User:Alexander Roidl/everything
I'm writing a lot
Sorting ideas out
the computational model reproduces complexity of reality
From the database the computer is being trained. While feeding these datasets into huge models we lose sight of it's connections, that make sense for only the computer now on. A model that tries to describe reality in order to generate, to analyze or predict.
I want to put a special focus on generative models and discriminative models of the world.
While these models simplify reality by trying to calculate probabilities and reduce to features, in the same time it makes reality more complicated. These computed models are black boxes. Mostly it is even impossible to see the data it is being trained on.
From features to reduced reality
So an image is being reduced to its most contrasting points, to its pixels that hold a certain array of color values. But what is it that makes an image? If I was to describe an image, I wouln't say: Oh, there is some contrast going on in the left corner, lots of brightness in the middle and
A sentence is being reduced to its words and connections. But can we describe the value of a sentence by only this features?
Training against myself
It becomes even weirder when we look at algorithms that create models by learning from themselves. So we do not only encounter the problem of creating an abstract model, even the database, that otherwise enables us to gather insights on why models act how they act, is incomplete.
How models are built
- model is built by computer scientists
- often incomplete / lacking / biased databases
- selecting features > generalizing
unsemantic everything
Semantic sorting can be filtered trough models > Unsemantic web
Models turn chaos into sense (for a machine)
Bending the model - an experimental approach in understanding
How can we understand machine learning models by their mistakes and missunderstandings
> what can we learn from that different view on reality?
How can we bring models to fail or to its limits, how can you abuse them?