Theoretical Impediments to Machine Learning: With Seven Sparks from the Causal Revolution | Judea Pearl
The claim of this paper is very clear:
- "Current machine learning systems operate, almost exclusively, in a statistical, or model-blind mode, which entails severe theoretical limits on their power and performance." (1)
- "Such systems cannot reason about interventions and retrospection and, therefore, cannot serve as the basis for strong AI." (1)
- "To achieve human level intelligence, learning machines need the guidance of a model of reality, similar to the ones used in causal inference." (1)
Statistical machine learning can be seen as "blind" evolution, whereas the success of humans seems to be "their ability to sketch and store a mental representation of their environment, interrogate that representation, distort it by mental acts of imagination and finally answer "What if?" kind of questions." (1)
Interventional: what if I act?
Retrospective/explanatory: what if I had acted differently?
Due to the mathematization of counter factuals, Pearl argues we should equip machines with causal reasoning tools.
I want to focus on the main hierarchy in used explanations:
Three layer causal hierarchy
Questions at any given level require at least information from that level or higher to be answered. But you cannot answer counterfactual questions based on intervention; neither can you answer an interventional question based on association.
Questions at lower points in the hierarchy are subsumed by higher ones. "For example, the interventional question, What will happen if we doulbe the price? can be answered by asking the counterfactual question: What would happen had the price been twice its current value?"
Associational questions are really interventional questions where we ignore the action.
Association requires no causal information, pure statistical correlation.
Intervention "involves not just seeing what is, but changing what we see" (2)
1. Association
P(y|x).
"What does a symptom tell me about a disease?" (2) "What does a survey tell us about the election results?" (2)
The main mode here is observation.
2. Intervention
P(y|do(x),z).
"What if I take aspirin, will my headache be cured?" (2)
Main mode here is doing.
(Heavily used in scientific experimentation).
3. Counterfactuals
P(y_x|x',y'), i.e. what is the probability of event Y=y given that X=x, if instead we would have observed x' and y'.
"For example, the probability that Joe's salary would be y had he finished college, given that his actual salary is y' and that he had only two years of college." (3)
"Was it the aspirin that stopped my headache?" (2)