the brain as a prediction machine — perception is not passive observation but active inference about the causes of sensory signals
the cortex maintains a hierarchical generative model. each layer predicts the activity of the layer below. only prediction errors propagate upward. learning adjusts the model to minimize these errors
the architecture
- top-down: predictions flow down the hierarchy
- bottom-up: prediction errors flow up
- lateral: precision weights modulate which errors matter
the system converges when predictions match observations — free energy is minimized. what remains is the model's best explanation of the world
connection to active inference
predictive coding is the neural implementation of active inference:
- perception: update predictions to reduce sensory errors (change the model)
- action: move to reduce proprioceptive errors (change the world)
- attention: adjust precision to weight errors by confidence (change the gain)
Karl Friston showed these are all gradient descent on the same free energy functional
in cyber
the cybergraph implements a distributed version:
- each neuron predicts local focus distribution based on its model of the graph
- cyberlinks that match predictions (confirm structure) are low-error
- cyberlinks that violate predictions (novel connections) are high-error — and potentially high-reward if they reduce free energy globally
see active inference for the framework. see free energy principle for the theory. see precision for the weighting mechanism