The Unified Causal AI Pipeline

• Sara Magliacana - prof at UvA

Intro

• predict the effect of actions and decide effective policies - and how
• dl - can learn representations from unstructured data
• predict effects of interventions
  • causal variables not observed directly
  • no labels
  • just high level observations
• tasks
  • identify causal variables from observations
  • learn causal relations between them

Recovering Causal Variables

• theoretical guarantees
• without supervision, cannot identify exact causal variables - but up to equivalence class (aka not perfectly)
  • somehow seems like trying to recover some variables from a tangled mess of the ground truth
  • 1-1 map?

Dynamic Bayesian Networks

• Markov Chain
• stationarity : transition model (edge) are same across pairs of time steps
• no instantaneous effects : no edges between vars at same timestep
• extension of Bayesian networks
• environment : AI Tutor

Temporal Intervened Sequences

• learn underlying causal process from temporal seq of high dim data (images) ${X^t}_{t-1}^T$
• Assume latent causal process in Dynamic Bayesian Networks with K multidim causal variables
• Assume interventions can happen, observe binary targets $I_i^t$ and $I_i\to C_i$
• targets are 1-1 with causal variables
• model with latent variable
• paper : CITRIS (phillip lippe, sara magliacane)
  • stochastic intervention : we dont know where the ball will be
  • simplified identification proof
  • minimal causal variables
    • part of the variable that is varying
    • identify them up to invertible component wise transforms if $I_i^t$ up to $I_j^t$
  • CITRIS - VAE
    • cant really use pretrained autoencoder since you assume its disentangling
  • normalizing flows : CITRIS NF
    • use a pretrained autoencoder
• BISCUIT - causal representation learning from binary interactions
  • extension of TRIS
  • effect encoded in binary interaction variables
  • assume action has to affect causal variables in a binary way
  • More BISCUIT - VAE and NF
• Environments - CausalWorld, iTHOR
• dynamic interaction map
• CRL World models

Why

• causal representation learning
• causal discovery
• inspired ml/rl research
• explainable ai
• causal effect estimation

Future

• BISCUIT + DAFT RL