Free Energy Principle and Consciousness

Theories of consciousness are diverse and aim to address different aspects of the phenomenon. Some theories attempt to tackle the “hard problem” of consciousness, which is the question of why and how brain functions are accompanied by conscious experience. Other theories focus on modeling how specific types of experiential content are implemented by neural structures and dynamics. The targets of these theories range from explaining the basic awareness common to all conscious beings to the more complex consciousness found in humans, which includes higher-order cognitive processes like interoception, selfhood, first-person perspective, and theory of mind. Recent trends in consciousness studies advocate for a move towards theoretical unification. The concept of “minimal unifying models” (MUMs) of consciousness has been proposed, which are models that specify only the necessary properties of conscious experience, provide well-defined descriptions that can be further analyzed and extended, and integrate existing theories by highlighting their common assumptions. The FEP-based approaches to consciousness each focus on mapping specific aspects or processes of consciousness. Here, we start a series of posts aiming towards showing how these models, premised on the FEP, can be integrated in a meaningful way to provide a solid foundation for the study of consciousness.

Consciousness as a Cognitive Function: Cognitive/Cortical theories

Exploring consciousness as a cognitive function requires understanding how organisms interact with their environment and how this interaction shapes conscious experience. This post is a first attempt to explain consciousness is closely linked to the concept of free energy. As mentioned in previous posts of this series, the Free Energy Principle (FEP) uses the idea of Markov blankets to describe how systems distinguish and engage with their surroundings. These blankets allow systems to make educated guesses about the external world, forming the basis for cognitive processing. By understanding how the brain’s hierarchical structure processes information, we can confer that consciousness arises through mechanisms like minimizing prediction errors and focusing attention. This post examines various cognitive and cortical theories of consciousness, including the Global Workspace Theory (GWT), Global Neuronal Workspace Theory (GNWT), and the innermost screen model. These theories explain how consciousness integrates, selects, and prioritizes information across different parts of the brain.

Identifying the fundamental causal processess that govern conscious experience necessitates researching the nature of reality itself and how organisms interact with the reality of their environment. The journey through the intricacies of consciousness leads us to a fundamental realization: the level of consciousness is tied to Free energy. Drawing parallels between thermodynamic concepts and neurophysiological processes, we have already put that arousal (yet free energy), is the cornerstone of wakeful interaction with the environment. This latter is underpinned by surprisal and uncertainty. In the Free Energy Principle framework, a Markov blanket serves as a mathematical construct representing the separation and interaction between a system and its surroundings. It separates the internal states of a system, which represent the system’s core, from external states, through the intermediary of sensory and active states. This separation gives the internal dynamics a structure that can be interpreted as making inferences about the external world, in a probabilistic sense. In other words, the internal states of a system can be seen as making educated guesses about what’s happening outside of it. Markov blankets encapsulate all the information available to an external observer about the system, but does not provide a complete description due to the inherent unknowability of the states concealed behind it. This includes information about internal states as well as nested Markov blankets within the system. Each level of the brain’s neuronal hierarchy functions as a processing stage, capturing relevant information about the environment and the self. The dynamic interdependencies between these layers can be conceptualized as a Markov blanket separating them. This perspective allows us to view the flow of predictions and prediction errors as a process of interacting with nested holographic screens. Ascending prediction errors correspond to retrieving information from a stage, similar to reading from a screen, while actions on the environment involve providing information to the surroundings, similir to writing to a screen. The hierarchical organization of the brain facilitates efficient processing of vast amounts of information through a process of coarse-graining, the process in which information is processed across successive stages of the brain’s hierarch. Lower-level details are abstracted and integrated into higher-level summaries as information is processed across successive stages of the hierarchy. In terms of neurobiology, this model implies that consciousness is a trait of brains with an irreducible Markov blanket. The internal states would be the neuronal groups that connect to these modulatory cells, forming a “dynamic core” or a “global neuronal workspace.” These internal dynamics are expected to show a kind of coordinated activity with the rest of the brain and body. Consciousness can only be expressed indirectly through active states, implying that consciousness is inherently linked to agency. This becomes clearer when considering the FEP’s view of this unique Markov blanket, which appears to take an intentional stance by acting to minimize expected free energy, which includes both knowledge-seeking and goal-directed components.

The original global workspace theory (GWT) was proposed by Baars and Newman in 1994. It was later developed into biologically-grounded models like the global neuronal workspace theory (GNWT) by Dehaene, Kerszberg. GWT is one of the most popular frameworks in the neuroscience of consciousness focusing on attentional processes and their relation to consciousness. GWT understands consciousness in terms of the availability and access to information within the brain. This is sometimes characterized as “access consciousness” rather than “phenomenal consciousness”. In the classical GWT, consciousness is associated with the availability of information in a “global workspace” in the brain. This allows different brain networks to broadcast information to each other, enabling the integration of information across modalities. GWT assumes a specific scale of brain activity, modeled as a neural network with recurrently-connected levels, where local processes self-organize into a larger network. The GNWT model further develops this, explaining conscious experience as arising from “ignition events” where information is aggregated and broadcast through the global workspace, potentially involving the inhibition of competing processes. GNWT identifies the prefrontal and parietal regions as the loci of these ignition events, with conscious experience resulting from the late activation of these regions. Empirical studies using neuroimaging techniques have supported GNWT, showing widespread activation of frontal and parietal regions during conscious perception, in contrast with more limited activation for unconscious stimuli. This large-scale ignition connects higher-order conceptual areas to sensory cortices, making information more broadly available between these systems. Bayesian global workspace theories are a set of theories that reframe global workspace theory through a Bayesian lens. They explore how information is processed and integrated in the brain, particularly focusing on attentional processes and their connection to consciousness. These theories suggest that consciousness arises from the availability of information within a “global workspace” in the brain, where various brain networks can share information, leading to coherent activity and the integration of information across different brain regions and modalities.

The next set of models includes the “winning hypothesis” account of consciousness, which extends to Bayesian global workspace theories. According to the winning hypothesis approach, the framework of active inference provides a mechanism for inhibition events, potentially integrating or even superseding the classical global workspace theory (GWT) model. The winning hypothesis leverages policy selection and meta-cognitive inference to explain how the contents of conscious experience are generated. In predictive coding, the brain evaluates competing hypotheses about the probable causes of its sensory states. Bayesian global workspace theories suggest that conscious experience corresponds to this kind of inference - the hypothesis with the highest posterior probability given the current data and prior beliefs determines the contents of experience. This view considers attention and perception as distinct but related aspects of prediction error minimization. The accuracy and expected precision of perceptual hypotheses form the “statistical dimensions of conscious perception” - accurate hypotheses minimize prediction error parsimoniously, while precise hypotheses can better ignore noise and pick out relevant inputs. Attention plays a dual role in this framework, adjusting precision weightings and biasing competition among hypotheses. However, conscious perception does not necessarily require both accuracy and precision at the cognitive level - a precise but inaccurate hypothesis could still determine conscious experience, and vice versa. Predictive global workspace theory suggests consciousness is a function of the brain’s ability to reconcile sensory information with the predictions generated by its internal models. The neuronal dynamics underlying this involve forms of inference with sufficient temporal depth to provide context to incoming cues, which can then be coordinated within the global workspace to enable goal-oriented behavior.

Attention schema theory proposes that consciousness arises from the brain’s ability to create a model or schema of its own attention processes. This allows the brain to understand and predict its own actions and behaviors, as well as those of others. The basic attention schema theory proposes that consciousness is a form of subjective awareness that arises from the attention an agent pays to a given stimulus. The “attention schema” is a simplified model the brain uses to represent its attentional processes to itself, and to inform and direct those processes. According to this theory, whenever we claim to be conscious of something, we are actually using higher-order cognitive processes to introspect on this attention schema and report the information it holds. The schema must be sparse in order to provide an efficient means of controlling more complex processes. A Bayesian version of attention schema theory has also been proposed, which incorporates elements of the “multiple drafts” view. This version builds on Bayesian global workspace theories, arguing that the mere fact a hypothesis has the highest posterior probability is not sufficient to account for the contents of conscious perception. Rather, attentional processes must first reduce the space of possible causes of sensations, and then probe these hypotheses. This ties the “fame in the brain” account back to attention schema theory, providing a deflationary theory of consciousness where apparently essential features of phenomenal consciousness are illusions created by reflexive access.

According to global workspace theory (GWT) and global neuronal workspace theory (GNWT), lower levels of hierarchical processing in the brain are unlikely to be experienced consciously. This is because these lower levels encode belief structures that unfold too quickly to be represented in the way required for conscious experience. We can understand this core aspect of GNWT by considering the hierarchical, nested structure of neural inference. Theories of consciousness based on GWT allow us to model how consciousness integrates, selects, and prioritizes inputs from lower hierarchical levels of the brain (corresponding to the “mental action” mediated by the “active sector” of the inner screen). According to GWT and its Bayesian implementations, this “access dimension” of consciousness is realized through “ignition” events and the concurrent inhibition of competing processes. In terms of the free energy principle and predictive coding, this would involve the precision-weighting of predictions and prediction errors. This selection of ascending messages is crucial for understanding attention from these perspectives. The scale of conscious contents follows from the hierarchical, nested composition of Markov blankets, where active states of the inner screen intervene causally in lower processing (i.e. mental action). An intermediate level of processing is seen as crucial for realizing mental action and attentional sets, in these intermediate levels, information selected by higher levels is prioritized as a natural part of Bayesian mechanics under hierarchical generative models. GNWT often emphasizes the role of cortical hierarchies in the genesis of conscious experience, though it does not necessarily assume consciousness requires a brain with neocortices or advanced phenotypic characteristics.

The innermost screen model of consciousness proposes that actions are informed by sensory feedback that travels through various hierarchical layers of the brain, making these layers important for maintaining the higher-level functions of the brain. Essentially, the innermost screen can only understand or remember aspects of the world by influencing these subordinate levels, a process central to both agency and consciousness. Interestingly, studies of the brain’s functional connectome show that networks involved in shifting from subliminal to conscious perception are part of a highly interconnected “rich club.” These networks, which have extensive connections to motor systems and are linked with intentional control. It is is noteworthy that brainstem nuclei, which are involved in arousal and modulate neurotransmitter systems, as part of the active states of this innermost screen, specially given their crucial role in regulating consciousness levels, including during sleep.

The exploration of consciousness as a cognitive function through cognitive and cortical theories offers a novel understanding of how consciousness arises from the brain’s interaction with its environment. By integrating concepts like the Free Energy Principle, Markov blankets, and hierarchical processing, we can put the brain’s ability to minimize prediction errors and focus attention as foundational to conscious experience. The Global Workspace Theory and its neuronal counterpart (GNWT) highlight the importance of information integration and selection across the brain’s networks, while Bayesian approaches further refine our understanding by incorporating probabilistic inference. Attention schema theory and the innermost screen model add layers of complexity by examining the brain’s self-modeling and hierarchical information processing. Together, these theories underscore that consciousness is not merely a byproduct of brain activity but a dynamic, structured process intimately tied to the brain’s predictive and adaptive capabilities, bridging the gap between sensory inputs and conscious perception.

References

• The neural correlates of consciousness under the free energy principle: From computational correlates to computational explanation
• Active Inference as a Computational Framework for Consciousness
• Mark Solms’ theory of consciousness
• The inner screen model of consciousness: applying the free energy principle directly to the study of conscious experience
• The Hard Problem of Consciousness and the Free Energy Principle
• The Hard Problem of Consciousness and the Free Energy Principle
• Artificial consciousness: A perspective from the free energy principle
• Integrated world modeling theory expanded: Implications for the future of consciousness