设计图模型时面临着以下选择:
Template based vs specific
Directed vs undirected
Generative vs discriminative
Hybrid(混合) model是很常见的,有时候会同时有Template based部分和非Template level的部分。:
Template-based | Hybrid | Specific |
image segmentation fairly small number of variable types | fault diagnosis elements unique and shared | medical diagnosis large number of unique variables |
Generative | Discriminative | |
no predetermined task(task shifts) not predetermined set of variables as input and disease as outputs medical diagnosis task every patient presents differently(symptoms&tests) easier to train in certain regimes(when data is not fully labeled) | a particular prediction task better solved by having richly expressive features avoid dealing with correlations between features ->high performance model | |
Directed | undirected | |
arrows correspond to causality |
latent model:简化模型结构
Discriminative models directly model P(Y|X) and avoid dealing with the joint distribution. Generative models handle missing data or partially labeled data better, so they are preferable if obtaining labels is hard. High-dimensional variables are in general more difficult to model, and discriminative models avoid this by conditioning on them. Generative models can be used to generate new observable data.
Causal ordering不仅更intuitive,而且使模型更加sparse,更容易parameterize。