Karl Friston's AI law: Is it a new step towards General Artificial Intelligence?

Have you ever thought about how our brain works, how we transform stimuli around us into knowledge and information? Neuroscientists have been researching these processes for a long time and, recently, the theory called 'Free Energy Principle' (FEP) by Dr. Karl Friston has been experimentally validated by a group of scientists at the RIKEN research institute in Japan.

The Free Energy Principle postulates that, biological systems, such as the human brain, are constantly trying to minimize the 'free energy' of their environment, their disorder. We can find a very similar term to entropy in deep learning where, like our brain, we try to minimize the entropy of the model so that it gets better results.

Friston suggests that the brain acts as an inference engine that makes predictions about the external world and compares that prediction with the perceived reality. When there are differences between the two, the brain adjusts its internal models to reduce free energy.

Towards a new form of AI: biomimetic AI

The experiment conducted in Japan has demonstrated this behavior in biological neural networks by creating microscale neural cultures grown from rat embryo cells. The use of embryos is based on the need to test the theory in a brain architecture that has not yet been 'trained'.

Electrical patterns mimicking auditory sensations were applied to these neuronal cultures. Initially, the networks began to react randomly, but gradually began to self-organize so that they could respond selectively to one speaker or another.

Doesn't it sound similar to when we train a neural network from scratch? At first the weights of the neurons are random, but as they see more and more data, many of them start to 'freeze' in order to optimize others.

Biomimetic AI is a form of artificial intelligence that mimics the natural functioning of the brain

Researchers were able to demonstrate that this self-organization matched the predictions of computer models based on the Free Energy Principle.

Figure 1. Source.

• a) scheme of the experiment and its corresponding generative model POMPD (Partially Observable Markov Decision Process)
• b) Equivalence between the neural network and Bayesian variational inference. We can observe how the learning process is equivalent, going towards the dark green areas (areas of lower free energy)
• c) Reverse engineering procedure of the generative model.

This discovery is giving way to a possible new form of AI - biomimetic AI. It builds on the brain's natural intelligence by allowing a neural network to self-optimize by continuously receiving real-time temporal data, thus unlocking a critical piece in advancing research towards General AI or superintelligence.

This new approach is also fully programmable, knowable and auditable, allowing it to scale along with human governance.