From pixels to planning: scale-free active inference.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Generative modeling is crucial for understanding and predicting dynamic systems.
• Partially observed Markov decision processes (POMDPs) are a standard framework but can be limited in modeling complex temporal dependencies.
• Existing models often struggle with compositionality across space and time.

Purpose of the Study:

• To introduce a novel discrete state-space model for generative modeling, generalizing POMDPs.
• To develop renormalizing generative models (RGMs) capable of learning compositionality over space and time.
• To demonstrate the versatility of RGMs through diverse applications in learning and inference.

Main Methods:

• Developed a discrete state-space model that incorporates paths as latent variables.
• Utilized deep or hierarchical forms inspired by the renormalization group.
• Applied variational principles for automatic discovery, learning, and deployment of RGMs.

Main Results:

• RGMs are shown to be discrete homologs of deep convolutional neural networks and continuous state-space models.
• The models exhibit scale-invariance and learn compositionality over space and time, modeling paths and orbits.
• Successful applications demonstrated in image classification, movie and music generation, and learning Atari-like games.

Conclusions:

• Renormalizing generative models offer a powerful new framework for generative modeling in dynamic settings.
• RGMs provide a principled way to learn hierarchical, compositional representations from data.
• The demonstrated applications highlight the broad applicability and effectiveness of RGMs across various domains.