrote:

> On Thursday 16 March 2017, James wrote:
> >
> > Seeing as facebooks computational power will rival that of super
> > computers, even if you wanted to recreate it, it would probably take
> > you 1000 years of home computing power to "train" the neural network
> > to recognise roads, at which point we will all be dead.
>
> That is not how this works although i understand how PR of Facebook and
> others makes people believe that.
>
> You can perfectly use neural networks and other 'artificial
> intelligence' techniques on hardware accessible to normal people.  The
> difficulty with getting useful results with such techniques is not the
> computing power required (this hardly ever is the case for anything
> these days), it lies in adjusting such a system for the task at hand
> and training it to produce useful results.  In contrast to an
> intelligent human who validates his/her conclusions with a large amount
> of life experience an artificial intelligence will happily produce any
> kind of nonsense if it is set up to do so - either voluntarily or
> accidently because of lack of expertise in using these techniques.
>
> Which is exactly why we need to evaluate the results in full before an
> import in OSM and not just a small sample area.
>
> This is also a kind of chicken-and-egg problem - people use 'artificial
> intelligence' techniques because they don't want to bother with
> analyzing and understanding the underlying mechanisms of a problem they
> are trying to solve so they say "let's just have the AI figure it out".
> However to really properly set up and train the AI they actually need a
> thorough understanding of the underlying mechanisms.
>
> If Facebook does not want to open their methodology for whatever reason
> that is perfectly fine but if they want to import their results in OSM
> they need to either open their process or open the data as a whole for
> us to evaluate before the import - as i already explained in my initial
> reply.
>
> --
> Christoph Hormann
> http://www.imagico.de/
>
> _______________________________________________
> Imports mailing list
> Imports at openstreetmap.org
> https://lists.openstreetmap.org/listinfo/imports
>



-- 
外に遊びに行こう！
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/imports/attachments/20170316/fcea274a/attachment.html>