causal representation learning

Classical causal inference assumes the variables are already given, neat columns labeled treatment, dose, outcome. But a model that sees raw pixels or sensor streams has no such columns; the causal factors are tangled inside the high-dimensional input. Causal representation learning aims to recover, from that raw data, a small set of latent variables that behave like genuine causal units, each one independently manipulable, so that downstream reasoning can intervene on 'object position' rather than on pixel 4,832.

The goal is to learn an encoder whose latent variables correspond to the true causal factors up to tolerable ambiguities, together with the causal graph among them. The central obstacle is identifiability: without extra structure, infinitely many latent parameterizations explain the same observations equally well, so the recovered factors are arbitrary. Progress comes from leveraging signals that pin the representation down, multiple environments, interventional or temporal data, sparsity of the mixing, or known mechanism changes, turning identifiability into theorems rather than hopes.

It is best read as the meeting point of disentanglement and causal discovery, scaled up to perception. The promise is models that generalize under distribution shift and support counterfactual queries about the world rather than about features; the present reality is that strong identifiability guarantees still require strong, often interventional, assumptions, and purely observational disentanglement is known to be impossible in general.