Overview
This article introduces a mathematical model for the Motivated Emotional Mind (MEM) cognitive architecture, designed for implementation in embodied intelligent systems. The model posits that such systems learn to maintain internal homeostasis through a generalized form of reinforcement learning, which is termed 'motivated learning' (ML). A core contribution is the rigorous formalization of a re-entrant loop, which integrates feedforward processing, lateral interactions, and feedback pathways. This loop, along with associated representational selection mechanisms, is posited to govern the adaptive responses of the system. The framework details how various signals—ongoing exteroceptive and interoceptive signals, the current bodily-motivational context, and memory traces—are integrated into associative memory structures known as 'semblions.' These semblions then engage in competition for access to subsequent processing stages and for top-down reconstruction. The formalization extends to secondary perception, processes of representational competition, curiosity, the identification of procedural gaps, and the selection of actions specifically directed toward limiting violations of allostasis.
Research Context
The research is situated within the development of cognitive architectures for embodied intelligent systems. It addresses the challenge of creating systems that can maintain internal homeostasis, a fundamental aspect for biological intelligence. The proposed motivated learning paradigm is presented as a generalized form of reinforcement learning, distinguishing it from standard reinforcement learning models by explicitly incorporating factors relevant to embodied agents. The model's formalization of cognitive phenomena aims to provide a basis for further theoretical analysis, computer simulation, and practical implementation within artificial intelligence systems that draw inspiration from biological processes.
Approach
The approach involves developing a mathematical model to formalize the Motivated Emotional Mind cognitive architecture. This formalization rigorously defines the components and interactions within the system. Key elements of the model's formalization include:
- Re-entrant Loop: Integration of feedforward processing, lateral interactions, and feedback pathways.
- Representational Selection Mechanisms: Governing adaptive system responses.
- Semblion Formation: Binding of exteroceptive signals, interoceptive signals, bodily-motivational context, and memory traces into associative memory structures.
- Cognitive Phenomena: Formalization encompasses secondary perception, representational competition, curiosity, procedural gaps, and action selection.
- Allostatic Regulation: Action selection is directed toward limiting allostatic violations.
- Motivated Learning (ML): Tailored for embodied systems, with dynamics shaped by needs, affect, and regulatory state.
- Model Incorporations: Unlike standard reinforcement learning models, this approach integrates need thresholds, goal generation, shifting goals, bodily state, resource constraints, and action uncertainty.
- Global Affect: Functions as a central control signal, modulating the learning rate, representational valence, and the balance between exploration and exploitation.
Findings
The mathematical model describes a cognitive architecture where a generalized reinforcement learning approach, termed motivated learning, enables embodied intelligent systems to maintain homeostasis. The rigorous formalization delineates a re-entrant loop that integrates feedforward processing, lateral interactions, and feedback pathways, alongside representational selection mechanisms that govern adaptive system responses. Exteroceptive and interoceptive signals, bodily-motivational context, and memory traces are bound into associative memory structures called semblions, which then compete for further processing and top-down reconstruction. The model specifically formalizes secondary perception, representational competition, curiosity, procedural gaps, and action selection mechanisms designed to limit allostatic violations. Within this framework, motivated learning is adapted for embodied systems, with their dynamics influenced by needs, affect, and current regulatory state. The model differentiates itself from standard reinforcement learning models by incorporating need thresholds, mechanisms for goal generation and shifting goals, consideration of the bodily state, resource constraints, and explicit representation of action uncertainty. This allows for a more adequate account of response selection when under regulatory pressure. Global affect is described as a central control signal, which functions by modulating the learning rate, the valence assigned to representations, and the balance between exploration and exploitation within the system.
Why This Matters
The model provides a rigorous formalization of complex cognitive phenomena, offering a structured basis for understanding how motivated emotional processes could operate in intelligent systems. This formalization supports further theoretical analysis, computer simulation, and practical implementation in artificial intelligence inspired by biological processes.