Actual causality for everything in life�thanks to Joe Halpern�
Hana Chockler
Department of Informatics
King’s College, London
IBM Research
© 2010 IBM Corporation
© 2010 IBM Corporation
1
Actual Causality
A theoretical concept from AI
Extends causal counterfactual reasoning
Enables us to reason causally
about a specific situation
that occurred in the past
©Halpern & Pearl, 2001
+
…
@Joe Halpern – many definitions and results
Then Joe spent a sabbatical at the Hebrew University –
a key moment in my academic journey (I was a PhD student)
What causes a system to satisfy a specification? ACM TOCL 2008
© 2010 IBM Corporation
2
Hana Chockler, Joseph Y. Halpern:�Responsibility and Blame: A Structural-Model Approach.
IJCAI 2003 and JAIR 2004
Quantification of causality,
allowing to rank causes by importance
To this day my most cited paper (I just checked again)
… also Joe’s 16th most cited paper (which is not saying much)
Allows to focus on the most influential causes – very important for large practical applications
© 2010 IBM Corporation
3
Hana Chockler, Joseph Y. Halpern, Orna Kupferman:
What causes a system to satisfy a specification? ACM TOCL 2008
Took a very long time to get published…
© 2010 IBM Corporation
4
Formal Verification
A huge and difficult
to understand system M:
A correctness specification φ
Does M satisfy φ?
no
counterexample
yes
the system is correct!
© 2010 IBM Corporation
5
Does M satisfy φ?
no
counterexample
yes
the system is correct!
Do we understand the counterexample?
A correctness specification φ
A huge and difficult
to understand system M:
Do we really believe it?
Formal Verification
© 2010 IBM Corporation
6
Does M satisfy φ?
no
counterexample
yes
Do we understand the counterexample?
A correctness specification φ
A huge and difficult
to understand system M:
Do we really believe it?
Everything is actual causality!
© 2010 IBM Corporation
7
A huge and difficult
to understand system M:
A correctness specification φ
Does M satisfy φ?
no
counterexample
yes
the system is correct!
Hana Chockler, Joseph Y. Halpern, Orna Kupferman:
What causes a system to satisfy a specification? ACM TOCL 2008
Everything is actual causality!
© 2010 IBM Corporation
8
A huge and difficult
to understand system M:
A correctness specification φ
Does M satisfy φ?
no
counterexample
yes
the system is correct!
Ilan Beer, Shoham Ben-David, Hana Chockler, Avigail Orni, Richard J. Trefler:
Explaining Counterexamples Using Causality. CAV 2009 (and FMSD)
My second most cited paper
© 2010 IBM Corporation
9
Explaining counterexamples using causality�(Red Dots)�part of IBM tool
A timing diagram of a buggy hardware execution
φ = always ((!START and !STATUS_VALID and END) ->
next(!START Until (STATUS_VALID and READY))
causes marked as red dots
© 2010 IBM Corporation
10
Black-box reasoning about AI models’ decisions
Intervene on inputs
inputs
outputs
Input transformation
Observe the outputs
Reason about the way AI makes its decisions
One-node causal model
© 2010 IBM Corporation
11
Ongoing theoretical work with Joe: harm, fairness, neural networks
… journal version under submission.
© 2010 IBM Corporation
12
Writing papers with Joe
© 2010 IBM Corporation
13