How would you get an epistemic uncertainty estimate out of an already-trained network?

## Why one network cannot answer this Epistemic uncertainty is disagreement among the many functions that fit your data. A single trained network *is* one of those functions; asking it how sure it is about itself gets you its output score, which is a ranking over classes and not a statement about its own ignorance. To measure disagreement you need more than one plausible model. Every practical technique here is a way of producing several models cheaply and reading their spread. ## MC dropout Dropout, during training, zeroes a random subset of units on every pass. Standard inference removes that randomness so predictions are deterministic. MC dropout deliberately does the opposite: it keeps the random masks switched on at prediction time and runs the same input through T times (thirty is a common working number), producing T different predictive distributions. Average them for the prediction; use the variance across them, or the disagreement measure below, for the uncertainty. The justification is that this approximates averaging over a distribution of models rather than committing to one. Each mask corresponds to a thinned sub-network, and averaging over masks approximates a posterior average. Its advantages are real and practical: no retraining, one set of weights to store, and you can tune T against your latency budget. Its weaknesses are just as real, and a senior candidate is expected to name them: - **The diversity is local.** Every sample is a perturbation of one trained weight setting. If that setting is confidently wrong in a region, all thirty samples tend to be confidently wrong together, and the measured disagreement understates the true epistemic uncertainty. - **The drop rate was chosen for the wrong objective.** It was tuned to regularize, by validation accuracy. Nothing made it the rate that gives faithful uncertainty, and raising it to widen the spread degrades the prediction itself. - **Placement decides everything.** A network whose only dropout sits immediately before the classifier head perturbs almost nothing; the shared representation is identical on every pass, so the spread is tiny and uninformative regardless of how unfamiliar the input is. - **Other stochastic layers must stay deterministic.** Normalization layers must keep using their stored population statistics; if they instead recompute statistics from whatever inputs share the pass, you are measuring batch composition, not model uncertainty. - **Cost is multiplicative at serving time.** T passes per prediction, in the latency budget of every request. ## Deep ensembles Train M complete networks — five is the usual working number — differing only in their random initialization and data shuffling, then average their predictive distributions. Independent initializations converge to genuinely different solutions, not perturbations of one, and that is the whole point: the members are *functionally* diverse. Where training data was dense they are forced into agreement; where it was sparse or absent nothing constrained them and they diverge. Feed the ensemble an input from a class no member ever saw and you typically see five confident but *different* answers — each softmax near one, but pointing at different classes. That pattern is precisely the signal a single confident softmax cannot give you. Ensembles are consistently the stronger baseline for epistemic uncertainty and for detecting distribution shift. The cost is blunt: M times the training compute, M times the storage, M times the inference cost — although the M passes parallelise across devices in a way the sequential structure of a latency budget sometimes tolerates better than it looks. Note that this means averaging the *outputs* of independently trained networks, which is a different operation from combining their parameters. ## Reading the spread properly With a set of member distributions in hand, classification admits an exact split. Let the members' predictive distributions be `p_1 ... p_M` and their average be `p_bar`. Then `H(p_bar) = mean_m H(p_m) + I` where `H` is entropy. The left side is **total** predictive uncertainty. The first right-hand term, the average of each member's own entropy, is the **aleatoric** part: uncertainty every member agrees is inherent in the input. The remainder `I` (a mutual information, also called the disagreement or the Jensen-Shannon divergence of the members) is the **epistemic** part. It is zero when the members agree exactly, however unsure each individually is, and it is large when confident members point in different directions. That identity is what makes ensembles operationally useful. An input where every member says `0.5 / 0.5` has high total but zero disagreement — genuinely ambiguous, and no amount of extra data helps. An input where members say `1/0`, `0/1`, `1/0`, `0/1` has the same average but large disagreement — the model is out of its depth, and this case belongs in the next labelling batch or in front of a human. ## Choosing between them Ask what you can afford and what you need it for. If the model is already trained, dropout is already present throughout the trunk, and you need a rough triage signal: MC dropout, with an honest note that it is a lower bound on disagreement. If the uncertainty gates a consequential decision or you need shift detection you can defend, budget for an ensemble — and if serving M models is impossible, train the ensemble anyway to establish the reference, then check how much of its behaviour a cheaper single-model proxy actually recovers. Whatever you choose, validate it on data the model has genuinely never seen: an uncertainty method that has not been shown to fire on held-out shifted data is decoration.