Beyond Bayes: Prediction Isn’t Understanding… But Is Understanding Necessary?

Why does the brain expect a coffee cup to fall before it actually slips? One popular answer is Bayesian: the brain combines prior beliefs with incoming sensory data to predict outcomes. But how satisfying is that answer if it doesn’t tell us how the brain implements the prediction? Dr. Madhur Mangalam echoes this criticism of Bayesianism in his recently written article “The Myth of the Bayesian Brain”. At its core, Bayesianism is a powerful mathematical approach that captures input–output relationships. It does this extremely well for a wide range of systems. Mangalam’s central claim is not that Bayesian methods are wrong or useless, but that they are too powerful. Precisely because Bayesian models can be tuned to fit almost any dataset, they risk explaining everything and, in doing so, explaining nothing. This flexibility becomes a serious problem in neuroscience, where the goal is not merely to reproduce behavior, but to explain the underlying mechanisms that generate it.

A basic schematic of a Bayesian brain. Sensory inputs are combined with prior beliefs to generate predictions. When predictions mismatch the sensory input, the brain uses this information to update its priors. From: "The Myth of the Bayesian Brain Figure 1"

Before engaging further with this critique, it is important to be explicit about my own biases. I bring strong priors to how I approach science, and those priors make me particularly receptive to Mangalam’s arguments. My background is in the use of ordinary differential equations and bifurcation theory to study neural systems. I am deeply uncomfortable with black-box explanations. I prefer approaches where a specific parameter,such as an ion channel conductance, can be directly manipulated. These manipulations unfold a bifurcation, and where the resulting change in system dynamics provides a mechanistic explanation of how a neural circuit functions. When I can trace behavior back to the structure of phase space and its bifurcations, I feel that I genuinely understand the system.

There is, of course, an important caveat to this perspective. Biological systems are highly redundant. Multiple ion channels, neuron types, and genetic pathways can often serve overlapping or compensatory roles. This redundancy makes biology robust, but it also means that any single bifurcation-based explanation risks being overly simplistic. Even so, with that caveat acknowledged, I still take pride in working within a framework that aspires to mechanistic clarity.

That said, Bayesianism undeniably provides a compelling framework for a different problem: given an input, or a sequence of inputs, how can one predict a particular set of outputs? Bayesian approaches excel at this task. Through repeated observations or trials,often mediated by sensory information, priors are updated, predictions improve, and input–output mappings become increasingly accurate. If the advent of ChatGTP and AI shows, this technique can be extremely powerful.

Given that predicting the outputs of neural systems based on their inputs mirrors a core function of the brain, especially in sensory processing and motor control, is it really such a surprise that this aligns well with the Bayesian framework. Learning and synaptic plasticity map onto this framework naturally. But this strength is also a double-edged sword. The algorithms we use are impressive and can reliably map nearly any input to an output. If an unknown input is within the distribution, these algorithms can easily predict what the brain will do in an experiment, giving the illusion that the brain is Bayesian. Without this predictive ability, the Bayesian mathematical framework would not appear to work.

Mangalam argues that using the Bayesian framework is effectively “curve fitting”, and “curve fitting” does not explain the mechanisms at play. It conflates a clean input–output relationship with the actual processes the brain uses. For example, as one moves up the visual hierarchy, receptive fields become more complex, and learning algorithms can reproduce these receptive fields. Similarly, when the brain encounters novel information, it adapts in a way that resembles updating priors. Yet this leaves many questions unanswered: how does this adaptation occur biochemically?

This figure shows how more data, along with stronger assumptions about how to process and fit that data, make it difficult to falsify the Bayesian perspective. Dr. Mangalam uses this to argue that Bayesianism is so flexible that it offers little real understanding of mechanism. He critiques its scientific merit somewhat harshly, since science requires falsifiability. From "The Myth of the Bayesian Brain Figure 2." 

Before the advent of ChatGPT and large language models, I would have wholeheartedly agreed with Mangalam. A black box like Bayesian prior updates does not explain anything; my mechanistic approach feels far more informative for understanding the brain. However, building extremely intelligent machines shows that we do not need to know the brain in mechanistic detail to achieve powerful AI. At a high level, AI uses information in its training data to construct models that update priors and predict outputs from in-distribution inputs. Minimizing a loss function is essentially equivalent to using inputs, predicting outputs, and minimizing the error between prediction and observation. AI has proven to be extremely powerful. While it is not perfect, we are still in AI infancy; its capabilities in 20 or 50 years could be far beyond what we see today.

Does this mean AI confirms the Bayesian approach? Not entirely. There are limitations. AI struggles with out-of-distribution inputs, does not learn novelty quickly, and handles real-world uncertainty poorly. As Mangalam points out, the brain's nonlinearities, like dendritic computation, behavioral timescale plasticity, and structured recurrent connections, particularly involving inhibitory interneurons—might hold clues to addressing these issues. Understanding these mechanisms could be the next major advance in AI, or AI may simply continue improving through scale and data without mechanistic insight. As a computational neuroscientist, I acknowledge that my perspective on AI is limited.

Where Bayesian frameworks clearly fall short is in treatment development. Unlike AI, humans require a mechanistic understanding to develop therapies. Mangalam’s critique of applying Bayesian models to schizophrenia is compelling: framing schizophrenia as a disorder of improper prior updating is elegant, but it does not reveal how dopamine and NMDA receptor dysfunctions cause specific symptoms or why antipsychotics selectively treat positive symptoms. Bayesian models alone cannot identify treatment targets; mechanistic insights tied to synaptic plasticity are essential.

Perhaps a compromise (or cop-out) is that Bayesian models and mechanistic neurodynamics answer different questions. Bayesian and ML methods are unmatched when the goal is prediction, while dynamical, mechanistic models are essential when we want causal explanations or to identify treatment targets. Personally, I think it’s wise to keep abreast of both frameworks: advances in one could inspire insights in the other. Using Bayesian/ML tools to build intuition and predictions, alongside dynamical systems to test hypotheses about mechanisms, is likely where real progress will come.

ChatGPT used to clean up my awful grammar. Ideas are mine.

Author: Alex White

Original: Mangalam, M. (2025). "The myth of the Bayesian brain." European Journal of Applied Physiology, 125(10), 2643-2677.

https://doi.org/10.1007/s00421-025-05855-6