An earlier entry argued that confidence is the wrong category: a scalar confidence, the aleatoric and epistemic decomposition, and even distribution-free conformal guarantees all answer a distributional question rather than certifying that a particular claim is sound enough to act on. That entry walked the top of the apparatus. This one walks the rest of it, because a supervisor is entitled to ask whether some other rung of machine learning's uncertainty machinery does what confidence cannot. The honest answer, from the field's own primary sources, is that it does not, and the survey that maps the whole apparatus says so in as many words.
The methods, and the question each one answers
The most-cited single method for representing what a model does not know is Monte Carlo dropout, which casts dropout at test time as approximate Bayesian inference and reads model uncertainty off the variance of the resulting predictions. Its authors are unusually direct about the failure it is meant to cure, and the sentence is the thesis in the field's own words: the softmax outputs at the end of the pipeline are often erroneously interpreted as model confidence, and a model can be uncertain in its predictions even with a high softmax output (Gal and Ghahramani, 2016). That is the wrong-category claim, stated by the people who built the standard remedy for it. Their remedy produces a better-behaved uncertainty estimate; it does not convert the softmax number into a warrant that the prediction's basis suffices for an action.
The scalable alternative to the Bayesian route is the deep ensemble, which trains several networks and reads uncertainty from their disagreement. The contribution is real, but the framing carries the same caution the review has leaned on throughout: a prediction may be accurate and yet miscalibrated, and vice versa, so accuracy and the quality of the uncertainty are orthogonal (Lakshminarayanan et al., 2017). The best the method claims on inputs unlike the training data is to express higher uncertainty on out-of-distribution examples, which is a distributional signal about the input's novelty, not a judgement about whether this claim may be acted upon.
A third route abandons the softmax altogether. Evidential deep learning places a Dirichlet distribution over the class probabilities, treating the network's output as accumulated evidence and giving it, in principle, a way to say it does not know. Its diagnosis of the standard approach is precise: softmax is not capable of inferring the predictive-distribution variance, so the distance of a predicted label is not useful for the conclusion drawn from it (Sensoy et al., 2018). Yet what it yields is an evidence-mass distribution over classes, a richer representation of uncertainty, and still not a per-action warrant. Even the method built expressly to represent I do not know represents it as a distribution, which a receiver must then interpret.
The rung below: is this input even in scope
Before any of these can be trusted, a prior question has to be answered: is the input the model is judging even the kind of thing the model was trained on. The foundational baseline for that question uses the maximum softmax probability to flag misclassified and out-of-distribution examples, and it works well enough to be a standard. But its authors record the finding that matters here: softmax classifier probabilities are not directly useful as confidence estimates, and, memorably, random Gaussian noise fed into an image classifier draws a predicted class probability of ninety-one percent (Hendrycks and Gimpel, 2017). A model will assert a class, with high confidence, about pure noise. This is the starkest possible statement of the problem the whole review turns on: fluent, confident output is emitted for an input that supports no claim at all, and nothing in the output distinguishes that case from a well-founded one.
The field's own verdict
The decisive source is not any single method but the survey that maps them all. The comprehensive survey of uncertainty in deep neural networks catalogues the sources of uncertainty, the Bayesian, ensemble, test-time-augmentation, and single-deterministic families of method, and the calibration literature that tries to make their outputs trustworthy. Its conclusion, stated plainly, is the capstone of this theme: estimating the predictive uncertainty is not sufficient for safe decision-making (Gawlikowski et al., 2021). The survey names the specific gaps, among them the inability to distinguish in-domain from out-of-domain samples reliably and the inability to provide reliable uncertainty estimates for a network's individual decision. The field that built the apparatus has surveyed it and reported that quantifying uncertainty, however well, does not by itself make a decision safe to take.
What this establishes
The uncertainty apparatus is broad and genuinely useful, and none of it is admissibility. Dropout and ensembles give better-behaved estimates of a distribution; evidential methods represent evidence mass; out-of-distribution detection flags novelty; and the survey that gathers them concludes that the sum is still not sufficient for safe decision-making. Each rung answers a distributional or novelty question about the estimation process; none carries, with a particular output, a representation of whether its basis suffices for a particular irreversible action. That is not a complaint about immature tooling that better uncertainty quantification will fix. It is a category difference, confirmed at the level of the field's own survey, and it is precisely the space this work occupies: not a better uncertainty estimate to feed the model, but a judgement, carried across the boundary, that the receiver can refuse on.
References
Gal, Y. and Ghahramani, Z. (2016). Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning. ICML 2016. arXiv:1506.02142. arxiv.org/abs/1506.02142
Lakshminarayanan, B., Pritzel, A. and Blundell, C. (2017). Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. NeurIPS 2017. arXiv:1612.01474. arxiv.org/abs/1612.01474
Sensoy, M., Kaplan, L. and Kandemir, M. (2018). Evidential Deep Learning to Quantify Classification Uncertainty. NeurIPS 2018. arXiv:1806.01768. arxiv.org/abs/1806.01768
Hendrycks, D. and Gimpel, K. (2017). A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. ICLR 2017. arXiv:1610.02136. arxiv.org/abs/1610.02136
Gawlikowski, J., Njieutcheu Tassi, C. R., Ali, M., Lee, J., Humt, M., Feng, J., Kruspe, A., Triebel, R., Jung, P., Roscher, R., Shahzad, M., Yang, W., Bamler, R. and Zhu, X. X. (2021). A Survey of Uncertainty in Deep Neural Networks. Artificial Intelligence Review. arXiv:2107.03342. arxiv.org/abs/2107.03342