Turing pattern, AI image creativity and nature of reality

 

In 1952 Alan Turing published a paper “The chemical basis of morphogenesis” that was largely ignored, he was trying to explain mathematical basis of patterns that we see in nature like spots on leopard or bands on snakes so on, this has deep implication on understanding the process by which living systems self assemble, for instance how group of cells organize itself into an organ or a limb. Self assembly is localized wherein individual cells take action or correct itself in response to signals from neighbours. Turing worked out simple mathematical rules in an attempt to create a model to understand basics of pattern formation in nature, the reaction-diffusion equation (range of patterns through activator-inhibitor interactions). And as researchers find out he was fairly successful but there are still complex patterns in nature that is not explained. Unfortunately for us the pettiness of the time he lived in, and much of the world still lives in (the reason why regressive systems and barbaric values must be resisted with all the urgency of humanity), he suicided, denying the world of his genius and therefore astounding possibilities of understanding deeper nature of reality.

Researchers in Stanford who were trying to understand the “black box of creativity” in image generator AI (like DALL-E, Imagen etc.) have found something that is truly amazing. AI is tapping on something far profound than expected. Yes it has Turing’s insight on nature of reality! Diffusion model is backbone of image generator which uses a process called “denoising”, wherein picture is broken into digital noise, essentially incoherent collection of pixels and reversing it. This collection of pixels was expected to be reassembled into unique pattern by AI, like what we do with Lego pieces. But something remarkable happened. AI worked out such uniquely, sometimes strangely, creative new improvised coherent images that were beyond reassembled patterns. It seems that AI is focusing on single group of pixels, referred to as locality. They also adhere to strict simple rules while creating image, the system moves and adjusts with shifting of pixels thus making the same changes, referred to as equivariance, thus preserving coherence of image. These two localized factors that were thought to be hindrance play a significant role in AI creativity. Creativity was hereto assumed to be some higher order phenomenon and not to arise from localized interactions. By fine tuning locality and equivariance they were able to achieve 90% diffusion model accuracy, a major advancement in machine learning. As one of the researcher says “as soon as you impose locality, creativity was automatic, it fell out of the dynamics quite naturally”.     

Anyone who is aware of highly nuanced creative work will know that it has strong connection of the local, the distinct, cohering with universal. Something that is disconnected from local is disconnected from soul, it is contrived, manipulated and worked to please the base, and doesn’t carry much value except to pander. Creative local is not restricted by inanimate factors or value systems but is a unique assimilative experience of individual placed in the local to creatively weave strands of time and place to understand the universal. So is creativity computable? Yes and No. Yes to an extent that it is pattern based in a narrow frame. Can AI take the local in its complex entirety as a human? A complexity that breaks pattern, and is unpredictable, untamed. A mutation is a break from the past to contain the urgency of present. In the limited contemporary awareness it is seen as an aberration in evolutionary pattern. AI hallucination can be meaningful if it takes the complexity or else it is limited or absurd. Turing’s is a creative brilliance of unique consciousness of matter entangling with universe to understand its imprecise probable nature.