To find out more about the podcast go to Audio Edition: Researchers Uncover Hidden Ingredients Behind AI Creativity.
Below is a short summary and detailed review of this podcast written by FutureFactual:
Diffusion Models: Creativity as a Byproduct of Locality and Equivariance
Overview and the diffusion-model paradox
Image generators such as DALL·E, Imagen and Stable Diffusion are built to copy their training data, yet they often produce images with coherent, novel meaning. Giulio Biroli notes the paradox behind diffusion models: if they worked perfectly, they should memorize, but they don't simply memorize, they create new samples.
"if they worked perfectly, they should just memorize. But they don't." - Giulio Biroli
The Equivariant Local Score and the main results
Kam and Ganguly present the Equivariant Local Score (ELS) machine, a set of equations that predicts the denoised image composition from locality and equivariance. The ELS matches trained diffusion-model outputs with about 90% accuracy across ResNets and Unets. This supports the idea that creativity is a deterministic byproduct of the architecture, not an emergent black-box property.
"as soon as you impose locality, creativity is automatic," - Mason Kam
Locality, equivariance, and a morphogenesis analogy
The narrative connects the model’s locality to morphogenesis, the self-assembly processes in biology that create patterns like Turing structures. Kam, whose work includes morphogenesis, notes that the same local rules that constrain denoising also generate creative outputs, explaining phenomena like AI images with extra fingers.
"The results were shocking across the board," - Surya Ganguly
Implications for AI research and human creativity
Experts argue that the work illuminates the mechanics of diffusion models and may shed light on human creativity, suggesting that creativity could be rooted in partial knowledge and local interactions. Benjamin Hoover emphasizes that human and AI creativity may be more similar than we think, given the way both assemble elements from experience and data.
"human and AI creativity may not be so different," - Benjamin Hoover
The piece situates these insights as a stepping stone toward better understanding diffusion models and the nature of creativity, with potential implications for neuroscience and AI design.
