liveBook · Manning

Independence of mechanism

As explained in Causal AI

Independence of mechanism is the causal principle that a mechanism generating one variable is invariant to changes in other, distinct mechanisms or in the variable's cause distribution.

Suppose a temperature model is trained in Kathmandu and used to reason about El Paso. The distribution of cloud cover may differ, but the physical mechanism relating altitude, cloud cover, latitude, and season to temperature may remain the same. Independence of mechanism is the assumption that changing one causal mechanism does not, by itself, change a distinct mechanism. Transfer still requires a correct causal model and enough training data to learn the mechanism that is meant to transfer. section 3.2.1 section 3.2.2

In a causal DAG, each child node has a causal Markov kernel describing how its parents determine it. For C → O, P(O|C) represents the cause-to-outcome mechanism, while P(C) describes how often causes occur. Learning how sunscreen use affects sunburn does not tell you how common sunscreen use is. By contrast, reasoning backward from a sunburn to possible sunscreen use depends on prevalence: P(C|O) is not itself the cause-to-outcome mechanism. section 3.5 section 5.3.1

The distinction has practical consequences:

  • Parameter modularity: separate kernels encode separate mechanisms, so their representations can be changed or trained separately. Their parameter sets can have independent priors, although they are generally dependent in the posterior. section 3.2.1
  • Causal transfer learning, data fusion, and invariant prediction can exploit a mechanism that remains stable when other parts of a system change; that stability must be justified for the domains involved. section 3.2.2
  • In semi-supervised learning, unlabeled features reveal information about P(X), but not directly about P(Y|X). Whether they help predict labels depends in part on the causal direction between features and labels. section 5.3.1 section 5.3.2

A labeling decision

A social-media image classifier may have many unlabeled images but few human-labeled examples of gratuitous violence. Before expecting the extra images to improve label prediction, ask what their distribution teaches about the joint model of images and labels. Independence of mechanism makes this a causal-direction question, not simply a question of how many unlabeled images are available. section 5.3.2

Figure 5.16 In causal learning, the features cause the label. In anti-causal learning, the label causes the features.
figure

Compare features causing labels with labels causing features when assessing whether unlabeled features could improve prediction. section 5.3.2

How Causal AI covers it

How deep it goes
Defines distinct causal kernels, derives parameter modularity, and applies invariance; leaves advanced transfer and fusion methods to references.
Examples
Uses temperature across cities, sunscreen and sunburn, and unlabeled images for violence classification.
What it assumes
Assumes familiarity with DAGs and conditional-probability notation, while explaining distinct mechanisms through concrete cause-and-outcome examples.
What you can do after
Evaluate/understand when a learned mechanism may transfer and whether unlabeled features might improve label prediction.
How it connects
Connects causal DAGs and causal Markov kernels to parameter modularity, causal transfer learning, invariant prediction, and semi-supervised learning.
Cover of Causal AI
liveBookmanning.com
Read Causal AI on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Causal AI. The section links point to the parts of the book it draws on.