The diagram below summarises relations between different types of ML
stateDiagram
KM : K-Means
LinR : Linear Regression
LogR : Logistic Regression
UL : Unsupervised Learning
PCA: Principal Component Analysis
RL : Reinforcement Learning
SL : Supervised Learning
SVM : Support Vector Machine
QL : Q-Learning
RL --> ML
SL --> ML
UL --> ML
Classification --> SL
Regression --> SL
SVM --> SL
KM --> UL
PCA --> UL
LinR --> Regression
LogR --> Regression
Binary --> Classification
KNN --> Classification
Multiclass --> Classification
QL --> RL
Uncertainty in deep learning
Reading suggestion: Uncertainty in Deep Learning: Epistemic Uncertainty and Bayes by Backprop. The method itself is from Blundell, Cornebise, Kavukcuoglu and Wierstra, Weight Uncertainty in Neural Networks, ICML 2015, arXiv:1505.05424.
Predictive uncertainty splits into two kinds:
- Aleatoric uncertainty is noise in the data-generating process itself, such as measurement noise or class overlap. It does not shrink with more data, and is modelled by making the network output a distribution, for example a mean and a variance.
- Epistemic uncertainty is uncertainty about the model parameters given finite data. It shrinks as data accumulates and is largest far from the training data. A network trained to a single point estimate cannot represent it.
A Bayesian neural network represents epistemic uncertainty by a posterior over the weights, , and predicts with . The posterior is intractable, so Bayes by Backprop approximates it variationally by a factorised Gaussian , with to keep it positive and . One minimises the variational free energy, the negative evidence lower bound,
which equals , so minimising it pulls towards the true posterior (see Kullback-Leibler divergence). Gradients come from the reparameterisation with : a Monte Carlo estimate of from a few samples of is differentiable in and is trained by ordinary backpropagation. With minibatches, the KL term is split across the batches of an epoch. The paper uses a scale mixture of two zero-mean Gaussians as prior.
At test time one samples weights and averages, . The spread of the individual predictions across samples measures the epistemic uncertainty, whereas the average predicted noise measures the aleatoric part. The cost is twice the parameters (a mean and a scale per weight) and a mean-field posterior that ignores correlations between weights and tends to underestimate uncertainty.