Skip to content
Active Inference Ontology
Full Ontology
Term
Tag
Ref_1
Ref_2
Ref_3
Ref_4
Ref_5
Def_1
Def_2
Def_3
Def_4
Def_5
Proposed_Def_1
Proposed_Def_2
Proposed_Def_3
Proposed_Def_4
Notes_1
Notes_2
Notes_3
Notes_4
Notes_5
Notes_6
Notes_7
Notes_8
Notes_9
Notes_10
Notes_11
Notes_12
Notes_13
SUMO superclass, instanceOf, superrelation
Subclasses, subrelations
Domains
Other SUMO relations
1
Accuracy
Core
Accuracy is simply the surprise about sensations that are expected under the recognition density
*Accuracy is a subclass of PsychologicalAttribute
An
instance
of
*Accuracy
is
the
number
2
argument
abstractCounterpart
.
*Accuracy is internally related to TruthValue
2
Action
Core
Action A : S × R → ℝ... corresponds to action emitted by an agent and depends on its sensory and internal states.
RECOGNITION models update internal parameters that correspond to external states (including hidden causes of the environment), blanket states, and internal states (meta-modeling). In contrast, GENERATIVE models take those same internal parameter estimates and emit expected or plausible observations.
*AbstractAction is a subclass IntentionalProcess.
3
Action Planning
Core
The requirement for an adaptive organism to predict the consequences in the future of engaging in this or that action. Also called the problem of mere versus adaptive active inference, or planning as inference.
Planning is a subclass of IntentionalPsychologicalProcess.
4
Action prediction
Core
In principle, the recognition of action goals might be implemented in perceptual and associative brain areas, similar to the way other events such as visual scenes are (believed to be) recognized, predicted and understood semantically.
We propose that the brain automatically predicts others’ future actions while perceiving their current actions. The human perceptual system constantly engages in this kind of reflexive prediction.
Predicting is a subclass of intentional psychological process
5
Active Inference
Core
(??)
Under the free energy principle, systems can be interpreted as engaging in active inference in order to minimize their free energy. A system can be described to engage in active inference in the sense of performing belief updating and acting such as to fulfil prior preferences about observations. Describing a self-organizing system in terms of active inference means that the system acts upon its external milieu to maintain itself in its preferred states (cf. homoeostasis). Active inference provides a mechanism to derive the dynamics of sensory and active states such that they minimize a variational free energy functional. This allows us to describe an agent as engaging in actions that will get them closer to their preferred sensory states. Belief updates, in turn, contribute to the optimization of internal states, which tightens the (free energy) bound on surprisal, thus enabling action to avoid (statistically) “surprising” sensations; and corresponds to perception.
An extension of predictive coding (and part of the free energy principle), which says that agents can suppress prediction errors by performing actions to bring about sensory states in line with predictions.
1. Computational process in which prediction error is minimized by acting on the world ("making the world more similar to the model"), as opposed to minimizing prediction error by changing the internal model, i.e. perceptual inference ("making the model more similar to the world"). 2. Also used as a generic term for the computational processes which underpin both action and perception, and, in the context of FEP, for all computational processes that minimize free energy.
The minimisation of free energy through changing internal states (perception) and sensory states by acting on the world (action).
active inference is a self-organising process of action policy selection.
Judging is a subclass of selecting
*ActiveInference is internally related to Judging.
6
Active Learning
Core
To reduce the latter type of uncertainty, agents can expose themselves to observations that complete ‘knowledge gaps’ and thereby learn the probabilistic structure of unknown and unexplored (novel) contingencies – hence active learning allowing for ‘model parameter exploration’.
Learning is a subclass of intentional psychological process
7
Active States
Core
Active states are a subset of blanket states that mediate the influence of internal states on external states. Conversely, sensory states are a subset of blanket states that mediate the influence of external states on internal states.
Which states are internal/external? Which are autonomous states?
Action vs Active states?
Partitioning of blanket states into incoming (sense) and outgoing (action) statistical dependencies.
Have outgoing statistical dependencies with external states
Have outgoing statistical dependencies towards external states
*ActiveState is a subset of PhysiologicProcess.
8
Active Vision
Core
Our vision incorporates both the information that falls on the retina and the consequences of eye movement, in particular the consequences of rapid or saccadic eye movements. We refer to this vision as active vision
*ActiveVision is a subclass of Looking. *ActiveVision is a subclass of Searching. Looking is a subclass of intentional process. Searching is a subclass of investigating.
9
Affordance
Core
Gibson defined affordances as action possibilities latent in the environment [3], objectively measurable and independent of the ability to recognize them but always in relation to the actor [4].
*Affordance is equivalent to resource. resource is an instance of case role . resource is a subrelation of patient.
10
Agency
Core
a sense of agency as a probabilistic representation of control that is distinct from the action actually emitted
a sense of agency rests upon prior beliefs about how one will behave
Intentional process is a subclass of process
*Agency is internally realted to IntentionalProcess
11
Agent
Core
Agents defined under active inference: A) sample their environment and calibrate their internal generative model to best explain sensory observations (i.e., reduce surprise) and B) perform actions under the objective of reducing their uncertainty about the environment.
Agent is a subclass of object . (relation) agent is a subrelation of involved in event . agent is an instance of case role.
Group
is
a
subclass
of
agent
The
number
1
argument
of
agent
is
an
instance
process
.
2
Agent.
12
Ambiguity
Core
expected uncertainty of observations, expected under a policy.Ambiguity can be thought of as the expected inaccuracy, where marginal likelihood is equal to accuracy minus complexity
ambiguity is the uncertainty about outcomes given the state of the world.
Thus, ambiguity is the expectation of the conditional entropy — or uncertainty about outcomes — under the current policy
Ambiguity is the loss of a precise or definitive mapping between external states of the world and observed sensory states (as quantified by entropy, denoted by H).
In this paper, we address
the issue of ambiguity of objective image quality assessment. We propose an approach
to obtain an ambiguity interval of an objective metric, within which the quality score
difference is not perceptually significant. In particular, we use the visual difference
predictor, which can consider viewing conditions that are important for visual quality perception. In order to demonstrate the usefulness of the proposed approach, we
conduct experiments with 33 state-of-the-art image quality metrics in the viewpoint of
their accuracy and ambiguity for three image quality databases. The results show that
the ambiguity intervals can be applied as an additional figure of merit when conventional performance measurement does not determine superiority between the metrics.
The effect of the viewing distance on the ambiguity interval is also shown.
Cannot describe when one thing becomes another -- e.g. forest and trees, or what makes someone rich, thresholds.
Some kind of undecidable uncertainty?
Only related to Observations~State mappings? Or other parameters? Does this have similar use as informal deployment of the term or not?
*Ambiguity is a subclass of StateOfMind .
13
Attention
Core
Inferring the level of uncertainty or precision during hierarchical perception.
Here, we pursue the notion that attention is the optimum weighting of prediction error in the context of action preparation
It is argued here that attention is frequently used to refer to two very different phenomena. One relates to salience, and is fundamentally a property of action plans and epistemic affordance. The other involves the biasing of inference towards sensory channels providing precise information.
Attentional states, s(2), modulate the confidence in sensory observations, and meta-awareness states, s(3), modulate the confidence in higher-order observations.
Machine learning // Conscious or Aware attention
Regimes of attention https://www.frontiersin.org/articles/10.3389/fpsyg.2016.01090/full https://www.frontiersin.org/articles/10.3389/fpsyg.2019.00679/full how is this related to motor/visual and salience "Attention"
Can non-cultural ActInf agents have shared regimes of attention
Attention and Information?
*Attention is a subclass of IntentionalProcess.
*Attention nominalizes the attends CaseRole
14
Autopoiesis
Core
*Autopoiesis is internally related to Reproduction. Replication is a subclass of OrganismProcess.
15
Bayesian Inference
Core
Minimisation of variational free energy with respect to a probability distribution over (fictive) hidden states causing sensory states (observations). Variational Bayesian inference is approximate because it minimises a (free energy) bound approximation to surprise. When free energy equals surprise, inference is exact.
Updating a model in accordance with Bayes' rule, i.e. computing the posterior distribution: p(c|s) = p(s|c)p(c)/p(s).
*BayesianInference is a subclass of PhysiologicProcess.
16
Behavior
Core
Here we use the term behavior to mean what a plant or animal does, in the course of an individual's lifetime, in response to some event or change in its environment
*Behavior is an near synonym of BodyMotion. *Behavior is an near synonym of Process.
17
Belief
Core
The term ‘belief’ is used in the sense of ‘belief propagation’ and ‘Bayesian belief updating’, which are just ways of talking about probability distributions or densities. ‘Beliefs’ in Bayesian terms are posteriors and priors, corresponding to the probability distributions (a world of possible states) that are shaped by physically realized states (i.e., the internal states that parameterize distributions over external states). In general, although we use the term ‘beliefs’ to describe the probability densities defined over external states, it is generally recognized that these densities are not themselves the same as propositional beliefs.
In short, self-evidencing appears to require perceptual inference, in the sense that a belief is formed that approximates the probability of the causes of sensations. Note that the term “belief” is used here in the technical sense of Bayesian belief updating—not to indicate a conscious, propositional belief.
Technically, these (Bayesian) beliefs are referred to as approximate posteriors, variational densities or recognition densities. The recognition model is the inverse of a likelihood model: it is a statistical mapping from obser?vable consequences to hidden causes. This explains why forming a posterior belief is often referred to as
model inversion, where Q(s) ’ P(sjo)
*Belief is a subclass of psychological process. (instance believes PropositionalAttitude)
18
Belief updating
Core
a parameterization of posterior beliefs about the past and future that makes state estimation
Belief updating mediates inference and learning, where inference means optimising expectations about hidden states (policies and precision), while learning refers to optimising model parameters. This optimisation entails finding the sufficient statistics of posterior beliefs that minimise variational free energy.
*BeliefUpdating is a subclass of IntentionalPsychologicalProcess. IntentionalPsychologicalProcess is a subclass of IntentionalProcess.
19
Blanket States
Core
Blanket states comprise active states and sensory states. Generally, external states do not influence active states and internal states do not influence sensory states.
If the states of a system, whose dynamics can be described with random or stochastic differential equations (e.g., the Langevin equation), possess a Markov blanket, then... the conditional independence in question means that a set of (internal) states are independent of another (external) set, when conditioned upon blanket states. The internal states can then be cast as representing, in a probabilistic fashion, external states. From this, one can elaborate a physics of sentience or Bayesian mechanics that would be recognised in theoretical neuroscience and biology.
"A Markov blanket is defined in terms of conditional dependencies entailed by the joint density over some states"
Friston: Blanket states comprise active states and sensory states. Generally, external states do not influence active states and internal states do not influence sensory states.
Dave asks, If we enforce as a definition the observation that "internal states do not influence sensory states," do we exclude some examples of predictive processing internal to the CNS? E.g. a wine-taster systematically scans (differentially activates) various olefactory centers (S. Barwich, Smellosophy: "Olfactory receptors, as the interface of the olfactory system, actively structure stimulus input;" Jordan et al., “Active Sampling State Dynamically Enhances Olfactory Bulb Odor Representation," Neuron 98.
Cognitive psychologist Ulric Neisser coined the term "perceptual cycling," to describe perception as a cyclical process in the brain, suggesting that search patterns in foraging behavior filter input information. Alternating oscillation phases mirror the periodic sampling of sensory input and govern the responsiveness of particular brain regions, including their connectivity. Several neural populations are actively competing at any given time. So the brain is primed by its own mechanisms of input selectivity.
What are Markov Blankets? What is the usage in FEP?
Interface / Boundary states for systems and their environments
(Sense and Action)
Values of parameters of the (Markov) Blanket
Blanket states mediate Internal and External states
Things or "Boundaries between things"
Partitioning, how does this relate to boundaries in the real world?
Thermodynamic & Homeostatic systems. H systems do have T properties. But they are not the same thing. Two poles of the analogy. Media & Message. Memeology.
https://pubmed.ncbi.nlm.nih.gov/33607182/ "Recent characterisations of self-organising systems depend upon the presence of a 'Markov blanket': a statistical boundary that mediates the interactions between the inside and outside of a system."
What's the alternative here?
*ThermodynamicBlanketStates are PhysicalStates. *HomeostaticBlanketStates are InternalAttributes.
20
Cognition
Core
Cognition is the updating of hierarchical, probabilistic models of the world.
In active inference, cognition is viewed as an action-inference loop. The beliefs (priors) of an individual, and the expectations that derive from them, lead to action. Action impacts the environment, producing some effects. The individual senses those effects (perhaps imperfectly), and compares sensations to those that were expected based on beliefs. If they differ—if observations are surprising—then learning (updating of the individual’s generative model) might be necessary.
Cognitive agent is a subclass of sentient agent
21
Complexity
Core
complexity is the divergence between posterior and prior beliefs.
complexity is defined in relation to empirical priors based on the approximate posterior expectations of the preceding (forward) and subsequent (backward) states
*Complexity is a subclass of ObjectiveNorm.
22
Cue
Core
In turn, the deontic cue is the probability of an outcome at any given state, which depends upon concentration parameters α that the environment learns as a function of how agents act on the world, which changes the value of the concentration parameters.
A *Cue is internally related to an instance of Perception . AgentPatientProcess is a subclass of Process .
23
Culture
Core
inter-group behavioural and cognitive variations that arise through social learning within members of the same species
Culture is an umbrella term which encompasses the social behavior and norms found in human societies, as well as the knowledge, beliefs, arts, laws, customs, capabilities, and habits of the individuals in these groups.
*Culture is a subclass of Proposition .
Agreement
is
a
subclass
of
Proposition
24
Data
Core
Data are units of information, often numeric, that are collected through observation.[1] In a more technical sense, data are a set of values of qualitative or quantitative variables about one or more persons or objects,[1] while a datum (singular of data) is a single value of a single variable.[2]
sensory data is given by Bayes rule (1)
InformationMeasure is a subclass of ConstantQuantity. Stating is a subclass of LinguisticCommunication.
InformationMeasure
is
a
subclass
of
ConstantQuantity
.
*Data is a near synonym of InformationMeasure. *Data is a near synonym of FactualText. *Data is a near synonym of Stating.
25
Decision-making
Core
This paper considers decision-making and action selection as variational Bayesian inference. It tries to place heuristics in decision theory (in psychology) and expected utility theory (in economics) within the setting of embodied or active inference.
Deciding is a subclass of Selecting. Selecting is a subclass of IntentionalPsychologicalProcess.
26
Ensemble
Core
Here, we consider simulations of a primordial soup reported in [11] to illustrate the emergence of active inference of a simple and prebiotic sort. This soup comprises an ensemble of dynamical subsystems, each with its own structural and functional states, that are coupled through short-range interactions.
ensembles; groups of neurons that tend to fire in synchrony. Importantly, spontaneously active ensembles are similar to those evoked by sensory stimuli suggesting that ensembles encode features of the sensory environment and that their spontaneous activation reflects an intrinsic capacity of the brain to generate an internal model of the environment
*Ensemble is a subclass of Collection .
27
Epistemic value
Core
Epistemic value is the expected information gain under predicted outcomes. In other words, it reports the reduction in uncertainty about hidden states afforded by observations
Interestingly, (35) tells us that maximizing the epistemic value of the policy maximizes opportunity, while at the same time minimizing risk. In the EFE (20), epistemic value is related with the mutual information between states and outcomes. In the CBFE [constrained Bethe Free Energy], the epistemic value of the policy is more inclusive, because it accounts for the information opportunity as well as the risk of the policy.
“epistemic value” (e.g., information gain or the resolution of uncertainty implicit in exploration or curiosity).
Epistemic value is......
From Karl: Epistemic value is the information gain or reduction in uncertainty about latent states afforded by the outcomes of a particular policy. It is variously known as Bayesian surprise, epistemic affordance, the value of information, intrinsic motivation and so on. Mathematically, it is the KL divergence between beliefs about latent states before and after the outcomes of a policy. Epistemic value is the value of a policy that is a functional of beliefs about the causes of sensations.
(Pragmatic and Epistemic) & (Extrinsic & Intrinsic) ---> Whats the relationship, are these the same?
Intrinsic motivation involves performing a task because it’s personally rewarding to you.Extrinsic motivation involves completing a task or exhibiting a behavior because of outside causes such as avoiding punishment or receiving a reward.
The main difference between intrinsic and extrinsic motivation is that intrinsic motivation comes from within, and extrinsic motivation comes from outside. While both types of motivation are important, they have different effects on how you work.
https://www.rochester.edu/emerging-leaders/understanding-intrinsic-and-extrinsic-motivation/
https://www.tandfonline.com/doi/abs/10.1080/17588928.2015.1020053?journalCode=pcns20 Minimizing expected free energy is therefore equivalent to maximizing extrinsic value or expected utility (defined in terms of prior preferences or goals), while maximizing information gain or intrinsic value (or reducing uncertainty about the causes of valuable outcomes).
From Karl: Epistemic value is the information gain or reduction in uncertainty about latent states afforded by the outcomes of a particular policy. It is variously known as Bayesian surprise, epistemic affordance, the value of information, intrinsic motivation and so on. Mathematically, it is the KL divergence between beliefs about latent states before and after the outcomes of a policy. Epistemic value is the value of a policy that is a functional of beliefs about the causes of sensations.
*EpistemicValue is a subclass of PsychologicalProcess . *EpistemicValue is a subclass of SubjectiveAssessmentAttribute . *EpistemicValue is an instance of InternalAttribute
The abstract counterpart of an *EpistemicValue is an *AbstractEpistemicValue. *EpistemicValue is a relatedInternalConcept to Investigating.
28
Ergodicity
Core
A process is ergodic if its long term time-average converges to its ensemble average. Ergodic processes that evolve for a long time forget their initial states.
29
Evidence
Core
Bayesian model evidence is effectively simplicity plus accuracy./Negative surprise is the same as log evidence; namely, the logarithm of Bayesian model evidence.
*Evidence is internally related to IntentionalPsychologicalProcess .
30
Expected Free Energy
Core
The expected free energy is a functional of posterior beliefs about states.
*ExpectedFreeEnergy is a subclass of RelationalAttribute .
*ExpectedFreeEnergy is internally related to InformationMeasure .
31
External States
Core
Statistically, the existence of a Markov blanket means external states are conditionally independent of internal states, and vice versa, given blanket states. Generally, external states only influence themselves and sensory states.
Here, circular causality is induced by separating the states of a random dynamical system into external and internal states, where external states are subject to random fluctuations and internal states are not.
External states correspond to environmental causes that generate sensory samples (also known as sensory input, sensory outcomes, sensory data, or evidence), which affect the system’s internal state.
Realist and Instrumentalist
External/Internal/Interface
Is it important that the partitioning be this way?
Qualify & Quantify
What is the relevance of this partitioning scheme (e.g. with a blanket separating Internal/External states)
The partitions are set by us on something else observed/modeled.
From Colombo & Wright: "For free-energy theorists, the dynamics of such systems will appear to place an upper bound on their informational entropy, and to maximize the evidence for a model M of external states “entailed” by their characteristic properties. This behavior—they would conclude—can be expressed as approximate Bayesian (active) inference about the causes of sensory states in terms of minimizing variational free energy.Footnote6"
What happens to the generative model when we are asleep? Do we have a generative model at that time?
Is it the same model being used differently, but then is it a different model?
Internal & External states -- they are a partitioning, have to be separate.
*ExternalState is a subset of PhysiologicProcess.
32
Free Energy
Core
An information theory measure that bounds (is greater than) the surprise on sampling some data, given a generative model.
In the context of Friston's FEP, free energy is not a thermodynamic quantity, but an information-theoretic quantity that constitutes an upper bound on surprisal. If this bound is tight, the surprisal of sensory signals can therefore be reduced if free energy is minimized by bringing about changes in the world.
*FreeEnergy is a subclass of PhysicalDimension . *FreeEnergy is a subclass of RelationalAttribute .
*FreeEnergy is internally related to InformationMeasure .
33
Free Energy Principle
Core
https://en.wikipedia.org/wiki/Free_energy_principlehttps://en.wikipedia.org/wiki/Free_energy_principle.
"The brain aims at reducing surprise, where this surprise (or surprisal) is quantified as accuracy (expected log likelihood) minus complexity (informational divergence between the posterior probability and prior probability). This complexity is also known as Bayesian surprise (or salience), and represents the extent to which the new data is 'surprising' to the prior model."
"Systems minimise a free energy function of their internal states, which entail beliefs about hidden states in their environment. The implicit minimisation of free energy is formally related to variational Bayesian methods."
A generalization of predictive coding (PC) according to which organisms minimize an upper bound on the entropy of sensory signals (the free energy). Under specific assumptions, free energy translates to prediction error.
"A self organising system—that entails a generative model— minimises the free energy of its sensory and internal states; where internal states encode a recognition probability distribution over (fictive) hidden states causing sensory states."
*FreeEnergyPrinciple is an instance of Proposition .
34
Friston Blanket
Core
For this reason, we propose to distinguish between ‘Pearl blankets’ to refer to the standard use of Markov blankets and ‘Friston blankets’3 to refer to the new construct. While Pearl blankets are unambiguously part of the map, Friston blankets are best understood as part of the territory. Since these are different formal constructs with different metaphysical implications, the scientific credibility of Pearl blankets should not automatically be extended to Friston blankets
*FristonBlanket is a subclass of ProbabilityRelation. *FristonBlanket is a subclass of Proposition .
35
Generalized Free Energy
Core
Equation 14
*GeneralizedFreeEnergy is a subclass of ProbabilityRelation. *GeneralizedFreeEnergy is a subclass of Proposition .
36
Generative model
Core
Generative model or forward model is a probabilistic mapping from causes to observed consequences (data). It is usually specified in terms of the likelihood of getting some data given their causes (parameters of a model) and priors on the parameters.
A probabilistic model that links (hidden) causes and data, usually specified in terms of likelihoods (of observing some data given their causes) and priors (on these causes). Generative models can be used to generate predictions of data , given their causes.
The joint probability distribution of two or more random variables, often given in terms of a prior and a likelihood: p(s,c) = p(s|c)p(c). (Sometimes, only the likelihood p(s|c) is called a "generative model".) The model is generative in the sense that it models how sensory signals s are generated by hidden causes c. Furthermore, it can be used to generate mock sensory signals, given an estimate of hidden causes.
A generative model is a probabilistic mapping from causes in the environment to observed consequences (e.g., sensory data);
A formalism that describes the mapping between inferred hidden states/causes, and expected outcomes/observations
RECOGNITION models update internal parameters that correspond to external states (including hidden causes of environmental states), blanket states, and internal states (meta-modeling). In contrast, GENERATIVE models take those same internal parameter estimates and emit expected or plausible observations.
Generative model of what?
Generative model is starting concept of Active Inference -- Distinction from FEP (?) -- Dynamics and behavior are starting from GM, action as well. All talks about terms should include this.
Generative Model &/of a Generative Process
Recognition models and Generative models.
Recognition model is from empirical observations to updates of inferred hidden states. Generative model is from inferred hidden states to plausible emitted observed states. This is the "tale of two densities" because models are distributions which are statistical densities.
"Processing" is often used in a uni-directional Recognition Model type way -- "Predictive Processing" entails and requires a Prediction
What is the relationship between conscious experience & Generative models? Why are some GM experienced or not? Meta-modeling?
GM of Sense + Action
Enactivism + Predictive Processing ---> PP does deal with action. However in Maria's perspective they do not pay as much attention to the environment, e.g. they are more focused on the organismal dynamics perhaps.
Generative Modeling is the key for perception?
Deep GM / Deep Inference.
Where is the body in GM?
Broadest most-applicable definition ----> Then we specify Computational, Embodied, Enacted.
*GenerativeModel is a subclass of Process.
37
Generative Process
Core
"a generative process 𝑝𝐺𝑃(𝑦, 𝑥𝐺𝑃) [captures] the actual causal structure of the environment where hidden states 𝑥 generate observations"
*GenerativeProcess is a subclass of ProbabilityRelation. *GenerativeProcess is a subclass of Proposition . *GenerativeProcess is a subclass of Process .
38
Hidden state
Core
Hidden states Ψ : Ψ × A × Ω → ℝ... constitute the dynamics of states of the world that cause sensory states and depend on action.
*AbstractHiddenState is a subclass of ProbabilityRelation.
39
Hierarchical Model
Core
Predictive Processing posits a hierarchy of estimators, which operate at different spatio-temporal timescales (so they track features at different scales). The hierarchy does not necessarily have a top level (but it might have a center — think of the levels as rings on a disc or a sphere).
*HierarchicalModel is a subclass of ProbabilityRelation. *HierarchicalModel is a subclass of Proposition . *HierarchicalModel is a subclass of Process .
40
Inference
Core
inference means optimising expectations about hidden states (policies and precision),
Inference in the generative model lies in finding posterior p(s|o) — the probability that the fruit is an apple (or orange) if it lies at a specific location.
*AbstractInference is a subclass of Learning. (This looks wrong. Abstract classes are non-temporal, and Learning changes across time.)
41
Information
Core
Information can be thought of as the resolution of uncertainty; it answers the question of "What an entity is" and thus defines both its essence and the nature of its characteristics. The concept of information has different meanings in different contexts.[1] Thus the concept becomes synonymous to notions of constraint, communication, control, data, form, education, knowledge, meaning, understanding, mental stimuli, pattern, perception, proposition, representation, and entropy.
This notion, known as Bayesian surprise, conceptualises a unit of surprise – a “wow” – in terms of the difference between the prior and posterior beliefs about the world. This allows us to formulate epistemic foraging in terms of the mutual information between an observation, and the unobservable (hidden) states of the world that give rise to it.
*Information is internally related to InformationMeasure .
42
Information Geometry
Core
The central idea that underwrites information geometry [9] is that we can define a space of parameters (a statistical manifold), where each point in that space corresponds to a probability density (e.g. the expectation and variance of a normal density).
One can take this metric treatment further and equip spaces of the sufficient statistics (i.e., parameters) of a density with an information geometry. In brief, information geometry rests on Riemannian metrics that can be used to measure distances on statistical manifolds (Amari, 1998; Ay, 2015)
43
Internal States
Core
Internal states R : R × S × Ω → ℝ... constitute the dynamics of states of the agent that cause action and depend on sensory states.
A system has a boundary that separates it from its environmental niche. It has internal processes (its parts interact). A system behaves. As it does so, there are changes to its internal state.
A common rhetoric used to unpack this is that the blanket states of a given internal state are the parents (things that cause it), children (things that it causes), and parents of its children. The parents of internal states are the sensory states that mediate the influence of the outside world, and their children are the active states that mediate their influence on the outside world.
The existence of a Markov blanket means that internal states will appear to minimize a free energy functional of the states of their Markov blanket.
Since the internal states of the Markov blanket are those states that constitute the system, we can think of the extended phenotype of the organism as literally embodying or encoding information that parameterises a recognition density.
What is an Internal state?
From Demakas et al 2020: "Imagine that every single state of being has a position in an abstract state space. There are 4 important kinds of states (i.e., dimensions) in this space: sensory states (e.g., the sound of a voice), active states (e.g., listening1), internal states (e.g., thoughts, feelings), and environmental states (e.g., location, context). These states are by definition the partitions afforded by Markov blanket."
Is this the only possible partitioning? Only one possible or only one for ActInf? ActInf with other partitionings?
Do partitions align with natural aspects/features/"systems" in the world?
Relationship with the topology of the action loop?
What are Internal States? Nesting of internal states depending on regime of attention, scale, Homeostatic, Cognitive
Utility of separating formal terms and definitions/notions -- Markov Blanket Action: is ...., separation from applications in domains (cell, psychology, social, computational)
Starting with the purely statistical Markov, Pearl, Friston, Beyond --> applications across domains.
Is enactivism realism?
Computational systems to test space of possible loops/partitions since empirical measurements are not always directly suggestive of particular partition
Tale of Two Densities: "The ‘causal bite’ of the generative model comes from the fact that it plays a role in policy selection by inducing free energy gradients (which then guide changes to beliefs about action). In other words, generative models are normative models of ‘what ought to be the case, given the kind of creature that I am’– they are realised physically through adaptive, belief-guided, normative actions that maintain the creature in its phenotypic states."
Inconstant or incompatible use of realism/instrumentalism, action-perception loops, interpretations of the priority of blanket vs. co-equal partitition, etc....
From Karl: Statistically, the existence of a Markov blanket means external states are conditionally independent of internal states, and vice versa, Given the blanket states. Generally, internal states can only influence active states.
*InternalState is a subset of PhysiologicProcess.
44
Latent cause
Core
The generative model shown in Fig. 2 links exteroceptive,proprioceptive, and interoceptive information by jointly repre-senting their hidden or latent causes (e.g., a cause can embody theprior knowledge that the sight of a burger will produce certainvisual and gustatory sensations, that it affords a grasping action,and ultimately that it increases the level of glucose in the blood).
The simplest encoding corresponds to representing the belief with the expected value, or expectation, of a hidden or latent cause. These causes are referred to as hidden because they have to be inferred from sensory consequence
45
Living system
Core
free-energy theorists assume that any living system possesses a random dynamical attractor—a set of states towards which a dynamical system tends to evolve for a wide variety of initial conditions of the system’s state.
In this paper we have argued that the autonomous organization of living systems consists of the hierarchical assembly of Markov blankets of Markov blankets through adaptive active inference.
*Attractor is internally related to SubjectiveAssessmentAttribute.
46
Markov Blanket
Core
A Markov blanket defines the boundary between a system of interest and its environment in a statistical sense. More specifically, it provides a statistical partitioning Into internal and external states that are separated by blanket statesto the system. In this context, a Markov blanket is a set of variables through which states internal and external to a system interact.
Markov blankets were first proposed by Judea Pearl (1988) in the context of graphical models. These models express the statistical dependencies (edges) between different factors or states (nodes). Given a state of the model, its Markov blanket is the set of sufficient states of the network needed to predict that state.
First, a 'thing' is defined stipulatively in terms of a Markov blanket, such that something’s internal states are independent of its external states, when conditioned on its blanket states. Blanket states can be further partitioned into active and sensory states that are not influenced by internal and external states, respectively. This partition is not part of the definition of a Markov blanket but describes a way of characterising the blanket states.
A Markov blanket is a statement of conditional independence between internal and external states given blanket states.
(Statistical) partitioning of system of interest, from environment, by an interface or boundary. A minimal Markov blanket is known as a Markov boundary.
Who is Markov / What is Markovian?
What does "blanket" or "blanketing" mean?
https://en.wikipedia.org/wiki/Markov_property A stochastic process has the Markov property if the conditional probability distribution of future states of the process (conditional on both past and present values) depends only upon the present state; that is, given the present, the future does not depend on the past.
Representing boundaries / boundary conditions -- where we have liminalitiy. Conditional independence separates 'things' out from their environment
Node partitioning scheme (where nodes are statistical variables) -- separating into Internal, External, and Blanket states. Blanket states render the Internal and External states conditionally independent.
Starting with System of Interest -- To define separation of System and Environment, we define the boundary of the system.
Difference between Physical boundary & Statistical insulation?
Separation of system from environment requires persistent boundary / Blanket. Will depend on the scale of analysis.
Nested Markov Blankets will have different realizations.
Examples of [Internal, External, Blanket states]
Cell [Cytoplasm Internal, Environment External, Membrane = Markov Blanket]
*ThermodynamicSystem is a subclass of *System
47
Markov Decision Process
Core
A Markov decision process (MDP) is a form of probabilistic generative model, defined in a discrete state space. The latent variables of an MDP are hidden states sτ, and policies, π
Deciding is a subclass of Selecting
48
Markovian Monism
Core
Markovian monism holds that Markov blankets are the type of thing/property that constitutes both mind and matter, and in this sense, their metaphysical monism is dubbed Markovian.
49
Model Inversion
Core
Model inversion uses available experimental observations of the output to determine the set of input parameters that maximize the predictive potential of a model.
When you calculate a prior from a posterior and a likelihood, is that an example of model inversion?
If you must re-calculate specific priors - and these are priors that CANNOT be altered (the evolutionarily-cast-in-base-pairs homeostatic set points) - does something special happen? - maybe fugue, dissociation, fainting, panic, depressive paralysis, shell-shock, repression, "cognitive dissonance," delusion, hysteria?
50
Multi-scale system
Core
Multiscale modeling refers to a style of modeling in which multiple models at different scales are used simultaneously to describe a system. The different models usually focus on different scales of resolution.
51
Narrative (model)
Core
Narratives are reports of real or imagined events, which can be presented in language (verbally or textually) or through sequences of images or other symbols.
52
Niche
Core
Niche construction refers to any (implicit or explicit) modification by organisms of the (biotic or abiotic) states of the niche that they and others inhabi
niche construction “[…] refers to the activities, choices, and metabolic processes of organisms, through which they define, choose, modify, and partly create their own niches”
In cognitive science, cognitive niche construction can be viewed as a form of instrumental intelligence whereby organisms “create and maintain cause–effect models of the world as guides for prejudging which courses of action will lead to which results”
Attribute subsumes *Niche
53
Non-Equilibrium Steady State
Core
https://www.sciencedirect.com/science/article/pii/S1571064517301409?
Under appropriate conditions, any system possessing a random dynamical attractor can be shown to be formally equivalent to any system at a steady state far from equilibrium, where the system’s “characteristic” variables are within homeostatic bounds (Friston 2012; Ramstead et al. 2018, p. 2).
living systems carve out and inhabit minima in free energy landscapes, precluding the dissipation of their states over phase space. This (nonequilibrium steady-state) behaviour differentiates living states from other states, like decay and death
Technically, a steady-state requires a solution to the Fokker Planck equation (i.e., density dynamics). A nonequilibrium steady-state solution entails solenoidal (i.e., conservative or divergence free) dynamics that break detailed balance (and underwrite stochastic chaos). In other words, The dynamics of systems at nonequilibrium steady-state are not time reversible (unlike equilibrium steady states, in which the flow is entirely dissipative).
*NonEquilibriumSteadyState is a subset of Attribute.
54
Novelty
Core
when resolving uncertainty about the contingencies, the corresponding epistemic affordance becomes novelty, i.e. the opportunity to resolve uncertainty about ‘what would happen if I did that?’
hidden contingencies encoded by the param-eters of the agent’s generative model, i.e. novelty
*Novelty is a subclass of SubjectiveAssessmentAttribute. SubjectiveAssessmentAttribute is a subclass of NormativeAttribute
55
Observation
Core
We assume the agent has sensory states that register observa- tions or outcomes ̃ o , where outcomes are a function of the state of the agent’s environment, or hidden states, ̃ s . These states are called “hidden” because they are “shielded off” from internal states by observation states.
but AI proposes a twist — rewarding observations are assumed to be likely under agent's innate beliefs.
In the free energy principle (FEP), proposed by Friston, it is supposed that agents seek to minimize the “surprise”–the negative log (marginal) likelihood of observations (i.e., sensory stimuli)–given the agents’ current belief.
*Observation is internally related to CognitiveAgent.
56
Particle
Core
each particle is distinguished from other particles, in virtue of possessing a Markov blanket.
57
Perception
Core
Perception is an inference about the causes of activity in sensory pathways.
Perception is a subclass of psychological process
58
Policy
Core
Policy is defined as a sequence of actions at time τ that enable an agent to transition between hidden states. ... From definition, in active inference, a policy is simply a sequence of choices for actions through time (i.e., a sequential policy)
Policy is a subclass of Proposition.
59
Policy selection
Core
In active inference, policy selection not only requires Bayesian belief updating; it also entails the imperatives for action. In brief, actions are considered more likely if they maximize the evidence expected under the consequences of that action. Mathematically, this means selecting actions that optimize expected free energy. This expected free energy comprises different terms, such that action policy selection depends on: (i) the potential for information gain about future states of the world (i.e., epistemic value or affordance), and (ii) to potential for fulfilling preferred sensory outcomes (i.e., pragmatic value or affordance).
The endpoint of this argument is that action or policy selection becomes a form of Bayesian model selection, where the evidence for a particular policy becomes the free energy expected in the future.
60
Posterior
Core
After specifying priors, Bayesian inference – also called estimation or inversion – furnishes a posterior probability over the parameters.These posteriors are related to the priors and likelihood by Bayes rule.
We can obtain the answer from the probabilistic model p0 by doing Bayesian inference, yielding the Bayes’ posterior
*Posterior is a subclass of Proposition.
61
Pragmatic value
Core
the expected value of a policy defined in terms of outcomes that are preferred a priori, where the equivalent cost corresponds to prior surprise.
our beliefs about states of the world are as close as possible to the true state of affairs, given the sensory evidence at hand. Based on these beliefs, we can then form beliefs about “what to do” by choosing those actions that minimize expected free energy. The important move here is to separate the divergence and evidence parts of free energy and understand what their expected values mean. It turns out that they correspond to epistemic and pragmatic (i.e., motivational) value respectively. This is remarkable because exactly the same separation emerges from the treatment of cognitive consistency, namely, into epistemic and motivational value. Furthermore, these two components appear to underpin nonspecific and specific closure. In other words, the maximization of epistemic value offers a formal description of nonspecific closure, whereas the maximization of motivational (pragmatic) value corresponds to specific closure
Finally, the pragmatic value just is the value of a policy with respect to its potential of fulfilling preferred outcomes (i.e., potential for supplying expected sensory states).
In active inference, behaviour has explorative (epistemic) and exploitative (pragmatic) aspects that are sensitive to ambiguity and risk respectively, where epistemic (ambiguity-resolving) behaviour enables pragmatic (reward-seeking) behaviour and the subsequent emergence of habits.
Pragmatic value is the benefit to an organism of a given policy or action, measured in terms of probability of a policy leading to a characteristic outcome (for the agent in question)
Pragmatic value describes the extent to which a given action is aligned with rewarding preferences over sensory outcomes.
What are the connections between Pragmatic/Epistemic Value and Affordances?
Genera