[Paper Review] Predictions in the eye of the beholder: an active inference account of Watt governors

Manuel Baltieri, Christopher L. Buckley+1 more|arXiv (Cornell University)|Jun 20, 2020

Embodied and Extended CognitionNeuroscience60 references30 citations

TL;DR

This paper formulates the Watt centrifugal governor as an active inference system by deriving a generative model that frames its behavior as minimizing prediction error through variational free energy minimization. It challenges the notion of generative models as internal representations, arguing instead they are epistemic tools for observers, and demonstrates that active inference provides a mathematically coherent, instrumental framework for modeling cognition without committing to strong representationalist metaphysics.

ABSTRACT

Active inference introduces a theory describing action-perception loops via the minimisation of variational (and expected) free energy or, under simplifying assumptions, (weighted) prediction error. Recently, active inference has been proposed as part of a new and unifying framework in the cognitive sciences: predictive processing. Predictive processing is often associated with traditional computational theories of the mind, strongly relying on internal representations presented in the form of generative models thought to explain different functions of living and cognitive systems. In this work, we introduce an active inference formulation of the Watt centrifugal governor, a system often portrayed as the canonical "anti-representational" metaphor for cognition. We identify a generative model of a steam engine for the governor, and derive a set of equations describing "perception" and "action" processes as a form of prediction error minimisation. In doing so, we firstly challenge the idea of generative models as explicit internal representations for cognitive systems, suggesting that such models serve only as implicit descriptions for an observer. Secondly, we consider current proposals of predictive processing as a theory of cognition, focusing on some of its potential shortcomings and in particular on the idea that virtually any system admits a description in terms of prediction error minimisation, suggesting that this theory may offer limited explanatory power for cognitive systems. Finally, as a silver lining we emphasise the instrumental role this framework can nonetheless play as a mathematical tool for modelling cognitive architectures interpreted in terms of Bayesian (active) inference.

Motivation & Objective

• To reframe the Watt governor—a canonical anti-representational system—as an active inference process using variational free energy minimization.
• To challenge the assumption that generative models in predictive processing must be internal, explicit representations within cognitive systems.
• To assess the explanatory power of active inference and predictive processing as a unifying theory of cognition.
• To demonstrate that active inference can serve as a formal, instrumental framework for modeling sensorimotor loops without ontological commitments to internal representations.
• To reconcile active inference with dynamical systems theory by showing how feedback regulation in the governor emerges from inference principles.

Proposed method

• Formulated a linear probabilistic generative model of the Watt governor from the perspective of an external observer, defining hidden states (e.g., flyball angle, angular velocity) and observations (e.g., valve position, engine speed).
• Derived a cost functional based on variational free energy, representing prediction error between sensory inputs and generative model predictions.
• Applied the assumption of overdamped dynamics to simplify the system and derive equations equivalent to the linearized dynamical model of the governor near equilibrium.
• Used Gaussian and Markovian assumptions to enable approximate Bayesian inference and action selection via free energy minimization.
• Connected the resulting equations to standard control-theoretic formulations of the governor, showing consistency with classical dynamical systems analysis.
• Reinterpreted the role of the generative model not as a mental representation, but as a descriptive, epistemic tool for modeling uncertainty and system properties.

Experimental results

Research questions

• RQ1 Can the Watt governor—a system often cited as anti-representational—be formally described using active inference and variational free energy minimization?
• RQ2 To what extent do generative models in active inference function as internal representations, or are they merely epistemic tools for observers?
• RQ3 Does the active inference framework offer genuine explanatory power for cognitive systems, or is it trivially applicable to any system?
• RQ4 How does the active inference formulation of the governor compare to traditional dynamical systems analysis in terms of descriptive adequacy?
• RQ5 Can active inference serve as a unifying mathematical framework for cognition without requiring strong metaphysical commitments to internal representations?

Key findings

• The active inference formulation successfully reproduces the linearized equations of motion for the Watt governor under standard assumptions, validating its consistency with classical control theory.
• The derivation shows that the system's behavior—regulating engine speed via negative feedback—emerges naturally from minimizing prediction error, without requiring explicit computation or offline inference.
• Generative models in this framework are not internal representations of the system but epistemic constructs used by an observer to model uncertainty and system dynamics.
• The paper identifies a 'mind projection fallacy' in assuming that generative models represent real internal states in cognitive systems, arguing instead that such models are observer-relative descriptions.
• Active inference provides a mathematically coherent and instrumentally useful framework for modeling cognition, even when the system itself lacks internal representations.
• The framework reconciles predictive processing with embodied and dynamical approaches by showing that inference and action can be unified under a single variational free energy minimization principle.