Active Inference: Framework, Model and Pattern

Introduction

This paper introduces and explores the concept of active inference as a universal framework in cognitive science and neuroscience. Active inference provides a structured approach to understanding intelligent behavior, emphasizing principles such as free energy minimization, Bayesian inference, and the interplay between action and perception. Within this framework, models serve as internal representations that guide perception, learning, and decision-making processes. The paper also discusses patterns emerging from successful applications of active inference models across various domains, highlighting their effectiveness in understanding phenomena ranging from emotion to psychopathology.

Framework

A framework serves as a structured and organized system that provides a conceptual skeleton for understanding and solving problems within a specific domain. It encompasses principles, rules, guidelines, and predefined components, providing a foundation for development, decision-making, and research activities. Thus, when confronted with unclear phenomena, the framework assists in understanding the boundaries and elements involved. In the context of active inference, a framework lays out the fundamental components and rules. Active inference, as a theoretical framework in cognitive science and neuroscience, is built upon a set of guiding principles and rules. These include free energy minimization, which posits that intelligent systems seek to minimize the difference between their internal model’s predictions and sensory input to avoid surprise or prediction error. Additionally, active inference emphasizes learning, Bayesian inference, action, and perception as integral processes through which agents update their internal models and interact with their environment. The framework also incorporates the notion of generative models, which enable agents to create internal representations of the environment based on hidden states and observations, further facilitating perception and decision-making. Moreover, active inference operates within the Markov space, where agents make decisions based on the current state of the environment and transition probabilities between states, contributing to the adaptive and dynamic nature of intelligent behavior within the framework. Overall, the active inference framework represents an agent to model the dynamics of perception, learning, and action within complex environments.

Active inference agents make a paradigm shift towards understanding cognition as an active process of minimizing surprise. The central thesis posits that such agents strive to reduce discrepancies between expected and actual sensory observations, a concept quantified as ‘surprise’. These preferred sensory states necessitate more than passive observation; active engagement with the environment is essential. Through adaptively controlling their action-perception cycles, agents aim to seek sensory experiences that align with their anticipatory models, thereby maintaining a state of minimized surprise [1, 2].

Furthermore, these agents are often conceptualized as possessing generative models, enabling them to predict the sensory consequences of their actions. This foresight is linked to the minimization of free energy, a concept associating surprise reduction with the avoidance of uncertain states. Underpinning this approach is a Bayesian inference, where agents, delineated by a Markov blanket, conduct a form of probabilistic representation and inference of the world’s states. Internal states of the agent reflect (on average) a Bayesian inference of external realities, facilitated by a generative model of sensory input. This perspective illustrates a symbiotic relationship between an agent’s internal cognitive processes and its external actions, promoting a view of cognition as fundamentally entwined with the dynamics of interaction with the environment [1, 2].

Model

A model is a structured representation or abstraction of a system, phenomenon, or process, designed to clarify the connections between its constituent parts. It acts as a tool for simulating, analyzing, or manipulating the system to achieve specific objectives. In the context of active inference, a model refers to an internal representation or belief system that an agent constructs to understand and interact with its environment. It encompasses various aspects, including the current state of the environment, predictions about future states, and the consequences of different actions. Models within active inference serve as the foundation for perception, learning, and decision making, enabling agents to navigate their surroundings effectively.
Indeed, as mentioned earlier, a model represents the interconnection of components identified within the framework. In Paper [2], four distinct models were presented, each comprising various components that contribute to different aspects of perception. These components form the building blocks of diverse models tailored to address specific facets of perception within the active inference framework. In Fig. 1, we present one of these shapes: a dynamic precision model. It’s a flexible policy selection within the active inference framework. This model incorporates several key components, including states (s), observations (o), likelihood (A), initial state priors (D), transition matrices (B), policies (π), expected free energy (G), expected free energy precision (), prior over policies (habits E), and prior distribution over observations (C​) encodes the agent’s preferences for different outcomes.

These models have been utilized in various papers, including the Model of Emotion Conceptualization [3], Model of Emotional Awareness [4], Computational Model of Emotional Valence [5], and Model for Concept Learning [6].

Pattern

A pattern can indeed refer to a successful model or solution that has been implemented and proven effective in a particular domain or context. When a model consistently achieves its intended goals and demonstrates its value, it may be formalized into a pattern. These patterns serve as reusable templates or guidelines for others facing similar challenges or seeking to achieve similar objectives. A pattern in the realm of active inference refers to a recurrent strategy or approach derived from successful applications of active inference models. It signifies a proven and effective solution that consistently achieves its intended goals within the framework of active inference. As models within active inference evolve and demonstrate their value through practical implementation, certain patterns emerge, encapsulating the underlying principles and strategies that contribute to their success. These patterns serve as reusable templates or guidelines for others encountering similar challenges or aiming to attain similar objectives within the active inference paradigm. Examples of patterns are found in various papers. For instance, active inference has been utilized to understand migraine [7] , interoceptive psychopathology [8], control and motivation [9], as well as a range of psychopathological conditions. These conditions include symptoms related to perception, schizophrenia, psychosis, hysteria, among others  [10].

References

1. Parr, T., G. Pezzulo, and K.J. Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. 2022: The MIT Press.
2. Smith, R., K.J. Friston, and C.J. Whyte, A step-by-step tutorial on active inference and its application to empirical data. Journal of Mathematical Psychology, 2022. 107: p. 102632.
3. Smith, R., T. Parr, and K.J. Friston, Simulating emotions: An active inference model of emotional state inference and emotion concept learning. Frontiers in psychology, 2019. 10: p. 489395.
4. Smith, R., et al., Neurocomputational mechanisms underlying emotional awareness: insights afforded by deep active inference and their potential clinical relevance. Neuroscience & Biobehavioral Reviews, 2019. 107: p. 473-491.
5. Hesp, C., et al., Deeply Felt Affect: The Emergence of Valence in Deep Active Inference. Neural Computation, 2021. 33(2): p. 398-446.
6. Smith, R., et al., An active inference approach to modeling structure learning: concept learning as an example case. Frontiers in computational neuroscience, 2020. 14: p. 41.
7. Sedley, W., et al., Migraine as an allostatic reset triggered by unresolved interoceptive prediction errors. Neuroscience & Biobehavioral Reviews, 2024. 157: p. 105536.
8. Paulus, M.P., J.S. Feinstein, and S.S. Khalsa, An active inference approach to interoceptive psychopathology. Annual review of clinical psychology, 2019. 15: p. 97-122.
9. Pezzulo, G., F. Rigoli, and K.J. Friston, Hierarchical active inference: a theory of motivated control. Trends in cognitive sciences, 2018. 22(4): p. 294-306.
10. Barca, L. and G. Pezzulo, Keep your interoceptive streams under control: An active inference perspective on anorexia nervosa. Cognitive, Affective, & Behavioral Neuroscience, 2020. 20: p. 427-440.