Artificial Intelligence: Causal learning
The Causal AI Revolution
•“You are smarter than your data. Data don’t understand cause and effect; humans do.” Judea Pearl
•“Machines’ lack of understanding of causal relations is perhaps the biggest roadblock to giving them human-level intelligence.” Judea Pearl
•“Causality is very important for the next steps of the progress of machine learning.” Yoshua Bengio, Turing Award winner and “Godfather of Deep Learning”
DeepMind: “Causal models have a unique combination of capabilities that could enable a deeper understanding of complex systems and allow us to better align decision systems with society’s values.”
IBM: Causal modeling is crucial to the effectiveness and trust of AI by ensuring that actions lead to intended outcomes. We study the inference of causal effects and relationships, as well as the application of causal thinking to out-of-distribution generalization, fairness, robustness, and explainability.
Microsoft: Machine learning based on correlational pattern recognition is insufficient for robust predictions and reliable decision-making. New approaches based on principles of causal reasoning provide a promising path forward. Causality methods promise robustness to exogenous changes and accurate modeling of counterfactual or “what-if” scenarios that are core to scientific experimentation, understanding, and decision-making.
Ladder of causation:
This three-level abstraction (see image) helps identify the key steps to move from an artificial narrow intelligence (ANI) to an artificial general intelligence (AGI), meaning the machine would be able to think like a human being.
Associational- rung1
- Bayesian network: P(y|x)
- Typical Activity ( Seeing)
- Typical Question
- What is? How would seeing X change my belief in Y?
- Example:
What does a symptom tell us about the disease?
Machine Learning (Supervised / Unsupervised Learning)
Interventional — rung 2
- Causal Bayesian network (causal graph): P(y|do(x), c)
- Typical Activity ( Doing)
- Typical Question …. What if? What if I do X?
- Example
What if I take aspirin, will my headache be cured?
- Machine Learning (Reinforcement Learning)
Counterfactual Imagining- rung 3
- Functional Bayesian network (functional dependencies): P(Yx|x’ , y’)
- Typical Activity ( Imagining)
- Typical Question … Why? What if I had acted differently?
- Example
Was it the aspirin that stopped my headache?
- Machine Learning …Not Yet!
Example:
Answer the questions of the three ladder’s rungs:
1- Association (what one fact tells us about the other, e.g If the prisoner is dead, does that mean the order was given?)
2- Intervention (What is the soldier A decided on his own initiative to fire, without waiting for C's command?
Get Seham Nasr’s stories in your inbox
Join Medium for free to get updates from this writer.
3- Counterfactual (If the prisoner is found dead, what is A don’t shoot is it possible he is not dead?
Causal Model
Causal models are mathematical models representing causal relationships within an individual system or population.
A causal diagram is a directed graph that displays causal relationships between variables in a causal model. A causal diagram includes a set of variables (or nodes). Each node is connected by an arrow to one or more other nodes upon which it has a causal influence. An arrowhead delineates the direction of causality. “Wikipedia”
Causal Model elements
Causal models have formal structures with elements with specific properties.
Junction patterns, chain, fork, collider, instrumental variable, mendelian randomization, and node types: (Mediator and Confounder)
Measure of causality
•Granger 1969 (a statistical concept of causality that is based on prediction)
X1 “G-causes” X2, then past values of X1 should contain information that helps predict X2 above and beyond the information contained in past values of X2 alone.
•Rubin Causal Model (RCM) 1974 ( the statistical analysis of cause and effect based on the framework of potential outcomes)
Consider whether changing treatment on X1 will result in different outcomes for X2.
From Bayesian Network to Causal Diagram
- Bayesian networks hold the key that enables causal diagrams to interface with data.
•Bayesian reading:
The missing arrow between A and C means that A and C are independent once we know the value of B.
•Causal reading:
B listens to A only, C listens to B only, and A listens to no one; it is determined by external forces that are not part of our model.
What can we use Bayesian Networks for?
The probabilities of variables in Bayesian Networks update as observations are added to the model. This is useful for inference or counterfactuals, and for predictive analytics. Metrics can help us understand the strength of relationships between variables.
•The sensitivity of nodes to changes in observations of other events can be used to assess what changes could lead to what effects;
- The active trail of a target node identifies which other variables have any effect on the target.
Structure causal model
Structure Causal Model (SCM)
- Structural causal models (Functional Bayesian networks ) represent causal dependencies using graphical models that provide an intuitive visualization by representing variables as nodes and relationships between variables as edges in a graph.
- SCM is a tuple of the following sets:
1. A set of Endogenous (V).
2. A set of Exogenous (U).
3. Variables connected by a set of functions (F)
that determine the values of the
variables in V based on the values
of the variables in U.
