I fixed all the Jacobian issues I hope (there was some i,j vs j,i confusion as well as the U vs V typos...)

Now I'm redetecting the optimum loop period at every iteration in case it changes, a little slower but gives smoother loops it seems...

archive.org/details/ntt088 I enjoyed this new release, CC BY-NC-ND

found a typo in my earlier jacobian code, no wonder it didn't work properly

but these are solutions to the fixed point iteration, not necessarily solutions to the reaction-diffusion time crystal problem...

trying simple fixed point iteration now, x <- mix(x, F(x-1`mod`n), g). g = 1/2 in my current test, convergence metric fluctuating around 0.005, want it to be around 1000x smaller to consider it ready....

in earlier tests I ended up with fixed points, scoring as low as 0.0001

Each iteration is faster, but the global convergence seems slower, maybe I made a mistake in the Jacobian calculations.

Perhaps it would be better to try to solve:

x_1 = F(x_n)
x_2 = F(x_1)
x_3 = F(x_2)
...
x_n = F(x_{n-1})

but doing it naively would give 10TB of Jacobian dense matrix data. Needs sparse methods.

The analytic derivatives make it 2-3x faster than the finite differences version, so it was worth the trouble working out the formulas.

I added Jacobian matrix of analytic derivatives for Newton's method solver, but the first time I forgot to take into account the derivatives for the max-min division so it exploded to infinity. After dinner I got it right, seems to be converging a bit better now.

Meanwhile I left the earlier version (with finite difference numerical derivatives) running, the best output it gave was not quite seamlessly looping, and was mostly static anyway. I guess I have to figure out another hack to stop it converging on a fixed point, and instead converge on an interesting cycle...

I'm trying to find repeating patterns in a reaction-diffusion system using GSL's multidimensional root solver.

I figured out a hack to try to stop it converging on a constant featureless image (divide target by max-min of the variables).

It tends to give a starting image outside the 0..1 validity range, which evolves to a featureless image despite my efforts.

So far only failure to report on this one...

I saw a call:

vectorfestival.org Vector Festival (Inter/Access Toronto) 2019 submissions open until 1st February 2019

I saw a call:

ntu.edu.sg/NTUdigitalartprize Nanyang Technological University (Singapore) Global Digital Art Prize "Fourth Industrial Revolution" submissions open until 15th February 2019

. - training neural networks

o - training generative adversarial neural networks

O - training generative adversarial neural networks to recognize their own weights

I dramatically miscalculated the number of weights: the true figure is 2920

or maybe I used a different momentum value, with 0.95 it overshoots badly, with 0.5 I get to 3% error in around 8mins consistently