생태계

우선순위와 도전 분야

생태계 개발은 구조, 성장, 사이버 및 인지 보안, 정보 흐름에 초점을 맞추며, Active Inference을 통해 해결할 수 있는 조직의 문제에 중점을 둡니다.

적용 분야

Active Inference은 생물학, 신경과학, 정신건강, 생태지역 모델링, 범주 이론, 계산, 경제학, 교육, 철학, 물리학, 로봇공학, 법적 시스템, 사회 시스템, 물류, 과학 방법, 그리고 분산 과학과 연결됩니다.

생태계 지원

생태계 지원에는 공개 커뮤니케이션, 학습 인프라, 프로젝트 가시성, 파트너 경로, 연구 자원 및 공유 오픈소스 도구가 포함됩니다.

Biology

Chris Fields and Michael Levin (2020) posit that Active Inference “provide(s) conceptual tools for reconceptualizing biology as the study of a unified, multiscale dynamical system”.

Ramstead et. al (2019) have leveraged active inference and the underlying free energy principle to characterize variational neuroethology, a theoretical ontology for living systems based on a recursively nested formulation of Markov blankets.

Friston et. al (2023) introduce a variational formulation of natural selection to explain how slow phylogenetic processes constrain—and are constrained by—fast, phenotypic processes.

Bioregional Modeling

An ongoing project, Biofirm currently consists of two main components in support of

1. Ecosystem Control System

• Active Inference-based multi-agent control framework
• Homeostatic regulation of ecological parameters
• Comparative analysis between random and controlled dynamics

1. BioPerplexity Analysis

• California county-level bioregion research using Perplexity.ai API
• Business case generation and pitch development
• Cross-document visualization and analysis

Category Theory

Active Inference, through the Free Energy Principle (FEP), provides a framework for understanding how systems make predictions and update their models based on sensory evidence. Category Theory, meanwhile, offers a formal mathematical language to describe the transformational processes that occur during these updates (see: Chris Fields 2024, "What is the Identity operator?").

Mathematical Bridge

Where Active Inference describes the necessity of prediction and error minimization, Category Theory provides the precise mathematical tools to track how these predictions and updates flow through a system. The power of this combination lies in Category Theory's ability to formalize the very transformations that Active Inference predicts must occur.

Creative Processes

This relationship becomes particularly relevant when examining creative processes. Active Inference explains why systems must make predictions and learn from surprises, while Category Theory's operators can formally map out the transformational paths taken - even in cases where the end state wasn't predictable from the initial conditions. The identity operator, in particular, helps us understand how systems maintain coherence while undergoing these creative transformations.

Practical Implementation

This growing theoretical bridge between and has practical implications for:

• Modeling learning processes
• Understanding system adaptation
• Tracking creative development
• Formalizing prediction errors and updates
• Maintaining system identity through change

The synthesis of these approaches provides a more complete picture of how systems learn, adapt, and create while maintaining their essential identity through transformative processes.

DeSci

Decentralized Science ( was explored in the work.

Domains of Application

The sub-sections of reflect some early collaborative efforts towards curating across different sectors and systems of interest. This section of the document is not presented as a comprehensive or exhaustive survey in any way, rather more of an invitation to those who would like to steward a section (keeping it updated and relevant) as we develop these synoptic capacities together. Later updates will more deeply reference and other resources where has been demonstrated across systems.

Along with other modern technical fields, Active Inference faces and addresses challenges of broad relevance such as (i) remote education, workforce development, and competency evaluation, (ii) user experience, ergonomics, and accessibility in a modern global context, (iii) availability, utility, reliability, and safety, (iv) participation in research and practice-oriented activities (v) cyber- and cognitive-security, (vi) theoretical and practical aspects of artificial intelligence explainability and safety, (vii) social and economic policy integration and management.

Integrations featuring Active Inference are increasingly being found across public and private sectors. These applications are enabled through common education around Active Inference themes, concepts, skills, practices, and tools. As such, there is potential for The Institute to facilitate both the study (theory and research) and professionalization (practice and implementation) of Active Inference within and across myriad sectors and disciplines, and to grow the incipient Active Inference Ecosystem and awareness of Active Inference by facing such challenges proactively and in a fashion aligned with our vision, values, and principles. We hope to achieve this through developing coordinated resources that are accessible to users at all backgrounds and levels of familiarity. Moreover, we aim to develop this nascent research arena by facilitating and/or mediating access to resources for an array of independent projects.

A core reason why is being adopted so rapidly is that it provides a flexible, agent- and action-oriented ontology which describes a great array of complex adaptive systems, up to and including human cognition.

The Active Inference framework can be used to describe systems at different nested scales. The applicability of Active Inference to multi-scale complex adaptive systems is a source of great explanatory power, and it is also a challenge for the framework’s coherence. Scholars from different disciplines or fields may read Active Inference concepts or constructs differently, and unknowingly build an error into their research ecology which is then propagated forward, thereby hampering progress in the field at large. To our knowledge, the Active Inference Institute is the first scaled attempt at directly tackling that risk by offering Active Inference education to learners of all backgrounds, and by working to specify an ontology that is both particular to Active Inference and broadly accessible. Furthermore, the institute offers accessible onboarding to current best practices in Active Inference research as well as the ability to drill down into specific topics across the broad array of disciplines that are implementing the framework.

Economics

Economics is a very broad field. From macro policy to econometric micro optimization. Here the focus is on conceptualizing the decision maker as it is relevant for deciding a relevant policy alternative from a potential set. Undoubtedly future work and potential authors will expand this section greatly.

The foundation of economics is to scale decision making to collective systems. Traditionally, decision makers are seen as utility maximizers (or regret minimizers). With the underlying assumption of full information and (bounded) rationality.

However, active inference nuances this view by positing that rational choice is a limit case of decision making. Only occurring during absolute certainty of observing one’s preferences (Friston et al., 2013). Instead a pragmatic turn entails information seeking as part of the decision process such that actions are both pragmatically and epistemically informed (Schwartenbeck et al., 2015).

Such a shift in perspective - all the way up to perspective swaps - may not be limited to traditional economics by expanding existing frameworks with new methodologies. Instead, this shift from viewing choice as static towards a dynamic process, means that multiple economic approaches to collective policy selection become feasible.

One such alternative economic approach is broad prosperity. It involves taking inventory of a set of value-neutral indicators, of which gross domestic product is just one. Unfortunately, it is very difficult to express the causal relationships between these indicators as these span a variety of domains like social, environmental and economic concerns. Additionally, what occurs locally has impacts globally and vice versa (TNO, 2024).

Active inference is poised to address these limitations. Given the nature of scale-free action perception loops; any self-organising system may be described as a sense-maker. In doing so solve the issue of not being able to sum free energy across agents. For example when planning a new public transport line. One could calculate the total utility obtained via preference elicitation (willingness to pay, stated and revealed choice experiments). Or one could instantiate a niche constructing digital twin. The entire urban region which is assumed to itself be a scale-free niche constructor will then have to share its niche with a synthetic artifact.

Evaluating the potential of a policy alternative, like building a tram or bus line, becomes a practice of understanding the generative model of the digital twin. Which is assumed to approximate a real niche constructor.

Education

The transdisciplinary nature and flexibility of Active Inference makes the framework ideal for practical, theoretical, and interoperable work across myriad use-cases. In the use case of learning in systematized settings (i.e. the conventional planning frames take on wheels (π, as in policy selection) in order to function as a platform enabling translational “spinning” (i.e. helicity) across contexts of greater scale (learning generalization as transfer). With the inertia from the spin as your stability mechanism, the addition of policy selection by the learner as a self-organizing system (i.e. learning agent) within larger variability retained settings, introduces uncertainties to test the which and the where of when trans-disciplinary experience (i.e. real world experience, real dynamism, real problems) requires practical/pragmatic (i.e. action) solution(s). Comparatively speaking, conventional frame containment as stabilizer, only provides a variability reduced-reductionism environment ubiquitously held up as constructing learning where the product is a wide base as “foundational” retentions, and relatively smaller “crowning” states, as in Maslow’s Hierarchy.

Before describing what the mechanics of this inclusion of policy selection is, and can look like for you, it is best to point out that going forward, the acceptance that policy selection plays a role in how we learn, is not necessarily easy to incorporate as strategy applied. "I find this policy selection part hard to understand" is often heard when something new and/or unfamiliar is introduced into a messaging exercise. This is understandable when a proposal uses terminology that isn’t part of the newcomer’s current lexicon (and sometimes even when the term is already used). To take up new labels (and the ideas behind them) requires taking a step back from centuries of the accepted definition of what providing an education...is: define and refine via a process of packaging and delivery of information (so deliver to me, the learner, what I can recognize). Sustainers and defenders of that (status quo) strategy will argue (correctly) there is much more going on than that minimum of two of define and refine, and the Active Institute’s argument would be...maybe, possibly, but not certainly.

There too many examples, practiced both currently and historically in academia, to deny that at the core of educational practice, there is a reinforcement and incentivisation firmly established around practices focused on defining (i.e. agreement around an external ontology/standards) and refining (i.e. moving to smaller and smaller divergence(s) from what we see/do, and what we think we're doing/seeing). That being the case, new terminology like Prediction Matter Expert is the surprise given that phrase’s like this that are introduced, lack consensus around meaning and precision. Time is then spent working through where the introduced term/label/idea can fit (appropriately) within contexts of particular study/focus/research. This is an effortful exercise, that can often lead people new to Active Inference and the FEP, to wonder “where exactly is the Institute going with this idea/terminology/set of formalisms?” That’s a fair question, and in asking, we open a portal to the navigational aspects of resolving the “where” of learning as orientation process. This is the “where am I?” action - not just wonder - as Active Inference.

Applying Active Inference and the FEP to educational programming - “you are now here, but you’re not staying here, you’re going back out there” - has thus far struggled to gain much traction in many legacy (read hierarchical Pyramid Model) educational systems. Given most education systems’ tendencies to want to place the certainty of keeping systems accountable ahead of determining how agents learn when prediction-as-skill under uncertainty is given equal priority with subject matter expertise (as skill), we continue to find that active inference as functional compliment needs time for mass academic uptake (to scale). One of the core differences between subject matters and prediction matters exists at the waypoint called Updating. Currently, legacy education systems interpret “updating” as a cumulative-constructive-classical exercise, and therefore it is surprising for those vested in that method, when someone with formal active inference priors, proclaims the need to incorporate statistical and probability functions into the praxis and pedagogy design. This non-binary nature of probability (i.e. could be zero, or one, or something between) aspect dependent on “what I as agent…thinks will happen,” does exist as a teaching strategy, but is only applied within the variability reduced frame, pre-selected by the course/activity/lesson plan designer who is the subject matter expert.

And, active inference prediction modelling begins with the concept that the learning agent is first and foremost a self-organizer, self-designer who wants (self-identifies) minimization of any divergence between their own model and what the niche continually signals. Under this circumstance, updating as a process may take on constructive attributes, but it will also require some exposure to de-constructing processes (i.e. the most basic being, when change in the situation is apparent, will the agent 1) accept that change and 2a) either modify their surroundings or 2b) modify their model?). This is a fundamentally different type of branching - change the model, change the environment, change both - to pass/fail or even rubric induced accounts. This then necessitates a different (second) definition of “updating” as a result of starting with a predictive probability of achieving an ad hoc and post hoc processing threshold (could be described as ALL moves cardinal vs. NEXT moves ordinal/sequential), before “right and wrong” or even “75% correct” as assessed (as the 25% “wrong” usually doesn’t carry forward past the filters of constructive practices).

So why does this difference matter? In arriving at a threshold minimum, the active inference learning agent needs to reconcile while also keeping records. That “25% wrong” for example, is actually valuable information (not to be discarded) if divergence minimization is one of the stated goals. Now the question becomes “do I let go of what I predicted wrong because it didn’t affect my pass/fail status, do I let what I got wrong change my aspirations because I haven’t achieved perfection, or, do I look at Right-Wrong as a proportional measure from which to make future decisions?” (more on this shortly). Taking accounts and making reconciliations, is the process of modifications byandto which the updating of the active inference generative model, evolves. The conventional view of update as build-up, build-forth (Subject Matter Expertise, SME), is now complimented with a Prediction Matter Expert (PME) view of “what can I as learning agent let go, in order to arrive at a new know?” as policy selection to be determined. Borrowing from Chris Fields’ Identity Operator presentation, PME’s cope better with the undecideability in the frame problem - what doesn’t change as a result of an action. Using Chris’ terminology, “circumscribing what I don’t have to worry about”…means “I” can now take my “eye(s)” off of certain contents so as to increase availability for new [to me as agent] contents. Under this condition, the forensics come before, and not just after, a learning episode, making policy selection (π) now one part agent domain, one part external plan designer/niche reducer domain - with All Moves now meaning all of the puzzle pieces are present, and each is connected regardless of order application.

Of course, once the differences between legacy systems perspectives and active inference perspectives are held up as the parameterized space, the ability for the learning agent to oscillate between perspectives (i.e. perspective swap as action) becomes available. This oscillating process - first back, then forth...and never forth-only - is not uncommon. Agents swap perspectives when pairing science with fiction, active with inference, math with art as comparative with collective proportional measuring (as minimum) processing (unit of) analysis.

Which leaves the Institute with a challenge: how do we continue to attract Subject Matter Experts and point to the fact that Subject Matter Expertise alone can only take one so far as a navigator in variability retained settings? Another way of putting this could be stated as, as an institute, can we afford to not talk about the gorilla in the room: how we learn (define and refine...and retain) needs a co-pilot (what can I let go...to arrive at a new know?). This being asked as AI and LLM's train on far more information than humans can, to derive that synthesis (here's your answer!) that defining and refining puts out (outputs).

Let’s look at a real world example already introduced to the officers of the Institute where subject matter expertise attracted agents to the institute, and, the institute had to find a way to help the “experts” let go of what they already know. In this case example, Active Inference has been linked to the process of early childhood education (Montessori programming). Under Montessori philosophy, teacher’s are described as “directors” with a focus on “independent learning.” Comparisons can then be made to other early childhood education approaches. The Reggio Emilia early learning method holds up their philosophy of teachers roles as “partners” and “guides.”

The question then becomes one of: as the learning agent ages - enters different “grades”, stages and phases of Updating as a result of predictive processing (probability now based on increased temporal depth) - does the teacher as multi-hat wearing director/partner/guide/coach/facilitator still fit the needs of the self-organizing learner going forward? Perhaps, if the learning is organized as an adventure as a proxy for authentic - where once again authentic is trans-disciplinary real world experience, real dynamism, real problems) requiring practical/pragmatic (i.e. action) solution(s), while an adventure is a simulation.

Or, as a PME enabler (Not trainer), does the teacher SWAP titles - by subjecting themselves to the Identity Operator process - of Teacher with Way Finder (navigator), initiating their own perspective exchanging process of self-identifying (minimizer of divergence between their own model, now as minimum(2) dual-state swap able [i.e. Gripper & Gripped - BY and TO - simultaneously], with what the niche continually signals) resulting in an SME + PME hybrid triangulating with ANY niche (not just their subject specialty)? This would require teachers to both teach and co-learn interchangeably.

As the reader can appreciate, this is a different condition than teachers staying close (closed) to what they know (SME dilemma) and thus self-selecting away from “what can I let go, in order to arrive at a new know?” This is where the Institute’s role as director/partner/guide/coach/facilitator ends, and a co-piloting triangulation exercise (i.e. simulations to actualizations and Back) begins.

Going forward, it is the Institute’s ambition to make clear that the channel (i.e. gap) between legacy systems developing subject matter experts and what we view as new affordances that can be realized when uncertainty-as-learning-tool is perceived as a feature - as prediction matter expertise - is a potential exponentiator of a learner’s predictive capacities within and beyond systematized and variability reduced settings. We choose to be partners in this enterprise, as we feel serving in that capacity is closer to co-piloting than co-hosting in a flight simulator. Every organization wonders where the “stay afloat” energy will come from. In our case, we policy select to work with people vested in research with a specialty focus who also want to be able to generalize (play in “Scale Free”) with higher degrees of confidence when necessary (be a trans-disciplinarian when the niche is open, and variability retained).

Implementations of Active Inference

In the project, we have curated dozens of at this link.

The sub-pages here go into more detail on several different toolkits for applying including in Python ( Julia ( Matlab ( and Prolog (

Legal

Cases mentioning active inference

Patents mentioning active inference&oq=%22active+inference%22)

Logistics

“Enhancing Population-based Search with Active Inference” (Dehouche and Friedman, 2024)

Mental Health

Being a theory of embodied and sentient behavior, Active Inference can contribute in knowledge sharing to better understand the intrapersonal and interpersonal dynamics involved in or implicated in mental health (Pezzulo et. al, 2024). Computational psychiatry (Friston 2022) serves to utilize models of cognition and behavior to predict and account for the above-mentioned dynamics. Being a model constrained by Bayesian principles and the free energy principle, Active Inference allows for one such attempt at better predicting treatment outcomes, nosology, and fundamental principles of cognition.

The Institute supports individual thoughts and projects designed to inquire on topics related to social sciences, psychology, and mental health. Such projects have included attempts at classifying and clarifying Active Inference ontology to better fit the lived experience of individuals with posttraumatic stress disorder.

Active Inference is a systems approach to psychological constructivism that offers a trans-diagnostic perspective to readers. One such benefit of a trans-diagnostic approach is that it identifies connections between different processes without the strict adherence to philosophical requirements. Areas of the theory that can be beneficial to mental health research include:

• Experiential quality of prediction error for patients (i.e., as a mismatch of one’s generative model) (defining “surprise” in therapy practice, Holmes & Nolte 2019)
• Homeostasis and role of consciousness as allostatic control (Krupnik 2024); as well as the dynamic interplay between these processes and mental health symptoms (cultural identity, Ramstead et. al 2016; social conformity, Constant et. al 2019)
• Mental health symptoms as under/over-reliances on a generative model (apathy, Hezemans et al. 2020)
• Requirement of interoceptive processes (body-based) and the roles these have in the make up of a Bayesian brain (Duquette 2016)
• Role of affect and ascribing confidence to one’s generative model (Hesp et al. 2021)
• Equal treatment of action policies as being direct manipulations of one’s environment (decrease free energy now) versus epistemic transformations (change your beliefs about the world to decrease free energy in the future) (PTSD & explore-exploit dynamics, Linson et al 2020; social cognition, Gallagher & Allen 2018)
• Hierarchical models of cognition that outline the dynamic interplay of predictions, actions, habituation, and environmental feedback (theory of constructed emotion, Barrett 2017; cognitivism versus autopoiesis, Allen & Friston 2016)

Conceptualizations are being offered that describe the experience of those with particular mental health symptoms (Parr et al. 2022, p. 186). Active Inference has also been applied to the study of depression (Barrett et al 2016), psychosis (Knolle 2023), schizophrenia (Jeganathan 202130527-7/abstract)), anorexia (Barca et al 2020), functional neurological disorder (Jungilligens et al 2022), and interoceptive dimensions of psychopathology (Paulus et al 2019; Barrett 2016). Conversely, Active Inference also provides a framework for understanding constructs of mental wellness, including subjective well-being (Smith et al 2022).

Chamberlin (2023) illustrates how Active Inference can be applied directly to one psychotherapy model, Coherence Therapy. This type of dialogue allows readers to see the neurological mechanisms and meaningful narratives at play in a framework that treats both equally. It is also good for readers to note that Active Inference offers a framework to reformulate agents as being cognitive, emotional, and embedded without adding other philosophical requirements. It can be beneficial to engage existing psychological theories [of cognition, emotion, personhood, agency, social relations] in order to emphasize constituent processes that Active Inference gives language for. For instance, the focus on sense making in the life of an individual highlights the existence of an agent’s generative model that has been determined within and a part from the generative process. In parallel, sense making can speak to themes of agency and emotional expression.

Neuroscience

Active Inference emerged from the field of theoretical neurobiology (Friston, 2005), where it was “first used to model the function, structure, and dynamics of the human brain” (Ramstead, 2024). It built upon foundational work in predictive coding (Rao and Ballard, 1999) and the Helmholtzian concept of perception as “unconscious inference” (Helmholtz, 1867).

Active Inference’s central premise that “all neuronal processing (and action selection) can be explained by maximizing Bayesian model evidence — or minimizing variational free energy” (Friston 2017) provides a unifying theory to explain and predict myriad aspects of brain function and behavior (Friston, 2010). As such, it has been applied to many areas of neuroscientific research.

Active Inference models are used to provide parsimonious explanations for neural mechanisms and motifs, such as canonical microcircuits and neural networks (Bastos et al 201200959-2), Isomura et al 2022).

Researchers have furnished Active Inference models for phenomena including motivated control (Pezzulo et al 201830022-6)), sense of agency (Friston et al 2013), modulation of uncertainty by the dopaminergic system (Friston et al 2012), and the computational relationship between interoceptive and exteroceptive neural systems (Allen, 2022).

Active Inference frameworks have also been used to explain the dynamics of a variety of neurological and psychiatric conditions, including depression (Barrett et al 2016) and schizophrenia (Limongi et al 2023).

Recent studies have shown that in-vitro neuronal networks self-organize in response to stimuli in ways that are consistent with, and predicted by, the Free Energy Principle (Isomura et al 2023). The FEP also provides theoretical commitments towards testable theories of consciousness (Whyte et al, 2024).

As a multi-scale theory, Active Inference aims to ground neurobiology in physics-as-information-processing, and links it to other domains of inquiry, including diverse intelligence (Levin 2023) and artificial intelligence (Friston et al 2024).

Philosophy

Active inference provides a mathematical model of sense-making. Philosophy is the study the components and dynamics captured by these mathematical models. It comprises of a broad literature in philosophy of mind spanning history. As such active inference is not an island but densely connected to other descriptions and mathematical models of sense-making. Each with their own components and dynamics. One such advancement is within the field of neurophenomenology that seeks to build a dialogue between neuronal processes and philosophical constructs as experienced (Sandved-Smith et al. 2021).

Developing new implementations and applications of active inference benefits greatly from its philosophical context. Theoretical advancement of how to interpret existing phenomena through the lens of active inference is not just to fill shelves with studies. Advancement is essential to improve algorithms and inform applications across domains. After all, there are many domains which have mathematical and conceptual models of sense-making. Each of which could potentially be evaluated through an active inference lens.

Physics

See by Chris Fields, and more references to come.

PyMDP

The package is an specifically “A Python package for simulating Active Inference agents in Markov Decision Process environments.”

Active Inference ModelStream 007.1 ~ Conor Heins & Daphne Demekas ~ pymdpActive InferenceModelStream 007.2~ pymdp

Robotics

• See JF Cloutier’s project, (project documentation)
• Second in 2022 had a focus on Robotics, see program.

RxInfer.jl

RxInfer.jl ( is a programming package of functions developed at BIASlab in Eindhoven, Netherlands. It attempts to commoditize making it suitable for engineering applications. Compared to existing like RxInfer is unique in the sense that it draws upon reactive message passing on Forney Factor Graphs (FFG). Whereas ‘traditional’ implementations rely on Bayes graphs in the form of Partially Observable Markov Decision Processes (POMDP). FFG’s using reactive message passing only perform calculations when necessary, hence there is no underlying clock which schedules calculations. The reactive paradigm thus may offer computational benefits in certain situations, and enable favorable scaling properties for Active Inference models.

The at the Institute collaborates with with the developers of in development, such as developing visualisation techniques of the FFGs within the code editor.

Core Capabilities

RxInfer.jl provides powerful features for probabilistic modeling, including:

• Streaming dataset processing through reactive message passing
• Hybrid models combining discrete and continuous latent variables
• Scalable inference for large models with millions of parameters
• Automatic differentiation support for parameter tuning

Scientific Method

See

Distributed Science - The Scientific Process as Multi-Scale Active Inference (Balzan et al. 2023)

Generative Research Teams: Active Inference Compositions For Research and Meta-Science (Friedman & Smekal 2023)

Social

Active inference research in the social domain tends to focus on modeling communication and the sharing of belief models within groups. Such topics can be understood as pertaining to normative processes of group cognition. Over time, we can expect research to extend further to pathological examples of group cognition, assessment of group cognition quality, steering of group cognition to improve quality, evaluation of the cognitive architectures used during group cognition (e.g., rules, policies, computational tools, communication tools, attention mechanisms), and evaluation of group cognition where the group is a political body (such as a city or nation).

Group cognition rests on the communication of (potentially dynamic and evolving) belief models—-the internal generative models that individuals use to predict and explain their world—and consensus building with respect to beliefs. As described by Albarracin et al., 2024 for a generic group in the normative setting, “group members can be seen as actively and implicitly aligning their beliefs and expectations through dialogue and interactions, thereby enhancing their ability to predict each other’s actions and intentions, and thereby coming to perceive and act in the world in similar ways.”

Most humans do not conceive of their own beliefs in terms of models, however. Rather, humans tend to experience their beliefs and make sense of the world in part through narratives (Bietti, Tilston, and Bangerter 2019; Turner et al. 2023; Fanti Rovetta 2023; Cordes et al. 2021). These can be internal narratives that a person constructs, adjusts, and recites to himself or herself, or social narratives that are shared within a group. In the active inference context, Albarracin et al. (2021) consider social scripts, which are widely-supported prescriptions about how one is to behave in various social settings, or what is important in those settings. Bouizegarene et al. (2020) consider shared narratives conceived of more broadly. Social scripts and shared narratives help humans to generate more accurate predictions about the world and to coordinate social behavior.

The CogNarr Ecosystem, an Active Inference Institute project, has as a goal the facilitation of group cognition at scale, through sharing of belief models (Boik, 2024a, Boik, 2024b). This is an extension of previous work that viewed core societal systems (e.g., economic, financial, and governance systems) as part of the cognitive architecture of political bodies (Boik, 2020a, Boik, 2020b, Boik, 2021)

A large body of active inference research, perhaps thousands of papers, at least mentions the social setting. In addition to some articles already cited, articles in which the phrases “active inference” and “social” appear in the title include the following:

• Gallagher and Allen, 2018. Active inference, enactivism and the hermeneutics of social cognition.
• Hipólito and van Es, 2022. Enactive-Dynamic Social Cognition and Active Inference.
• Constant et al., 2019. Regimes of Expectations: An Active Inference Model of Social Conformity and Human Decision Making.
• Cheadle et al., 2024. Active Inference and Social Actors: Towards a Neuro-Bio-Social Theory of Brains and Bodies in Their Worlds.
• Kirchner et al., 2022. Better Safe than Sorry?-An Active Inference Approach to Biased Social Inference in Depression.
• Tehrani-Safa et al., 2024. Learning Risk Preferences Through Social Interaction: An Active Inference Approach
• Fox, 2021. Active inference: Applicability to different types of social organization explained through reference to industrial engineering and quality management.
• Bezzazi, 2021. Social Active Inference.
• Ohata and Tani, 2020. Investigation of the Sense of Agency in Social Cognition, Based on Frameworks of Predictive Coding and Active Inference: A Simulation Study on Multimodal Imitative Interaction.
• Matsumura et al, 2023. Social Emotional Valence for Regulating Empathy in Active Inference.
• Solymosi and Schulkin, 2024. Creative Resilience. Flourishing and Valuation through Social Allostasis and Active Inference.
• Tani, 2019. Accounting Social Cognitive Mechanisms by the Framework of Predictive Coding and Active Inference: A Synthetic Experimental Study using Robotics Interaction Platforms.

Some of these topics and papers were explored in the 2023

SPM (Statistical Parametric Mapping)

Statistical Parametric Mapping (SPM, homepage) represents a pivotal development in the history of active inference and computational neuroscience. Created by Karl Friston at the MRC Cyclotron Unit in the late 1980s, SPM began as a statistical technique for analyzing brain imaging data, particularly fMRI, PET, and EEG data (Wikipedia and History).

The development of SPM marked a crucial shift from simple region-of-interest analyses to whole-brain statistical approaches. Originally written in MATLAB, SPM91 (also known as SPMclassic) became the community standard for analyzing neuroimaging studies within a few years of its release. The software's success stemmed from its rigorous approach to making valid statistical inferences about brain responses without prior knowledge of where those responses would occur.

SPM's theoretical framework evolved to incorporate increasingly sophisticated statistical methods, including the general linear model (GLM) and Gaussian field theory. This evolution paralleled and supported the development of active inference theory, as many of the mathematical principles underlying SPM - particularly those involving free energy minimization and Bayesian inference - became foundational to active inference. Today, while dedicated toolboxes exist in various programming languages (like in Python, in Julia), SPM remains significant as both a historical cornerstone and practical tool in the field.

Symbolic Cognitive Robotics

Symbolic Active Inference, developed by Research Jean-Francois Cloutier, represents an innovative approach to combining symbolic reasoning with active inference principles. The framework aims to bridge the gap between traditional symbolic AI and the free energy principle by implementing active inference using symbolic representations and logical reasoning. This implementation allows for explicit modeling of beliefs, goals, and actions using symbolic structures while maintaining the core mathematical principles of active inference - namely the minimization of variational free energy and expected free energy.

The approach enables systems to perform goal-directed reasoning and planning through symbolic manipulation while grounding these processes in the formal theory of active inference. Key aspects include the representation of generative models using symbolic structures, belief updating through logical inference, and action selection based on expected free energy minimization. This synthesis provides several advantages: it makes active inference more interpretable through explicit symbolic representations, enables complex reasoning about abstract concepts and relations, and allows for more efficient computation compared to purely numerical implementations. The framework has been demonstrated through implementations in domains like robotic planning and symbolic problem-solving, showing how symbolic representations can be effectively integrated with active inference's information-theoretic principles. This work represents an important step in developing hybrid AI systems that combine the strengths of both symbolic and probabilistic approaches to intelligence.

All information on can be found at

Ecosystem Development: Structure and Growth

Community Growth and Development

Here, we present the community growth and development model for built on the following 5 core components:

1. Awareness. Promoting and fostering awareness and use of Active Inference, and developing partnerships with well-aligned organizations and communities.

2. Developing and disseminating educational materials, contributing to competency, capability, and common language within the community.

3. Common Forum. Offering and maintaining an inclusive and accessible common forum for discussing, sharing, and hosting relevant work and opportunities, finding collaborators, and networking (i.e., an informational commons).

4. Support. Providing support for emergent teams and projects which align with The Institute’s mission, in the interest of innovation and impact

5. Governance. Maintaining stable governance for cultivating and sustaining partnerships, technical infrastructure, and sponsors.

Ecosystem Structure

The Institute cultivates an active and engaged ecosystem around the scientific modeling framework of Active Inference. This vibrant Ecosystem and community drives innovation on the research front and makes significant strides in providing accessible education. The Institute ensures that efforts are well-aligned, impact-focused, relevant, and meaningful in advancing research and education for the betterment of society by forming partnerships and by engaging with and growing the Active Inference community. Our community development model emphasizes facilitation over management, and distributed as opposed to command-and-control strategies. More importantly, our model moves beyond the provision of networking and discussion space to support emergent, collaborative work.

In these regards, the Institute functions as a seed crystal that can help to foster phase changes across a variety of information system domains and applications. The Institute does not directly manage all of the systems upon which it has an influence, but instead seeks to leverage its influence by providing coherent multiple tools and practices from which communities of shared interest can optimize their local information system dependencies for active inference efficiencies.

As opposed to a linear “funnel” growth model, The Institute will implement a cyclical model of organic growth pursued through the incubation among participants of (i) self-efficacy, or a sense of personal capability, (ii) a sense of support and safety, and (iii) a sense of investment and impact in participants, as a basis for forming a sense of community and providing the foundation for development of relationships within the community through positive, repeated contact. The support of these senses leads to productive, emergent collaboration, which in turn leads to emergent community narrative, norms, roles, and “scripts”. Participants are reinforced in their feelings of capability as a part of a team, assured that they will be provided with support in a reasonably safe environment, and that results will have a lasting, positive impact on their community. Resulting research and educational artifacts and documentation constitute shareable content which can then be used to bring awareness about Active Inference and The Institute to non-community members.

Where a “funnel growth” model focuses on awareness alone as a basis for developing a user-base, our model’s focus on education, knowledge sharing and presentation of work, and support for teams allows for non-community members of all backgrounds and interests to engage with and contribute to the community, thus affirming membership through a sense of shared investment, impact, and competency. Further, where online learning communities anticipate members terminating participation following completion of coursework (or after achieving feelings of self-efficacy in the material), our model’s provision of support and opportunities for sharing of work with professionals and academics provides incentives for continued engagement and participation to those who feel they have already become reasonably familiar with all available educational material.

Below, background is provided on the (i) structure of the community (i.e., user segmentation), (ii) our information storage and dissemination technology (“tech”) stack, (iii) our communications plan, (iv) the education, support, and infrastructure and governance functions we provide and/or intend to provide as a part of this model, and (v) our intended approach toward evaluating quality control and growth.

Community Participants

The Institute is a formal organization that has been constituted to serve some of the organizational and operational needs of the expanding active inference ecosystem. The Institute and its staff recognize that the energy and knowledge value relating to the further understanding and development of active inference resides in the broad active inference community, which is supported, fostered, convened and cross fertilized through the activities of The Institute. The reach and potential implications of active inference across domains and sectors is sufficiently broad that parties can choose from among many different ways to engage. A partial list of categories of participation is presented below to provide a sense of the variety of participants.

Direct Institute Participants (Members and Learners)

Many individual participants interact directly with The Institute and its resources and programs. Participants include members of the Active Inference Ecosystem, or those who engage directly with and contribute to These participants include students, educators, researchers, and professionals from around the world who may benefit either from awareness of Active Inference and its implications, developing related competencies and having opportunities to network and collaborate with individuals who do, or from opportunities to collaborate and share work and insights which would be valuable to the Active Inference Ecosystem.

Participants also include learners at various levels of involvement and expertise that engage directly with The Institute as part of their learning process. The Institute seeks to support all learners, from the academic expert to those individuals who are not, and everyone in between. The Institute seeks to facilitate access by all learners to tools and materials and narratives that can help people at all levels access information that can help them to enjoy the direct and indirect benefits of active inference thinking and approaches.

Users (Adopters and Beneficiaries)

For individual and organizational users that explicitly adopt Active Inference-based [organization and operation] of their information processing and synthetic intelligence practices, policies and tasks, the Institute’s productive outputs provide support and opportunities for engagement with a broader community. The Institute maintains an online resource center that includes software, tools, and materials that convey methodologies and practical pathways for instantiating Active Inference-derived structures in a variety of community settings and institutional contexts, and includes [practical suggestions for] the facilitation of Active Inference itself as an open source and open standards set of products and practices. As such, the community using Active Inference and related products requires documentation, clear messaging regarding updates, and guidelines on fair and best practices. By considering such beneficiaries of Active Inference as “users,” The Institute may leverage existing best practices from other domains, such as user experience, requirements engineering, and software engineering. Potential users include professionals, researchers, educators, and engineers.

Beyond direct “users” of active inference, there are many groups of parties that benefit from the use of active inference who won’t interact directly with such systems, nor be aware of it. Comparison is made to people who fly in airplanes, but haven’t studied Bernoulli’s hydrodynamics principles.

Research Partners (External Research Organizations and Working Groups)

The Institute’s ReInference unit collaborates with external research partners, universities, institutions, and subject matter experts. These partnerships involve joint research projects, data sharing, and knowledge exchange to enhance the depth and breadth of research efforts. Collaborations with research partners create an opportunity to enrich The Institute's research capabilities and resource access, thereby accelerating the generation of new knowledge and helping us to address complex research questions, validate findings, and extend the reach of our research impact. Potential research partners include organizations working on or faced with problems that may be solved by Active Inference, and organizations which are working on or have solutions to problems which The Institute and the community are facing.

Educational Partners (Universities and Educators)

The Institute’s EduActive Unit collaborates with educational partners to influence, instantiate, share, and get access to educational programs, teacher training, and learning resources. By partnering with educational institutions, The Institute extends its educational reach and impact and fosters effective delivery and dissemination of its educational content. Potential educational partners include universities, tutors, educational institutions, and educators.

Funders (Donors, Supporters, and Funding Agencies)

The Institute requires in order to keep pace with community needs, maintain information infrastructure, and assist researchers in finding their own financial support for relevant research initiatives. Potential donors and funders include generous community members and beneficiaries, government funding agencies, private philanthropic donors, and sponsors of events, programs, and initiatives.

Ecosystem Priorities and Challenge Areas

We look to continued engagement with the Ecosystem, to better curate and refine the

Education: Scientific Literacy and Workforce Development

Active Inference relies on mathematical formalisms and is loaded with abstract conceptual challenges that transcend disciplinary boundaries. We hope to model educational processes such as pedagogy, competency evaluation, and professionalization in Active Inference. Thus, the Institute catalyzes workforce development, seeks to stabilize the "research to practice" gap, and contributes to the broader project of participation in scientific ecosystems.

Research: Grounding the Cognitive Sciences in Physics

Research across the natural sciences suffers from a lack of theoretical integration and practical collaborations. Active Inference is gaining traction as a rigorous attempt at a unifying first-principles accounts of vital features of biological systems, transcending disciplinary boundaries. At The Institute we promote this theoretical integration through various educational programs, supporting learners of all backgrounds.

Information Science and Diverse Intelligences

The interaction frequencies of modern information environments are higher and more complex than ever. At The Institute we apply Active Inference to understanding, monitoring, evaluating, refining, and developing artificial and synthetic (e.g., human-machine interface, organizational, crowd) intelligence systems. In this way, active inference helps to identify, analyze and optimize various forms of "interoperability" across various forms of intelligent system, making possible a form of "mutual socialization" among such systems. This work is enacted by projects currently related to information science, ontology, data quality control, artificial intelligence explainability, and knowledge engineering.

User Experience, Accessibility, and Sociotechnical Design

It remains an open challenge how to most effectively, efficiently and fairly enable sustainable engagement in digital systems consistent with all parties expectations and needs. At The Institute we map cognitive frameworks as a framing for design, user experience, ergonomics, and requirements engineering, as well as implementation and operational guidance, to offer new methods and tools to a wider community of professionals and scholars.

Business Applications

Business and commercial interactions are typically characterized by party attention to reduced set of abstracted variables as compared with biological and social systems. Notwithstanding the "management" and regulation or variables, active inference can still help to improve the competitive insights and risk mitigation strategies and other variables that are the focus of business and commercial parties. Active inference research and analysis promises to substantially enhance and improve critical business elements such as risk strategies, insurance markets, banking (lending criteria), identity authentication, and authorization and a host of other business interaction decisions.

Welfare

The scale independence of active inference analysis causes it to be well suited to framing issues in settings where different parties experience different levels of information and resources. This includes various programs of local and global social welfare that seeks to enhance the local and global fairness of resource allocations of various kinds and to offer a pathway to easing the consequent burdens that unbalanced resource related interactions place on precarious populations.

Cyber and Cognitive Security

Individuals and organizations today are confronted with a rapidly-evolving landscape of threats to digital and cognitive security. At The Institute we work to unify cognitive frameworks with existing cyber security and emerging cognitive security concepts and frameworks, to understand, measure, and address local and global information technology risks and impacts more effectively at multiple scales.

Scaling the Active Inference Ecosystem

The nascency of the Active Inference Ecosystem enables us to take a proactive approach towards various areas of consideration. At The Institute we create synergy among the efforts applied to the above challenge areas, and emerging needs of the Active Inference Ecosystem. This approach creates an opportunity to learn by doing and to embrace convergence research, where implementations are developed in parallel with theory, supported by regular information sharing and collaboration among practitioners and researchers.

Applying Active Inference

The Institute brings insights from empirical and theoretical Active Inference research into practice by designing new projects or communicating with existing projects that design and implement social system infrastructure, such as health infrastructure and cultural technologies that support human well-being. We also support that design and implement solutions to various collective problems, such as climate change, threats to democracy, armed conflict, or overall polycrisis.

Additional Focus Areas

• Software usability and accessibility
• Information system optimization and efficiency
• Cultural heritage and progress
• Legal and regulatory consistency and compliance

Ecosystem Projects

There are many — here we include the subset which have completed a form at to increase their visibility and participation.

See for all projects by members, and

The Active Inference Ecosystem

The is a vibrant, global community of researchers, practitioners, and enthusiasts united by their interest in — a powerful framework for understanding cognition, behavior, and complex adaptive systems. The ecosystem extends far beyond the formal boundaries of the encompassing a wide array of individuals, organizations, and projects that contribute to the development and application of Active Inference across

At its core, the Active Inference Ecosystem is characterized by its open, collaborative nature. It brings together experts from fields as varied as neuroscience, artificial intelligence, philosophy, physics, and social sciences, fostering cross-pollination of ideas and innovative approaches to complex problems. The ecosystem thrives on the collective efforts of its participants, who engage in research, education, software development, and practical applications of Active Inference principles.

The ecosystem is not just an academic or theoretical construct; it is a living, evolving network of interactions and initiatives. It includes among organizations, educational programs, products, events like the various community-driven efforts. The Active Inference Institute serves as a hub within this ecosystem, providing infrastructure, coordination, and support to facilitate the growth and impact of Active Inference across disciplines and sectors (see for how this has unfolded over the years).

As the document transitions into detailing the Active Inference Ecosystem, readers can expect to explore the across