A computational framework may successfully predict behavior without accurately describing the mechanisms that produce it. How does a scientific tool become a model of cognition, and when does that model justify a theory about how the mind actually works?

Models: Tools, Not Truths

Throughout this chapter, we have used the term “model” in several ways, referring to Bayesian inference, reinforcement learning, and probabilistic generative models as computational models of cognition.

But we have also been careful with our language. A framework may capture certain aspects of cognition; cognitive processes may operate as if they were performing probabilistic inference; and dopamine neurons may fire in a pattern functionally analogous to a reward-prediction-error signal.

The fact that processes such as perception, learning, or self-representation can be modeled mathematically does not mean that these frameworks are literally true descriptions of cognition. A formalism can serve as a scientific tool without describing the mechanism responsible for the phenomenon being studied. When a successful tool is reinterpreted as a theory of the mind itself, Gigerenzer (1991) refers to this transition as the tools-to-theories heuristic.

A classic example is the proposal that cognition functions like a statistician, inspired by the rise of inferential statistics in the early twentieth century. In vision science, a similar shift occurred when Bayesian methods, originally developed for statistical inference, were reinterpreted as claims about how the visual system processes uncertain information.

A contemporary version of this pattern appears with deep neural networks. DNNs are highly successful engineering tools for object recognition, classifying images by detecting and combining hierarchical features. They also do more than perform well at classification: the internal representations of some networks predict neural responses across multiple stages of the primate visual cortex with considerable accuracy.

This makes them candidates for computational models of biological vision, not merely convenient algorithms. Yet predictive fit does not establish shared mechanisms. Standard ImageNet-trained convolutional neural networks, for example, often rely more strongly on texture cues for object recognition, whereas human observers rely more strongly on shape. Such discrepancies between model and organism may reflect differences in training conditions rather than fundamental architectural limitations, but they illustrate the risk of moving too quickly from predictive success to mechanistic claims.

When Models Inform Theory: Dopamine and Prediction Errors

Although a good fit to behavioral or neural data does not establish shared mechanisms, computational models can play an important role in theory development. The Rescorla–Wagner model was introduced as a formal model of Pavlovian conditioning. It describes learning as an updating process driven by the difference between expected and observed outcomes. Later, Schultz, Dayan, and Montague (1997) found that dopamine neurons in the primate midbrain exhibit activity resembling the reward-prediction-error signals described by reinforcement learning models. This suggested that the broader computational principle of prediction-error learning might capture something important about neural function. The reward-prediction-error hypothesis subsequently became one of the most influential computational ideas in neuroscience.

Nearly thirty years later, however, the picture has become considerably more complex. Dopamine neurons appear to do more than signal reward prediction errors, and whether the original framework is sufficient remains an open and actively debated question. The dopamine case illustrates that science is an iterative process: models inform theories, theories motivate experiments, and new evidence reshapes both.

The Ambiguity of Computational Frameworks

So which of the frameworks encountered in this chapter are tools, and which are models? The answer depends on the claims attached to them. Deep neural networks are engineering tools for classification, but they become candidate models of cognition when their internal representations are compared with neural activity and interpreted as hypotheses about biological computation. Bayesian inference can serve as a statistical tool for analyzing data, but it becomes part of a model of cognition when we propose that the brain performs something functionally resembling probabilistic inference over a generative model. The Rescorla–Wagner learning rule is a formal model of conditioning. The broader prediction-error framework gained additional credibility as a candidate model of neural computation when dopamine neurons were found to exhibit functionally analogous signals. In each case, the transition from tool to model is not automatic. Predictive success demonstrates that a formalism is useful, but stronger mechanistic claims require independent evidence that the model captures relevant representations, operations, or physical processes within the organism. The ambiguity in how we use the terms tool, model, and theory is therefore not accidental. It reflects the persistent hope of cognitive scientists that the computational frameworks they construct will turn out to describe not only the data, but something about the mind itself.

Authors: Fabian Mueller & William Palmer

References

Gigerenzer, G. (1991). From tools to theories: A heuristic of discovery in cognitive psychology. Psychological Review, 98(2), 254–267.

Box, G. E. P. (1979). Robustness in the strategy of scientific model building. In R. L. Launer & G. N. Wilkinson (Eds.), Robustness in Statistics (pp. 201–236). Academic Press.

Wichmann, F. A., & Geirhos, R. (2023). Are deep neural networks adequate behavioral models of human visual perception? Annual Review of Vision Science, 9, 501–524. https://doi.org/10.1146/annurev-vision-120522-031739

Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., & Brendel, W. (2019). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. Proceedings of the International Conference on Learning Representations (ICLR 2019).

Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., & DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the National Academy of Sciences, 111(23), 8619–8624.

Khaligh-Razavi, S.-M., & Kriegeskorte, N. (2014). Deep supervised, but not unsupervised, models may explain IT cortical representation. PLoS Computational Biology, 10(11), e1003915.

Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599.

Adam, D. (2026). Dopamine takes a hit: how neuroscience is rethinking the “feel-good” chemical. Nature, 651, 572–574.