**4.1 Active inference: from Aristotle to Friston**

The opening line in Aristotle's *Metaphysics* states that "humans desire to understand" [48]. Lack of understanding engenders puzzlement, and failure to identify causes leads to undesirable self-evaluation

*".. men of experience know that thing is so, but do not know why, while the others know the 'why' and the cause…. and man who is puzzled and wonders thinks himself ignorant" (Aristotle, Metaphysics).*

In a penetrating insight, Aristotle captures relations between experiences, surprise (puzzlement) and self-directed activities motivated by the desire to reach beyond the appearances (identify causes). Arguably, principles of active inference and variational free energy minimization advanced in [13] are congruent with those early insights. The principles assert that life in all its forms, from unicellular organisms to humans, is predicated on the organisms ability to use sensing to predict conditions in its environment and to conduct activities reducing the difference between the predicted and the actual experiences. Predictions require models of the environment, the variational free energy value determines, roughly, the (information-theoretic) distance between the current and the desired states that takes into account the difference between the predicted conditions and those that were actually sensed and the surprise experienced under the model (the smaller the probability assigned by the model to the condition, the higher the surprise).

Emphasis on activities directed at minimizing variational free energy underlies the notion of 'active inference,' which is best appreciated if contrasted to the idea of 'passive' inference expressed in Plato's allegory of the cave, as follows. Prisoners are chained to the floor inside a cave where they can see nothing of the outside world except shadows on the wall they are facing. The message is that people are caged inside their minds, senses are the only window into the world, and that window can be distorting. The allegory defines passive inference: prisoners can make guesses about the outside world but have no means to validate them or to use in any fashion.

Active inference differs from passive inference in that it incorporates iterative actions on both the outside world and the model of that world that can lead to progressively improving guesses. Understanding involves a form of model manipulation that is best defined within the active inference theory through the notion of a Markov Blanket - the third conceptual pillar in the theory integrating ideas about emergence of life, evolution, and brain operation into a seamless whole.

'Markov Blanket of node x' is a graph-theoretic term denoting a set of nodes in a directed graph that are connected to x by links incident to and from x. More loosely, the term can be used to denote a group of nodes in subnetwork X1 separating it from the rest of the network X. If links denote some form of interaction, Markov Blanket of X1 can be viewed as an interface through which internal nodes in X1 interact with their surrounds in X. On that view, Markov Blanket accords X1 a degree of (conditional) independence from X - a critical concept in the overall theory, as follows.

The theory of life attributes emergence of life to spontaneous phase transitions in molecular networks ('primordial soup'), resulting in the formation of subnetworks that remain connected to their surrounds but acquire a degree statistical

#### **Figure 7.**

*Passive inference, active inference and comprehensive active inference. (A) The allegory of the cave (passive observation without action: sensory input is neither solicited nor acted upon. (B) Observation-action iterations guided by feedback produce (deposit) a model that adjusts subsequent iterations and gets adjusted by them. (C) Second regulatory loop manipulates structures formed by the first loop to construct models, the process is decoupled from the motor-sensory feedback.*

independence (autonomy) from it. In that context, Markov Blanket denotes interface (a 'membrane') between such quasi-autonomous formations and their environment [14]. As more complex forms of life develop, the Markov Blanket expands to incorporate the entire sensory-motor periphery, as suggested in **Figure 3**. Finally, **Figure 7** separates Markov Blanket from the nervous system to illustrate the notions of active inference and comprehensive active inference (incorporating the understanding capacity).

The process in **Figure 7A** is an idealization; **Figure 7B** depicts associative learning (e.g. the hypothetical salamander associates elevated temperature with successful hunting, entailing search for hot spots); **Figure 7C** depicts active construction of mental models that underlies understanding. Learning yields "knowledge that a thing is so", understanding defines causes.

In summary, different facets of the ideas depicted in **Figure 7** have been addressed in numerous sources in psychology, physiology, neuroscience and philosophy of the mind. The active inference framework offers a synthesis of some of the key insights in these disciplines, integrating them in a coordinated conceptual structure expressed in a unifying mathematical formalism. The central notion is that of activity: an organism is actively seeking sensory inputs, constructs models and acts on the environment. These contentions will be revisited in the discussion.

#### **4.2 Neuronal mechanisms**

The proposal in this section stems from five assumptions about the nature of neuronal processes that underlie intelligence and its special form, understanding. The proposal will be presented in three sections: first, the assumptions are formulated, along with some clarifications; next, the key points in the theory are formulated and applied to answer questions posed at the end of Section 2; finally, these key points are re-visited and related to experimental findings and other ideas in the literature.

**87**

*them.*

**Figure 8.**

features

α

*Brain Functional Architecture and Human Understanding*

*4.2.1.1 Cognition involves active deployment of neuronal resources*

(I shall return to these exciting experiments at the end of the section).

notion of "neuronal packets" that is pivotal in the theory.

asserting existence (perceiving) some bounded entity (object)

from the surrounding associative network.

*4.2.1.2 Progressively improving deployment requires relative stability of neuronal* 

Deployment strategy progresses from deploying individual neurons to deploying neuronal groups, to deploying groups of groups, etc., which requires a degree of stability in all the elements of the growing organization. This intuition entailed the

A neuronal packet is Hebb's assembly (i.e., comprises neurons connected by associative links) that is synergistic and is separated by a boundary energy barrier

It was hypothesized that packets form as a result of phase transition in associative networks, not unlike raindrops form in vapor. Accordingly, energy barrier is determined by surface tension, that is, the amount of free energy per unit surface (presumably, surface comprises cell membranes in the boundary neurons. Accordingly, surface energy is determined by the distribution of membrane potential across the surface). Neurons at the packet boundary constitute packet's Markov Blanket, surface tension in the boundary holds neurons together. Mapping these notions on the process in **Figure 4** will help appreciating its crucial consequences: first, combining neurons responding to A, B, C, D, E… in a quasi-stable bounded packet amounts to

= {A, B, C, D, E…} and, second, synergistic packets allow 'tuning' to their

*Neurons xi and xj are selected in the neuronal pool and tuned to stimulus C in the stimuli stream. Neuron xi responds to A, B, C stimuli, tuning amplifies its response to C. Sensing and motor actions are both products of active deployment (e.g. one sees color C because some neurons were selected, mobilized and tuned to C). Imagining color C involves the same process. Imagining A, or B, or C involves shifts in tuning, which can be expressed as rotating neuron's response vector. Co-firing of xi and xj establishes an associative link between* 

α

comprising

Brain is a synergistic system that selects, mobilizes and deploys (fires) neurons. Mobilization involves activities that precede firing and are centered on tuning, as

Consider the following three experiments: raising your right hand and touching your nose with the index finger, doing the same with your eyes closed, and imagining the same without doing anything. The first run involves coordination in the external space, the second involves coordination in the mental space (you know where your nose and your finger are, without reference to external coordinates), the third demonstrates coordination in the neuronal space that underlies the other two

*DOI: http://dx.doi.org/10.5772/intechopen.95594*

*4.2.1 Assumptions*

shown **Figure 8**.

*groups*

### *4.2.1 Assumptions*

*Connectivity and Functional Specialization in the Brain*

independence (autonomy) from it. In that context, Markov Blanket denotes interface (a 'membrane') between such quasi-autonomous formations and their environment [14]. As more complex forms of life develop, the Markov Blanket expands to incorporate the entire sensory-motor periphery, as suggested in **Figure 3**. Finally, **Figure 7** separates Markov Blanket from the nervous system to illustrate the notions of active inference and comprehensive active inference (incorporating the understanding

*Passive inference, active inference and comprehensive active inference. (A) The allegory of the cave (passive observation without action: sensory input is neither solicited nor acted upon. (B) Observation-action iterations guided by feedback produce (deposit) a model that adjusts subsequent iterations and gets adjusted by them. (C) Second regulatory loop manipulates structures formed by the first loop to construct models, the process is* 

The process in **Figure 7A** is an idealization; **Figure 7B** depicts associative learning (e.g. the hypothetical salamander associates elevated temperature with successful hunting, entailing search for hot spots); **Figure 7C** depicts active construction of mental models that underlies understanding. Learning yields "knowledge that a

In summary, different facets of the ideas depicted in **Figure 7** have been addressed in numerous sources in psychology, physiology, neuroscience and philosophy of the mind. The active inference framework offers a synthesis of some of the key insights in these disciplines, integrating them in a coordinated conceptual structure expressed in a unifying mathematical formalism. The central notion is that of activity: an organism is actively seeking sensory inputs, constructs models and acts on the environment. These contentions will be re-

The proposal in this section stems from five assumptions about the nature of neuronal processes that underlie intelligence and its special form, understanding. The proposal will be presented in three sections: first, the assumptions are formulated, along with some clarifications; next, the key points in the theory are formulated and applied to answer questions posed at the end of Section 2; finally, these key points are re-visited and related to experimental findings and other ideas

**86**

in the literature.

capacity).

**Figure 7.**

thing is so", understanding defines causes.

*decoupled from the motor-sensory feedback.*

visited in the discussion.

**4.2 Neuronal mechanisms**

## *4.2.1.1 Cognition involves active deployment of neuronal resources*

Brain is a synergistic system that selects, mobilizes and deploys (fires) neurons. Mobilization involves activities that precede firing and are centered on tuning, as shown **Figure 8**.

Consider the following three experiments: raising your right hand and touching your nose with the index finger, doing the same with your eyes closed, and imagining the same without doing anything. The first run involves coordination in the external space, the second involves coordination in the mental space (you know where your nose and your finger are, without reference to external coordinates), the third demonstrates coordination in the neuronal space that underlies the other two (I shall return to these exciting experiments at the end of the section).
