Concept
Uncertainty in learning and prediction
4 studiesEvidence last moved Sep 27, 2026
Uncertainty is a model's or learner's representation of how unsure it should be: about noisy outcomes, about whether the world has changed, and about its own predictions. The evidence here spans Bayesian models of human learning in changing environments (a lab experiment and a model-fitting study), a benchmark of uncertainty quantification for protein property prediction, and a Bayesian parameter-estimation pipeline built on a neural emulator of the heart.
A prediction without a sense of how much to trust it is hard to act on. These papers separate kinds of uncertainty and show that being accurate and being well calibrated are different properties.
Studies
4
Findings
5
5 supporting · 0 challenging · 1 qualifying citations
Open tensions
1
Latest change
Concept page published
Uncertainty in learning and prediction
Currently
What we know
- Accuracy and calibration can come apart.
- Good uncertainty estimates do not automatically improve downstream decisions.
- People can behave as if they track distinct kinds of uncertainty, and tend to avoid the unknown.
- Learning rates that rise with estimated volatility describe many people's behaviour.
- Fast emulators make full posterior uncertainty practical for expensive simulators.
Largest unresolved question
Bayesian volatility tracking is not universal: when participants were not told probabilities could jump, the Bayesian model no longer beat reinforcement learning, and in one reversal dataset about 30% of people were best fit by a plain Kalman filter.
Common misconceptions
The most accurate model also gives the most trustworthy uncertainty.
In the protein-engineering benchmark, ensembles were often among the most accurate yet poorly calibrated, while simpler probabilistic models were often better calibrated.
If a Bayesian model fits behaviour best, the brain must be computing that algorithm.
Both human-learning studies are model comparisons on choice data; a better fit shows the model describes behaviour well, not that neurons implement it, and fits changed with task instructions.
Related
Claim ledger
What the evidence shows
Drawn from 4 studies in this library. Mix labels say which citation roles are present; they are not a strength score. Supports means evidence for a finding; Challenges means evidence against a stated position; Qualifies marks scope.
Accuracy and calibration can come apart.
In a benchmark on three protein sequence-function landscapes, no uncertainty method was best across datasets and metrics; CNN ensembles were often among the most accurate yet among the worst calibrated, while Gaussian processes and Bayesian ridge regression were often better calibrated.
Good uncertainty estimates do not automatically improve downstream decisions.
In the same retrospective benchmark, uncertainty-based sampling often beat random sampling in later active-learning rounds, but in Bayesian optimisation no uncertainty-aware acquisition beat a greedy strategy.
People can behave as if they track distinct kinds of uncertainty, and tend to avoid the unknown.
When undergraduates were fully informed that reward probabilities could jump, a Bayesian model that tracks unexpected change fitted their choices better than reinforcement learning, and people penalised options with high estimation uncertainty rather than exploring them.
Learning rates that rise with estimated volatility describe many people's behaviour.
A two-level volatile Kalman filter came within a few percent of a particle-filter benchmark while the hierarchical Gaussian filter erred by about a fifth, and the volatile Kalman filter was the best-fitting model for 37 of 44 and 102 of 161 participants in two reversal-learning datasets.
Fast emulators make full posterior uncertainty practical for expensive simulators.
A latent neural ODE emulator of whole-heart simulations let Bayesian inference recover parameters of unseen simulations inside their 95% credibility regions, about 1,718 times faster than using the original model.
- Can a small neural network stand in for a supercomputer heart model?— One patient's anatomy, simulated rather than clinical observations, and only 5 test simulations.
Debates
Tensions and limits
Some items are genuine disagreements on the same question. Others mark different assays, populations, or outcomes.
Bayesian volatility tracking is not universal: when participants were not told probabilities could jump, the Bayesian model no longer beat reinforcement learning, and in one reversal dataset about 30% of people were best fit by a plain Kalman filter.
Bayesian volatility tracking is not universal: when participants were not told probabilities could jump, the Bayesian model no longer beat reinforcement learning, and in one reversal dataset about 30% of people were best fit by a plain Kalman filter.
- Do people learn like Bayesians when rewards suddenly change?
- How should a learner speed up learning when the world keeps changing?
Study Role Design N Population Outcome Do people learn like Bayesians when rewards suddenly change? Supports Human experimentModel comparison (Bayesian learner with forgetting vs Rescorla-Wagner and Pearce-Hall) on choices from a six-arm restless bandit, plus a new experiment manipulating how much task structure participants were told. 43 undergraduates in the no-information treatment; the numbers in the other two treatments are garbled in the extracted text, and treatments 1 and 2 pooled gave 75 cases. Reanalysed data from an earlier experiment are also used. Undergraduates at Ecole Polytechnique Fédérale Lausanne Model fit (log-likelihood, BIC) of learning models to trial-by-trial choices; debriefing reports of noticing jumps How should a learner speed up learning when the world keeps changing? Supports Computational / modellingNew Bayesian learning algorithm (VKF) tested in simulations against ground truth, a particle filter and the HGF, then fitted to two existing human choice datasets with hierarchical model comparison. Two reanalysed human datasets: 44 participants in the first task and 161 analysed (of 174) in the second; simulations used 1000 or 500 generated sequences. Simulated volatile environments; adult participants in two published probabilistic reversal-learning tasks. Tracking error relative to a near-optimal particle filter; learning-rate/volatility relationship; model evidence for fitting human choices.
PaperFren reads this as a limit on how far one study travels — different assays, populations, or outcomes — not a forced fight between papers.
Timeline
How understanding moved
Study years are when the paper was published. Evidence edits are dated changes to this page's claims. Explanations are when PaperFren added a Discovery — not a claim that the science happened that day.
2026
Concept page published
Uncertainty in learning and prediction
Change log
What changed
Dated edits to this page's evidence: studies added or removed from a claim, claims added or withdrawn, and new explanations tagged here. Rewordings are not listed.
- Concept page published
Papers
4 studies in this library bear on Uncertainty in learning and prediction, ordered by citations.
- Do people learn like Bayesians when rewards suddenly change?
People chose like Bayesian learners who dislike uncertainty, but only when told that reward odds could suddenly jump; otherwise simple trial-and-error learning fit just as well.
- How should a learner speed up learning when the world keeps changing?
A simple extension of the Kalman filter that also tracks how fast the world is changing approximates optimal learning more accurately than the popular HGF and explains most people's choices better.
- Can a small neural network stand in for a supercomputer heart model?
A compact neural ODE learned to reproduce a detailed whole-heart simulator's pressure and volume curves with a few percent error, making heavy analyses feasible on a laptop.
- Which uncertainty methods help ML guide protein engineering?
No uncertainty-estimation method was best across protein tasks, and using uncertainty to choose which proteins to test next never beat simply picking the highest predicted ones.
Compare studies
Select 2–10 studies. Design and N are labels, not a ranking.
Nothing selected yet.
Questions
What is still open
Bayesian volatility tracking is not universal: when participants were not told probabilities could jump, the Bayesian model no longer beat reinforcement learning, and in one reversal dataset about 30% of people were best fit by a plain Kalman filter.
Ask PaperFren about Uncertainty in learning and prediction
Study this conceptflashcards and short-answer questions
Explain the difference between accuracy and calibration using evidence from protein engineering.
Accuracy measures how close predictions are to the truth; calibration measures whether stated uncertainty matches actual error rates. A benchmark on GB1, AAV and thermostability landscapes found CNN ensembles often among the most accurate but worst calibrated, and Gaussian processes often better calibrated. Accuracy dropped as domain shift grew, while calibration showed no clear pattern. Because it was retrospective, it does not show what happens in real lab campaigns.
What is 'volatility' in learning, and how should it affect a learning rate?
Volatility is how quickly the underlying reward contingencies change. A learner that estimates higher volatility should weight recent outcomes more, i.e. raise its learning rate. The volatile Kalman filter does this and fitted most participants best in two reversal-learning datasets, whose fitted learning rates rose after switches. But a sizeable minority were fit by a plain Kalman filter, and an experiment found Bayesian change-tracking only beat reinforcement learning when people were told changes could happen.
Flashcards
0 of 8 answers reviewed