13 Apr
2022

Why we need active inference?

Research

Our scientific challenge

“Haven’t you ever worried that when you say that the Earth is a planet, that it is a globe, you actually have to mentally position yourself as if you were considering it from out in space?” à— Critical Zones. The Science and Politics of Landing on Earth

What does it mean to be human? For a long time, we thought that it meant being caused by some entity greater than us, God or Nature. Then, somewhere around the 17th century, as Westerners became “modern”, we decided that being human meant standing tall and autonomous at the center of our universe, us, on planet earth, which itself was part of something greater — the universe.

A conception of what it means to be human comes with a conception of where we are being human. Depending on whether you see yourself as being caused by nature or being the autonomous cause of yourself, you will view planet Earth — where you are being human, as something that you can cause, or that causes you. If you see Earth as something that you can cause, you will also tend to think that it is something that you are, in some sense, outside of.

Being outside of planet Earth, in an “objective” relation to it, is the modern way of thinking about our planet. The posture of being “outside” is at the core of what I will refer to as the scientific challenge of “being a modern on Earth”: How can we study and solve planetary problems with some degrees of scientific objectivity while being nonetheless an integrative part of those problems?

Partly problem, partly solution

One way to start addressing the scientific challenge of being a modern on Earth is by taking seriously the fact that we are an integrative part of this planet. Concretely, this means that we must use a concept of “Earth” that captures the manner in which humans and non humans, biotic and abiotic entities coexist on Earth. That layer where biotic and abiotic entities coexist is what the authors of the quote in the introduction called a “Critical Zone”.

The concept of the Critical Zone is “a way to bring different disciplines together in order to refresh the study of the thin skin of the living Earth”. In short, the Critical Zone is this multidisciplinary object of enquiry that corresponds to our planet, which is the space where multiple biotic and abiotic agents conspire, actively and passively, directly and indirectly, through biochemical, cultural, and geophysical processes to make life possible... or impossible. How are we going to provide a unified methodology to study such a Critical Zone, capturing all its ecological and socio-cultural complexity?

Active inferencing into the Critical Zone

The goal of this blog series is to discuss how we can start addressing the scientific challenge of being a modern on Earth — the challenge of developing a science of the Critical Zone we live in and are part of, using the modeling methods of Active Inference (ActInf). ActInf was born from research in theoretical neuroscience and was initially aimed at studying human cognitive functions, such as perception, learning and action.

ActInf assumes that an agent’s brains embody a — Bayesian — generative model of the cause of the sensory impressions they receive. Based on that model, agents manage to infer sensory causes and to generate expected, future sensory impressions (a.k.a. perception, which is sometimes presented as “controlled hallucination”). As they infer sensory causes, agents also optimise their — prior — beliefs about those causes (a.k.a. learning). Agents can further infer the course of action, or policy (a.k.a. action) likely to yield desired sensory impressions in the future. Over time, cycles of action, perception and learning attune, or adapt the agent to the environment. Sensory entries become increasingly predictable, and adequate courses of action are readily performed.

Following the publication of ActInf’s seminal Nature article, more articles started rolling out on how to apply ActInf methodology to systems beyond humans and their brains such as plants, ecological niches and sociocultural ensembles. If ActInf could be used to simulate life and cognition in humans, in principle, why couldn’t it be used to model and simulate life and cognition in any system? What could it mean for non human entities to perceive, act and learn? The conceptual space for the application of ActInf was decomparmentalised, and researchers started to develop behavioral models of entities such as ant colonies  and even climate and the biosphere itself.

Mutual attunement in the Critical Zone

The ability to conceptualise what ActInf meant outside the context of brain and cognition was the main limitation on the scope of active inference, since from a formal point of view, an ActInf model could already apply to systems of any spatial or temporal scale, and could be used to compare systems in terms of a metric that can be summed, irrespective of the nature of that system.

Indeed, a system S1 of any size whose dynamics unfold over any timescale can be modeled with ActInf, and be compared with a system S2, of different size and temporal dynamics, and their respective assessment can be added. This is why we use ActInf to model Critical Zones. Because it can apply to biotic and abiotic, human and non-human systems, across temporal and spatial scales, and those models can be stacked so as to reflect the complexity of the earth system.

Modeling planet Earth means modeling our critical zone, where life is possible, in all its complexity. And I believe that active inference — ActInf, as a scale free modeling strategy, is a promising candidate methodology for achieving a model of Earth understood as an agent that sees us, learns about us, and acts on us, the same way that billions of other agents see it, learn about it, and act on it, in a search for mutual attunement and adaptation.

What’s next

In our next blog, we will take a deep — but not too deep — dive into Active Inference. We will describe how it can work as a general purpose strategy for modeling living systems. In following posts, we will discuss how ActInf can be applied to model specific systems, such as soils and forests, and how ActInf models could even be used to bridge earth science and tokenomics.

About the Author

Axel Constant is an Innovation Fellow at OpenEarth focusing on new practices of Artificial Intelligence as applied to whole Earth level. He is a PhD candidate in the Philosophy of Biomedicine at the University of Sydney working on evolutionary, cultural and computational approaches to Psychiatry.

Highlighted news

Research
-
Jul 2022

Stay up to date

Sign up for our newsletter and stay in touch.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
topographic lines in the background