Hacker News Reader

Comment by chestervonwinch | Hacker News Reader

Do I understand correctly? You have f:R^2 -> R^3, where the input is the image coordinate, the output is an RGB value, and f is a neural net with activations, weights, etc, chosen so that the image (literal and mathematical) of the map looks aesthetically pleasing?

View on HN

Nevermind. Just read the docs of the `get_func` function:

    '''return function of [0,1]^2 -> intensity \in [0, 1]^c'''

... so yes.

The weights are random. No attempt has been made to make sure the generated image looks aesthetically pleasing. That's the surprising bit: a random neural network generates good-looking images.

> That's the surprising bit: a random neural network generates good-looking images.

Despite the randomness in the weights, there's structure and coherence imposed by the activation functions and the manner in which activations are combined (the network architecture). There's nothing inherently magical going on: take a rational function with random coefficients and do domain coloring [1], and you'll get something interesting, too.

[1]: https://en.wikipedia.org/wiki/Domain_coloring

Reply on news.ycombinator.com