Thread by @Icarus98378742

The @acm_ci2021 presentation of the @InferenceActive Collective Intelligence working group is finally up !

Our core argument is that Active Inference, and more specifically the enactive flavour articulated by @mjdramstead @MdKirchhoff @ineshipolito and others, provides an *integrative* and *grounding* framework for the study of collective intelligence from first principles.

This means that Active Inference gives us a intuitive model (and a formal ontology) that can be fitted to basically any intelligent system, while also explaining how exactly a system becomes intelligent.

Let's get into details !

Active Inference agents minimise an informational quantity known as Variational Free Energy.
Due to mathematical black magick, VFE is accessible to the subjective perspective of an agent, but also bounds the future volatility of its sensorimotor states.

A mathematical result known as the Free Energy Principle states that all system with individuated boundary states synchronise with their environment so as to minimise VFE, given that the agent-environment system has an attractor distribution.

This means that basically any thing whose structure is stable enough to constitute a system in its own right does Active Inference, and integrates structural information about their environment through dynamic synchronisation.

This allows us a principled understanding of intelligence in terms of synchronisation, but also tells us what kind of dynamics drive the emergence of intelligence through developmental and evolutionary time.

We can therefore look for signs of VFE minimisation at any scale we see fit : a single neuron, a brain, an organism... or a team of workers.

The big advantage of Active Inference, with regard to preexistant physical theories of intelligence, is that it affords a deflationary account of semantics.

The idea that brains operate near a critical phase transition does explain their wild fluctuations, but not how those fluctuation could retain a structure, and much less have "meaning".

Synchronisation under the FEP, in contrast, entails the integration of information in the system's internal states through developmental and evolutionary time.

Organisms become embodied models of their niche through learning and natural selection, and their phenotype are therefore imprinted with a pragmatic kind of meaning.

As such synchronisation dynamics act simultaneously at multiple nested scales, we should fully expect multi-scale fluctuations such as those displayed by living brains.

ActInf therefore reproduce the prediction of Critical Brain Hypothesis, but additionally explains their origin in terms of cognitive behaviour !

We suggest to developp this intuition through a formal model of the collective brain displaying :
- minimal ActInf agents ;
- self-organising multiscale structure ;
- open-ended evolution of meaning.

We strongly expect such a model would display the wild fluctuations patterns typical of critical phenomenonology...
But also display much better capability at integrating information (by any meaningful structural/functional measure) that a topologically similar critical network.

Thanks for reading, and/or watching the video 🙏

unroll @threadreaderapp