Active Inference — Predictive Coding with Gabor Edge Dictionary and Interactive Controls
a browser‑based simulation of active inference and predictive coding in a 2‑D visual environment. The model maintains separate latent vectors for low‑resolution shapes, high‑resolution shapes, and oriented edge features; the edge dictionary consists of true Gabor filters with configurable orientation count and carrier frequency. The free‑energy cost includes sensory prediction error, L2 and temporal regularisers for each latent group, and a coupling term tying high‑ and low‑resolution shape representations. Inference updates all latent variables via gradient descent on the free energy, while the action policy samples pixels from a foveated view to minimize expected uncertainty.
The interface exposes sliders for sensor noise, precision weights, learning rate, number of basis functions in each dictionary, L2 and dynamic penalties, and the Gabor dictionary orientation count and frequency. It also allows toggling motion, occlusion and foveation, switching among various active‑sensing policies (random, novelty, uncertainty, residual, hybrid), and editing the environment by adding, removing or moving blobs, bars, triangles and 5‑point stars. Multiple live displays show the predicted image
𝑔(𝜇)
g(μ), the ground truth, a novelty/coverage map used for action selection, and a running free‑energy trace.