Big AGI Breakthrough! From Active Inference to Renormalising Generative Models
Friston et al. (2024) recently introduced "renormalising generative models" (RGMs) as a means for scaling active inference across both time and space, with applications to image recognition, image/video/audio compression and generation, as well as to planning. We show how the evolution from active inference to RGMs (2010 - 2024) parallels the evolution of energy-based neural networks from simple (restricted) Boltzmann machines to deep learning (1983 - 2012), through the common theme of renormalization group theory. Action Perception Divergence (APD) still has a role in AGI, and we also identify some key comparisons between RGMs and JEPA (Joint Embedding Predictive Architectures).
The Blogpost for this Vid is at:
https://themesis.com/2024/08/19/big-agi-breakthrough-leveling-the-playing-field/
The Themesis three-week short course, "Top Ten Terms in Statistical Mechanics," is designed to give students a gentle-yet-intensive immersion into sufficient statistical mechanics vocabulary so that the student can go on to read papers in generative AI and AGI that were previously inaccessible.
Here's the link to the Themesis Academy, where you can traverse a link to the actual course offerings on themesis.thinkific.com:
https://themesis.com/academy/
Prior to this, we introduced Action Perception Divergence in THIS YouTube:
https://www.youtube.com/watch?https://www.youtube.com/watch?v=436l0cmnEi0
We then did the derivation for Eqn. 3 of the Action Perception Divergence in THIS YouTube:
https://www.youtube.com/watch?v=ae_-qumFxrs
Please OPT-IN with Themesis on the About page to get word AS SOON as new YouTubes, blogs, and short courses are released:
Opt-In HERE: www.themesis.com/themesis/
Subscribe to the Themesis YouTube channel easily - click this link: https://www.youtube.com/@themesisinc.4045?sub_confirmation=1
Friston et al. (2024) recently introduced “renormalising generative models” (RGMs) as a means for scaling active inference across both time and space, with applications to image recognition, image/video/audio compression and generation, as well as to planning. We show how the evolution from active inference to RGMs (2010 – 2024) parallels the evolution of energy-based neural networks from simple (restricted) Boltzmann machines to deep learning (1983 – 2012), through the common theme of renormalization group theory. Action Perception Divergence (APD) still has a role in AGI, and we also identify some key comparisons between RGMs and JEPA (Joint Embedding Predictive Architectures).
The Blogpost for this Vid is at:
https://themesis.com/2024/08/19/big-agi-breakthrough-leveling-the-playing-field/
The Themesis three-week short course, “Top Ten Terms in Statistical Mechanics,” is designed to give students a gentle-yet-intensive immersion into sufficient statistical mechanics vocabulary so that the student can go on to read papers in generative AI and AGI that were previously inaccessible.
Here’s the link to the Themesis Academy, where you can traverse a link to the actual course offerings on themesis.thinkific.com:
https://themesis.com/academy/
Prior to this, we introduced Action Perception Divergence in THIS YouTube:
https://www.youtube.com/watch?https://www.youtube.com/watch?v=436l0cmnEi0
We then did the derivation for Eqn. 3 of the Action Perception Divergence in THIS YouTube:
https://www.youtube.com/watch?v=ae_-qumFxrs
Please OPT-IN with Themesis on the About page to get word AS SOON as new YouTubes, blogs, and short courses are released:
Opt-In HERE: www.themesis.com/themesis/
Subscribe to the Themesis YouTube channel easily – click this link: https://www.youtube.com/@themesisinc.4045?sub_confirmation=1
