Reminds me a lot of the wavefunction collapse procedural generation technique. Of course that algorithm isn't entirely local like this, and it doesn't require training a model either. This begs the question, how big and how complex can you make the images before the model starts failing to reproduce them with low error? If you increase the hidden state size and model size, how does that change things? Could you use this as a way to measure image complexity in a more nuance way than resolution?
True. It would be interesting to analyze how much of the overall structure is encoded in each cell's hidden state. Maybe this whole thing is just trivial and all global information is hidden in there.
Crucial insight buried in the video here about aging (not a new idea, but nice to see it implied here)... how the update rules' result seems to go awry after the number of steps it has been trained for.... maybe this is why our bodies start breaking down and going wonky after we live past the typical reproductive period of our ancestors! Maybe part of aging is that our morphogenic+morphostatic system has not been trained/evolved to maintain our structure past a certain point. The methods they used to overcome this might inspire distant future attempts at combating or preventing aging. Really really excellent video...made it simple for dumb people like me to understand and nerd out on, thanks!
That is a very interesting thought! A few months ago Veritasium made a video about some of the new insights on body aging: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-QRt7LjqJ45k.html
Very interesting research. This kind of model definitely gets close to what must be going on in a body growing out of a lump of cells. Clearly, there cannot be much reliable communication (with chamicals or otherwise) at great distances, especially not before the blood and nervous systems are in place, so the body-shaping decisions must be made locally. However, each cell has the same complex program so "The network parametrizing this update rule consists of approximately 8,000 parameters. " does not sound excessive. "Typical cellular automata update all cells simultaneously. This implies the existence of a global clock, synchronizing all cells. Relying on global synchronisation is not something one expects from a self-organising system. " is also another important consideration.
Really inspired that they backpropagate through time instead of using evolutionary methods like most "artificial embryology", that was a suitable update rule could be discovered much more quickly and perhaps more accurate rules could be discovered.
It is amazing, god said "To build an eye", then the eye comes out. The only problem is when organs combine together, the cell on the border can not make clear which organ should it belongs to.
Although the website is non functional now as the original patterns were replaced with black dots and it is NOT configured to grow from that, it’s only configured to have a preexisting pattern, I know this is an old video, but can someone fix that?
Decent explanation of the article. One point of confusion for me though - when you mentioned the 'residual connection' being the key to making the whole thing work ; did you mean that the fact that they output a delta pixel and add that to the input pixel to get value of that pixel at the next time step?
Sort of, since it computes deltas rather than the new value itself, there's an "identity" connection from a timestep to its previous one, which could let gradients pass nearly undisturbed/without being shrunk/exploded due to matrix multiplications & activation functions
Perhaps make it a 1D cellular automata? With amplitude in place of the RGB channels. Do you wanna use it to recover sounds? Because it seems that this algorithm will only output a specific pattern it's been trained on. I wonder if a GAN could generalize better
Is cnn the reason why people are so hooked on watching the news, which loops? I think it creates anxiety and a false sense of reality in people. Just a "dumb" question.