BLUEPRINTS OF THE MIND
4.11
From Models to Theories – and Back Again
This chapter traced how cognitive science uses computational ideas to describe the mind, from the earliest stages of visual processing to probabilistic models of the world, other agents, and the self. But when a computational model captures behavior well, does that mean the mind actually works that way?
We began with visual processing as a hierarchical system that transforms light into structured neural representations. From early feature detection to more complex combinations, the visual system extracts regularities in the structure of sensory input and transforms them into increasingly structured representations.
Pattern recognition offers an account of how these representations support object classification under variation in viewpoint, illumination, and context. Deep neural networks have become a powerful mathematical framework for modeling this capacity, classifying images by detecting and combining hierarchical features in ways that parallel certain properties of the ventral visual stream.
Yet not all of perception is best explained by a discriminative approach. Because the same sensory input can arise from multiple external causes, perception also involves an inverse problem. Bayesian approaches describe its resolution as probabilistic inference, in which sensory evidence is integrated with prior expectations to estimate the most likely state of the world.
Analysis-by-synthesis and predictive coding extend this account by modelling perception as a continuous process of prediction and error correction: internal models generate expectations about sensory input, and mismatches between predicted and actual signals drive updating.
Discriminative and Generative Models
A fundamental distinction that matters for understanding computational models of the brain, is between discriminative and generative approaches.
A discriminative model learns a direct mapping from input to output. Given a retinal image, it predicts the object category. It does not need to know how images are generated; it only needs to find statistical regularities that distinguish one category from another. Deep neural networks trained for classification are the prototypical example.
A generative model, by contrast, learns the process by which observations are produced. It captures the statistical structure of the environment and inverts this model to infer the most likely causes of a given observation. Bayesian approaches to perception are built on this logic.
The distinction matters because the two approaches carry different theoretical commitments. A generative model implies that perception involves reconstructing the hidden causes of sensory input. A discriminative model implies something closer to pattern classification: extracting task-relevant features without building a causal model of the world.
In biological vision, both modes may coexist. Early sensory processing may operate more like a feedforward discriminative system, as suggested by the speed of object recognition. Slower, recurrent processing may implement something closer to generative inference, as suggested by effects of context, expectation, and ambiguity on perception. The question is not which framework is correct but which aspects of cognition each framework best explains.
The chapter then shifted from perception to learning. Reinforcement learning models describe how expectations are updated through prediction errors when outcomes differ from prior estimates. The Rescorla-Wagner model provided a formal account of this principle, and later work on midbrain dopamine neurons suggested that the brain may implement functionally similar prediction-error signals.
The discussion then broadened to probabilistic generative models of intuitive physics, theory of mind, and even self-perception. Across these domains, internal models were presented as enabling the simulation of physical events, the inference of other agents’ mental states, and the representation of the self as an acting agent in the world.
Finally, we considered the limits of computational explanation. The Chinese Room argument questions whether formal symbol manipulation is sufficient for understanding. Embodied cognition highlights that the body’s physical structure shapes behavior in ways that pure computation does not capture. Phenomenology points to the qualitative character of conscious experience, that is the felt quality of seeing, hearing, or being a self, all of which might resist formalization.
Models: Tools, Not Truths
Before closing, another important distinction should be made explicit. Throughout this chapter, we have used the term ‘model’ loosely, referring to Bayesian inference, reinforcement learning, and probabilistic generative models as computational models of cognition.
But we have also been careful with our language: these frameworks can be described as capturing certain aspects of cognition, they process information as if performing probabilistic inference, dopamine neurons fire in a pattern functionally analogous to a prediction error signal.
But the fact that processes such as perception, learning, or even self-representation can be mathematically modelled successfully does not mean that these frameworks are literally true descriptions of cognition. They are, first and foremost, formal tools. And when a tool becomes powerful or fashionable, it can easily be reinterpreted as a theory of the mind itself. Gigerenzer (1991) called this pattern the tools-to-theories heuristic.
Tools, Computational Models and Theory
What exactly is the difference between a tool and a theory? Wichmann and Geirhos offer a useful clarification of this distinction: a tool, they argue, is an algorithm or method that serves the scientific process but is not itself of scientific interest. A computational model, by contrast, is a concrete instance of a theory and is itself of scientific interest. The difference lies in whether the formalism carries a theoretical commitment about cognition or simply gets the job done.
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 data analysis, were reinterpreted as claims about how the visual system computes. A contemporary version of this pattern appears with deep neural networks. DNNs are highly successful at object recognition, classifying images by detecting and combining hierarchical features. And they do more than perform well as engineering tools: their internal representations predict neural responses across multiple stages of the primate visual cortex with considerable accuracy.
This makes them candidate computational models of biological vision, not merely convenient algorithms. Yet predictive fit does not establish shared mechanisms. Standard deep neural networks, for instance, rely predominantly 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 Tools Become Theory: The Case of Dopamine and Prediction Errors
Although good fit to behavioral and neural data does not guarantee shared principles, computational tools can play an important role in theory development.
Consider the Rescorla-Wagner model, which began as a mathematical tool for describing how animals learn from prediction errors.
Decades later, Schultz, Dayan, and Montague (1997) found that dopamine neurons in the primate midbrain fire in a pattern closely resembling the model’s error term. The tool, it seemed, had captured something real about neural mechanisms, and the reward prediction error hypothesis 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 far 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: tools inform theories, theories motivate experiments, and new data reshape both.
Models, Tools, and the Ambiguity of Computational Frameworks
So which of the frameworks we encountered in this chapter are tools, and which are models?
The answer is not always clear, and it depends on the claims we attach to them. Deep neural networks, for instance, are engineering tools for classification, but they become models when their internal representations are compared to neural activity and taken to reflect something about biological computation.
Bayesian inference can serve as a statistical tool for analyzing data, but it is sometimes accounted as a model when we propose that the brain actually performs probabilistic inference over a generative model.
The Rescorla-Wagner learning rule began as a tool for describing conditioning data, and became a candidate model when dopamine neurons were found to fire in a pattern resembling its error term. In each case, the transition from tool to model is not automatic. It requires independent evidence that the formalism captures something about the underlying mechanism, not just about the data.
However, this ambiguity in language is not accidental. It reflects the persistent hope of cognitive scientists that the computational frameworks they build will turn out to describe not just the data, but the mind itself.
Author: Fabian Mueller
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.