The Bayesian brain hypothesis represents one of the most ambitious unifying frameworks in the cognitive sciences: the proposal that neural computation implements probabilistic inference, weighing prior beliefs against incoming evidence according to Bayes' theorem. Under this view, the brain is fundamentally a hypothesis-testing machine, continuously updating posterior distributions over states of the world to guide adaptive choice.

The theoretical appeal is considerable. Bayesian inference provides a normative account of optimal reasoning under uncertainty, and empirical work across perception, motor control, and higher-order cognition has documented behavioral signatures consistent with probabilistic integration. Yet the hypothesis faces persistent challenges, most notably systematic deviations from Bayesian optimality observed in probabilistic reasoning tasks, base-rate neglect, and conservative belief updating.

Reconciling these observations requires a more nuanced formulation. Rather than treating the brain as an ideal Bayesian observer, contemporary theorists increasingly conceptualize cognition as implementing approximate Bayesian inference under computational constraints. This shift transforms the debate from whether the brain is Bayesian to how it approximates probabilistic computation, and what algorithmic strategies it deploys to achieve resource-rational behavior. The stakes extend beyond theory: understanding these mechanisms illuminates why choices sometimes appear irrational while nonetheless serving adaptive functions.

The most compelling evidence for Bayesian decision-making comes from tasks in which observers must combine noisy sensory evidence with structured prior knowledge. Ernst and Banks' seminal work on visuo-haptic integration demonstrated that humans weight modalities in inverse proportion to their variance—a hallmark of Bayesian cue combination. Similar patterns emerge in speed estimation, where perceived motion is biased toward slower velocities consistent with a slow-motion prior.

In value-based decision-making, priors manifest as base rates, reward expectations, and structural beliefs about the choice environment. Behavioral and neural evidence indicates that these priors modulate choice in ways consistent with posterior computation: as evidence quality degrades, priors exert stronger influence on decisions, while sharper evidence dominates when likelihoods are precise. This dynamic weighting is difficult to reconcile with non-probabilistic accounts.

Neurally, orbitofrontal and parietal cortices encode signals resembling posterior probabilities and log-likelihood ratios. Population activity in lateral intraparietal cortex during perceptual decisions tracks accumulated evidence in a manner formally equivalent to sequential probability ratio testing, providing a mechanistic bridge between Bayesian theory and neural implementation.

Yet the elegance of these findings should not obscure their limits. Priors in laboratory tasks are often explicitly specified or trained over hundreds of trials, sidestepping the harder question of how naturalistic priors are learned, structured, and retrieved. Real-world choice involves hierarchical priors over uncertain generative models, not merely fixed distributions over known states.

The empirical signature of prior integration thus supports a Bayesian framework more than a strict Bayesian mechanism. What the brain achieves is not literal posterior computation but a functional approximation that captures the essential logic of evidence weighting.

Takeaway

Rationality under uncertainty is not about being right—it is about weighting what you already believe against what you now observe in proportion to their relative reliability.

Systematic departures from Bayesian optimality—conservatism in belief updating, base-rate neglect, representativeness heuristics—have historically been marshaled as evidence against the Bayesian brain. The resource-rational reframing proposes instead that these deviations reflect optimal computation under constraints of time, memory, and metabolic cost.

Formally, resource-rational analysis reconceptualizes the objective function. The agent maximizes expected utility net of computational cost, yielding a bounded-optimal policy that may diverge from unbounded Bayesian ideals. Under this framing, apparent irrationality is often the signature of an algorithm making principled trade-offs between accuracy and effort.

Consider anchoring effects. A Bayesian ideal observer would treat an irrelevant anchor as uninformative, but a resource-limited agent using iterative adjustment must start somewhere, and terminating adjustment before convergence produces systematic bias. The bias is not a bug in the algorithm—it is the algorithm's cost-conscious behavior.

This perspective reframes classical findings from Kahneman and Tversky. Heuristics such as availability and representativeness cease to be evidence of non-Bayesian cognition; they become candidate approximations whose deviations from optimality are predictable from the structure of the computational problem. Griffiths and colleagues have formalized this by deriving heuristics as rational solutions under specified resource constraints.

The empirical program is demanding. Resource-rational models must specify the cost function, the class of admissible algorithms, and the environmental statistics with sufficient precision to yield falsifiable predictions. Without such constraints, the framework risks becoming unfalsifiable—capable of rationalizing any observed deviation post hoc.

Takeaway

The distinction between error and efficiency dissolves when we account for cost: what looks like irrationality may be optimal reasoning under a budget you failed to notice.

If the brain performs approximate Bayesian inference, what algorithm does it use? Monte Carlo sampling has emerged as a leading candidate because it offers a neurally plausible substrate for probabilistic computation without requiring explicit representation of full posterior distributions.

Under sampling accounts, beliefs are represented implicitly by the frequency with which the neural circuitry generates hypotheses or scenarios. A choice is made not by integrating over a densely represented posterior but by drawing a small number of samples and acting on their aggregate. This is computationally tractable and matches known constraints on neural representation.

The account predicts specific behavioral signatures. With few samples, choices should exhibit probability matching, variability across repeated queries, and biases toward high-likelihood hypotheses. Vul and colleagues have shown that human probability estimates behave as if generated from small numbers of posterior samples—sometimes as few as one to seven—consistent with a sample-based approximation.

Neural evidence is accumulating. Spontaneous cortical activity in visual cortex resembles samples from a learned generative model of natural images, and hippocampal replay has been interpreted as trajectory sampling for evaluating potential future outcomes. Such findings suggest that sampling may be a general-purpose neural computation rather than a domain-specific trick.

Critical questions remain. Which sampling algorithm—Metropolis-Hastings, particle filtering, importance sampling—best captures neural implementation? How are samples generated in autocorrelated sequences, and how does this shape choice dynamics? Answering these questions requires tighter coupling between algorithmic theory and neural measurement than the field has yet achieved.

Takeaway

Belief may not be a distribution the mind holds but a process the mind runs—thoughts are samples, and thinking longer means sampling more.

The Bayesian brain hypothesis has evolved from a bold claim about optimal inference into a more mature theoretical framework centered on approximation, resource constraints, and algorithmic implementation. This evolution has strengthened rather than weakened the paradigm, aligning it with the messy realities of biological computation.

The productive question is no longer whether decisions are Bayesian but which approximation strategies the brain deploys, how those strategies are selected and calibrated, and how their computational signatures manifest in neural activity. Sampling algorithms provide one compelling class of answers, but the space of admissible approximations remains large and underdetermined.

What emerges is a view of choice as neither ideally rational nor fundamentally flawed, but as computation shaped by the twin pressures of environmental structure and biological cost. Understanding decision-making within this framework requires the integration of formal theory, algorithmic analysis, and neural measurement—a program still in its early chapters.