8 - Seven Sparks from the Causal Revolution - Pearl (2018)
Dit paper bespreekt de opkomst en voordelen van causale modellen in AI:
• Wat zijn de drie niveau's van causale modellen? In welk niveau past het model van
Maguire uit het vorige artikel dat jullie hebben gelezen?
• Wat kunnen causale modellen wel dat non-causale modellen niet kunnen?
• De auteurs pleiten ervoor dat causale modellen erg belangrijk zijn voor de verdere
ontwikkeling van AI. Voor deze causale modellen is representatie volgens de auteurs
cruciaal. Maar het paper 'Intelligence without representation' dat jullie hebben gelezen,
pleitte er juist voor dat AI's zonder representatie cruciaal zijn voor de verdere
ontwikkeling van AI. Hoe kan dit? Werken de auteurs van deze papers misschien met
verschillende definities van AI?
- to achieve human level intelligence, learning machines need the guidance of a model of
reality, similar to the ones used in causal inference tasks
→ causal inference is the term used for the process of determining whether an observed
association truly reflects a cause-and-effect relationship
- current machine learning systems operate in a statistical, or model-free mode: they improve
their performance by optimizing parameters over a stream of sensory inputs received from the
environment
- humans can imagine alternative hypothetical environments for planning and learning (and
animals cannot)
, 1. association: invokes purely statistical relationships, defined by the naked data (bottom)
→ P(y|x)=p: the probability of event Y=y given that we observed event X=x is equal to p
2. intervention: involves not just seeing what is, but changing what we see (middle)
→ P(y|do(x),z): the probability of event Y=y given that we intervene and set the value of
X to x and subsequently observe event Z = z
→ interventional expressions cannot be inferred from passive observations alone
3. counterfactuals: necessitating retrospective reasoning (top)
→ P(yx|x’,y’): the probability that event Y = y would be observed had X been x, given that
we actually observed X to be x’ and Y to be y’
- the hierarchy is directional and the top level is the most powerful
- each layer has a syntactic signature that characterizes the sentences admitted into it
- the hierarchy explains why statistics-based machine learning systems are prevented from
reasoning about actions, experiments and explanations
→ it also informs us what extra-statistical information is needed, and in what format, in order to
support those modes of reasoning
- the rules of cause and effect were denied the benefits of mathematical analysis (not anymore)
→ a mathematical language is developed for managing causes and effects (causal revolution)
- structural causal models (SCM): the mathematical framework that led to the revolution
1. graphical models: serve as a language for representing what we know about the world
2. structural equations: serve to tie 1 and 3 together in a solid semantics
3. counterfactual and interventional logic: help us to articulate what we want to know
Dit paper bespreekt de opkomst en voordelen van causale modellen in AI:
• Wat zijn de drie niveau's van causale modellen? In welk niveau past het model van
Maguire uit het vorige artikel dat jullie hebben gelezen?
• Wat kunnen causale modellen wel dat non-causale modellen niet kunnen?
• De auteurs pleiten ervoor dat causale modellen erg belangrijk zijn voor de verdere
ontwikkeling van AI. Voor deze causale modellen is representatie volgens de auteurs
cruciaal. Maar het paper 'Intelligence without representation' dat jullie hebben gelezen,
pleitte er juist voor dat AI's zonder representatie cruciaal zijn voor de verdere
ontwikkeling van AI. Hoe kan dit? Werken de auteurs van deze papers misschien met
verschillende definities van AI?
- to achieve human level intelligence, learning machines need the guidance of a model of
reality, similar to the ones used in causal inference tasks
→ causal inference is the term used for the process of determining whether an observed
association truly reflects a cause-and-effect relationship
- current machine learning systems operate in a statistical, or model-free mode: they improve
their performance by optimizing parameters over a stream of sensory inputs received from the
environment
- humans can imagine alternative hypothetical environments for planning and learning (and
animals cannot)
, 1. association: invokes purely statistical relationships, defined by the naked data (bottom)
→ P(y|x)=p: the probability of event Y=y given that we observed event X=x is equal to p
2. intervention: involves not just seeing what is, but changing what we see (middle)
→ P(y|do(x),z): the probability of event Y=y given that we intervene and set the value of
X to x and subsequently observe event Z = z
→ interventional expressions cannot be inferred from passive observations alone
3. counterfactuals: necessitating retrospective reasoning (top)
→ P(yx|x’,y’): the probability that event Y = y would be observed had X been x, given that
we actually observed X to be x’ and Y to be y’
- the hierarchy is directional and the top level is the most powerful
- each layer has a syntactic signature that characterizes the sentences admitted into it
- the hierarchy explains why statistics-based machine learning systems are prevented from
reasoning about actions, experiments and explanations
→ it also informs us what extra-statistical information is needed, and in what format, in order to
support those modes of reasoning
- the rules of cause and effect were denied the benefits of mathematical analysis (not anymore)
→ a mathematical language is developed for managing causes and effects (causal revolution)
- structural causal models (SCM): the mathematical framework that led to the revolution
1. graphical models: serve as a language for representing what we know about the world
2. structural equations: serve to tie 1 and 3 together in a solid semantics
3. counterfactual and interventional logic: help us to articulate what we want to know
