Generative inference: a unifying principle for the integration of learned priors in natural and artificial intelligence
Leveraging learned knowledge to interpret ambiguous or unusual sensory inputs is a hallmark of biological intelligence, yet artificial systems typically cannot utilize their learned priors beyond narrowly defined training experiences. I developed "Generative Inference", a computational principle that enables any artificial neural network to access and deploy its implicitly learned statistical regularities through feedback pathways. Applied to vision, we demonstrate that Generative Inference accounts for both perceptual experiences and neural signatures across a spectrum of previously disparate phenomena—from illusory contours to figure-ground segregation, brightness illusions, and even hallucination-like pattern formation—all within systems capable of rapid object recognition.
Project Page
Interactive demonstrations and detailed project information
Related Publications
Illusions as features: the generative side of recognition
Workshop on Scientific Methods for Understanding Deep Learning, Advances in Neural Information Processing Systems (NeurIPS) (2024)
Generative inference in object recognition models—A unifying framework for discriminative and generative computations in vision
From Neuroscience to Artificially Intelligent Systems (2024)
Generative perceptual inference in deep neural network models of object recognition induces illusory contours and shapes
Cognitive Computational Neuroscience (CCN) (2024)
Invited Talks
Generative inference in object recognition models—A unifying framework for discriminative and generative computations in vision
TigerBrain Research Symposium, Princeton University (2024)
Generative inference in object recognition models—A unifying framework for discriminative and generative computations in vision
From Neuroscience to Artificially Intelligent Systems, Cold Spring Harbor Laboratory (2024)
Generative perceptual inference in deep neural network models of object recognition induces illusory shapes
Swartz Foundation Meeting, University of Washington (2024)
Emergence of illusory contours in robust deep neural networks by accumulation of implicit priors
Object Recognition: Models, Vision Science Society Meeting, St. Pete Beach Florida (2024)
Cortical computations underlying the integration of perceptual priors and sensory processing
Brain Science External Postdoc Seminar Series, Brown University (2024)
Harnessing feedback pathways: Integrating perceptual priors in sensory processing
SYNAPSES Seminar Series, Yale University (2024)
Public Demos
"How do I see myself if I develop dementia?"
Presented at Vision Science Society, Demo night, 2024