[OSM-dev] Loc2Vec

[Christoph Hormann]

On Monday 18 June 2018, Mateusz Konieczny wrote:
> > One interesting thing about this kind of work is the legal side and
> > the question if a process as described there is a derivative work
> > of the training data used.
>
> I am surprised that somebody may think that it is not a derivative
> work of training dataset.

Given how rarely you see attribution for data sources in the 'big data' 
sector and how widespread attribution requirements are in the open data 
world OTOH it is pretty likely that use as training data in adaptive 
algorithms is widely not considered to be copyright relevant.  If this 
view is correct or not is a whole other story.

The topic is also further complicated because if there is a relevance of 
copyright/database rights here there are still two options:

* The trained algorithm is considered to be a derivative of the training 
data but the output of applying the trained algorithm is not.
* The output of the trained algorithm is considered to be a derivative 
of the training data as well.

I think it is probably possible to create scenarios where all three 
viewpoints (the two above and the 'no copyright issues involved' case) 
are plausible interpretations.  So there are no simple yes/no answer to 
the whole subject - you would need to draw a line somewhere - 
preferably one that, as i phrased it in legal-talk recently, "supports 
desirable and harmless use cases but does not create loopholes against 
the spirit of the license."

-- 
Christoph Hormann
http://www.imagico.de/