Architecture of the mind

The architecture of the mind is its overall organisation: what kinds of representations it uses, what processes operate on them, and how they are arranged. Four families of account compete. The classical view holds that the mind literally is a computing system that manipulates symbols (Rescorla, 2024). Connectionism holds that information is stored in the connection weights of networks of simple units, not as symbols (Garson & Buckner, 2019). Embodied, dynamical and predictive accounts tie cognition to the body and to time, and describe cognitive agents as dynamical systems rather than digital computers (Nguyen et al., 2024; Allen & Friston, 2016). Bayesian models describe cognition as optimal probabilistic inference, a modelling style that is itself contested (Jones & Love, 2011). Each family has a published critique, and the behavioural evidence that decides between them is thin. Unified cognitive architectures, such as Soar and ACT-R, attempt to build a single working model of the whole mind; the most recent proposal, the “Standard Model of Cognition”, is a consensus statement rather than a tested theory (Laird et al., 2017).

History

Not covered by the research: the cognitive revolution, the rise of connectionism in the 1980s and of embodied cognition in the 1990s.

Newell’s programme of unified theories of cognition, realised in the Soar architecture, sought a single system that would account for all of cognition. In 1995, Cooper and Shallice concluded that Soar fails as a unified theory, on methodological and empirical grounds (Cooper & Shallice, 1995). They did not rule out unified theories, but said they need rigorous empirical testing (Cooper & Shallice, 1995). The dynamical hypothesis, set out by van Gelder, positioned itself against the computational view; one commentary argued that van Gelder overstated how dominant the computational view was and how neglected the dynamical one (O'Brien, 1998).

Core concepts

Classical computation

The classical computational theory of mind does not hold merely that the mind is like a computer; it holds that the mind literally is a computing system (Rescorla, 2024). Fodor and Pylyshyn argued that this view explains the systematicity of thought, which classical models get for free (Garson & Buckner, 2019).

Connectionism

Connectionists store information non-symbolically in connection weights (Garson & Buckner, 2019). Hybrid views hold that symbolic processing in the brain may be a matter of degree (Garson & Buckner, 2019). The approach and its critiques are covered in Connectionism.

Embodied, dynamical and predictive accounts

Dynamical accounts hold that cognitive agents are dynamical systems, not digital computers; a 2024 review reports that the hypothesis has developed theoretically, methodologically and empirically since 1998 (Nguyen et al., 2024). Within predictive-processing theory, accounts range from cognitivist, through strongly connectionist, to dynamical and non-representational views (Allen & Friston, 2016). See Embodied cognition.

Bayesian models

Rational Bayesian models treat cognition as inference that combines prior knowledge with evidence. They are criticised as unconstrained by data about mental processes and as often recapitulating existing mechanistic theories (Jones & Love, 2011). See Bayesian models of cognition.

Unified cognitive architectures

A cognitive architecture is a fixed set of mechanisms intended to underlie all cognition, implemented as a computer program that can perform tasks. Laird, Lebiere and Rosenbloom proposed the Standard Model of Cognition as a community consensus drawn from three architectures, ACT-R, Sigma and Soar. It is not itself an architecture, and its authors acknowledge disagreements and gaps (Laird et al., 2017). Stocco and colleagues, the model’s authors among them, report that an architecture based on it fitted brain-connectivity data from 200 participants (measured with functional MRI) better than six alternative architectures (Stocco et al., 2021).

Questions, controversies, and new developments

Each account and its critique

Can architectures be tested?

Cognitive architectures are criticised as not directly falsifiable. Newell appealed to Lakatos’s notion of research programmes, which are judged by their progress over time rather than by single tests, and Cooper argues that the lack of direct falsifiability need not undermine architectures (Cooper, 2007). Critics of the Standard Model point to the representational limits of current architectures, whose knowledge bases are small and uniform (Chella et al., 2018). The fMRI test of the Standard Model described above was run by its proponents (Stocco et al., 2021).

Rules versus connections

The English past tense became the central empirical battleground between symbolic rules and connectionist associations, in the 2002 exchange between Pinker and Ullman and McClelland and Patterson (Yoon, 2023). The evidence points both ways and is set out in Connectionism and Language. The dynamical account of infants’ A-not-B error has been tested mainly by its own proponents, and competes with other explanations of the same data (Spencer et al., 2001).

Broader connections

Which level of explanation an account works at, in Marr’s sense, shapes what evidence can test it; see Methods of cognitive psychology. Dual-process theory is another cross-cutting account, concerned with types of processing rather than with the format of representations.

Further reading

• Rescorla (2024): the classical computational theory of mind.
• Garson and Buckner (2019): connectionism and the systematicity debate.
• Laird et al. (2017): the Standard Model of Cognition.
• Cooper and Shallice (1995): the case against Soar as a unified theory.