Causal abstraction

As explained in Causal AI

Causal abstraction is the creation of high-level variables from lower-level variables while preserving causal assumptions about the data-generating process.

Causal abstraction creates high-level variables from lower-level data while preserving causal assumptions about the data-generating process. It chooses variables at the level where meaningful causal questions can be asked. section 3.1.1

Raw data may be too detailed for a causal model. A pixel-level DAG for an image is unwieldy and rarely answers useful questions such as the causal effect of one pixel on another. section 5.1.2

Build a causal abstraction by:

• Identifying the high-level variables required by the causal question. section 3.1.1 section 5.1.2
• Mapping low-level observations to those variables through domain knowledge or analytical methods such as mathematics and logic. section 3.1.1
• Checking that the mapping preserves assumptions about the underlying data-generating process. section 3.1.1

From pixels to causal variables

A medical-imaging dataset may contain pixel matrices while the causal model uses a binary Tumor variable. In image datasets such as MNIST and TMNIST, a compact causal model is more useful than treating all 784 pixels as separate causal nodes. A learned representation Z is not automatically a good causal abstraction: reconstruction-focused autoencoders preserve information broadly, whereas causal representations should retain causal information and ignore non-causal information. section 3.1.1 section 5.1.2 section 5.2.5

Figure 5.7 A plate model representation of the causal DAG. Plates represent repeating variables, in this case 28 × 28 = 784 pixels. X_j is the j^th pixel.

Use the plate model to inspect how 784 repeating pixel variables can be represented without making the causal model operate at an impractical pixel-by-pixel level. section 5.1.2

How Causal AI covers it

How deep it goes

Explains the definition, motivation, mapping methods, and causal-preservation requirement, but defers formal identification procedures.

Examples

Uses transportation variables, tumor labels from medical images, MNIST and TMNIST pixels, plate models, and latent representation Z.

What you can do after

Apply it to derive high-level variables from low-level observations while preserving causal assumptions about the data-generating process.

How it connects

Connects causal abstraction with causal DAGs, causal representation learning, domain knowledge, interventions, and semi-supervision.

liveBookmanning.com

Read Causal AI on liveBook Get the book at manning.com

Related concepts

• causal representation learningcontrasts with · Causal representation learning is a representation-learning approach that seeks latent representations encoding causally meaningful properties of modeled objects or processes.
• causal graphical modelused by · A causal graphical model is a probabilistic or structural model whose graph represents causal relationships and whose mechanisms support reasoning about interventions.

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.