Trends in Cognitive Sciences
OpinionGetting ahead: forward models and their place in cognitive architecture
Section snippets
Two roles for forward models
There is a great deal of evidence that people predict both themselves and other people 1, 2. But how do they do it? Recent proposals suggest that people use forward models 3, 4, 5 to make predictions (see Glossary). However, there are two different types of account of how they might be used. The first account (Figure 1) assumes a dedicated prediction mechanism implemented by additional circuitry distinct from the core mechanisms of perception and action. Those core mechanisms involve one or
AFMs
When I plan an action, for instance moving my arm to a target, I construct an action command, use that command to perform the action, and experience the sensory (including proprioceptive) consequences of that action. If I repeatedly perform the action, I can learn from my mistakes (e.g., changing the plan slightly if my arm just misses the target). Over time, I can predict that if I instigate an action command, I will subsequently experience a particular result. In the same way, if I decide to
IFMs
The alternative IFM approach originates in work on the role of prediction in perception 10, 11, 12, 13, 14, 15, 16. In these accounts, perception itself involves the use of a forward (generative) model whose role is to construct the incoming sensory signal ‘from the top down’. Mismatches between the predictions issued by the forward model and the sensory flow result in ‘prediction error’ signals that refine and alter the predictions, until the system settles into a coherent multilevel state.
Comparisons between the accounts
AFMs and IFMs both implicate forward models in the production of fluent motor action. Both invoke prediction error-minimising schemes that are either identical with or formally related to familiar schemes such as Kalman filtering (see [25]; for a review and some formal comparison, see [26]). The key difference between the accounts lies in the need (or lack of it) to explicitly compute an inverse model. According to the AFM account, the forward model is distinct from the inverse model, because
Testing grounds
The AFM account, we have seen, invokes two distinct models: an inverse model that converts intentions into motor commands and a separate forward model that converts motor commands into sensory consequences. IFMs, by contrast, automatically invert a single forward or generative model that converts intentions into sensory consequences. At the most abstract level, AFMs thus depict movements as driven by descending motor commands and simulations as handled by a further (efference copy-driven)
Trading complexity
Despite their many similarities, the IFM and AFM accounts represent fundamentally different views of the shape and functioning of the human cognitive architecture. In common is the core emphasis on the need to predict our own upcoming sensory states. Such predictions can power learning, help finesse time delays, and enable a suite of potent capacities for motor imagination and simulation-based reasoning. In common too is the resulting emphasis on the learning and use of a forward (generative)
Glossary
- Active inference
- the combined mechanism by which perceptual and motor systems conspire to reduce prediction error using the twin strategies of altering predictions to fit the world and altering the world (including the body) to fit the predictions.
- Corollary discharge
- often (incorrectly) used synonymously with ‘efference copy’, this names the output of the forward model (the predictor mechanism), which may be used to influence further processing.
- Efference copy
- a copy of the current motor command
References (41)
- et al.
Forward models: supervised learning with a distal teacher
Cogn. Sci.
(1992) - et al.
Forward modeling allows feedback control for fast reaching movements
Trends Cogn. Sci.
(2000) Internal models for motor control and trajectory planning
Curr. Opin. Neurobiol.
(1999)- et al.
Multiple paired forward and inverse models for motor control
Neural Netw.
(1998) - et al.
Computational mechanisms of sensorimotor control
Neuron
(2011) Hierarchical mosaic for movement generation
Int. Congr. Ser.
(2003)Canonical microcircuits for predictive coding
Neuron
(2012)The free-energy principle: a rough guide to the brain?
Trends Cogn. Sci.
(2009)What is optimal about motor control?
Neuron
(2011)- et al.
Incremental interpretation at verbs: restricting the domain of subsequent reference
Cognition
(1999)
Sensitivity to syntax in visual cortex
Cognition
Tongue corticospinal modulation during attended verbal stimuli: priming and coarticulation effects
Neuropsychologia
M1 corticospinal mirror neurons and their role in movement suppression during action observation
Curr. Biol.
Learning to predict the future: the cerebellum adapts feedforward movement control
Curr. Opin. Neurobiol.
The cerebellum optimizes perceptual predictions about external sensory events
Curr. Biol.
Cerebellar rTMS disrupts predictive language processing
Curr. Biol.
An integrated theory of language production and comprehension
Behav. Brain Sci.
The case for motor involvement in perceiving conspecifics
Pscyhol. Bull.
The emulation theory of representation: motor control, imagery, and perception
Behav. Brain Sci.
Perceptions as hypotheses
Philos. Trans. R. Soc. Lond. B: Biol. Sci.
Cited by (137)
Schematic processing and emotional change: Implications for treatment
2024, New Ideas in PsychologyEmotion regulation from an action-control perspective
2023, Neuroscience and Biobehavioral ReviewsMotor invariants in action execution and perception
2023, Physics of Life ReviewsMeaning makes touch affective
2022, Current Opinion in Behavioral Sciences
- *
The authors contributed equally to this work.