Skip to content

Latest commit

 

History

History
136 lines (107 loc) · 5.29 KB

abduct.md

File metadata and controls

136 lines (107 loc) · 5.29 KB
Raw

 home :: syllabus :: timetable :: groups :: moodle(591, 791) :: video tbd :: © 2021

Explaining Explanation

Explanation and Ethics

"Models induced from data must be liable as liability will likely soon become a legal requirement.

  • "Article 22 of the General Data Protection Regulation (GDPR) sets out the rights and obligations of the use of automated decision making.
  • "Noticeably, it introduces the right of explanation by giving individuals the right to obtain an explanation of the inference/s automatically produced by a model, confront and challenge an associated recommendation, particularly when it might negatively affect an individual legally, financially, mentally or physically. "

"By approving this GDPR article, the European Parliament attempted to tackle the problem related to the prop agation of potentially biased inferences to society, that a computational model might have learnt from biased and unbalanced data."

So explanation and ethics are linked.

Lets talk about theories of explanation.

Background

  • Induction: (learn where we can run) Give lots of examples of (premise,conclusion), make a rule
  • Deduction: (run forward) Give rules and a premise, make a conclusion
  • Abduction : (run backwards) Given a rule and a conclusion, assume a premise

Abduction == "guessing" and guesses can be wrong

  • Rule1: grass is wet cause it rained
  • Rule2: grass is wet cause of the sprinkler
  • Conclusion: grass is wet:
    • what is our premise?

Abduction == the logic of guessing

  • Work backwards over the rules.
  • Make assumptions
  • Group together consistent assumptions
  • Find the base assumptions (those that decedent on nothing else)
  • Monitor for evidence for the base
  • Jump between assumption sets based on that evidence

JTMS vs ATMS

  • JTMS: only every hold one world of consistent believes
    • Spend too much time jumping to another world?
  • ATMS: weave all the worlds of belief together, in indexed by the base assumptions.
    • All known ATMS exponential runtime

Next

Notes on learner types

Generation methods

  • Expert systems: lots of rules, often built manually via expert intervie
    • These days, ES usually augmented with leaners

Learner goals:

  • regression: target class is numeric
    • multi-regression: N target classes
    • multi-objective: targets have weights (things to avoid or leap towards)
  • classification: target class is symbolic
  • REcommender systems:
    • not what is but
    • what to change.
    • think "recommender system" is analogous to "optimizer"
      • but for symbolic systems.

Learner outputs:

  • Bayesian networks
    • directed graph
    • child parent beliefs are initialized with a prior
    • then updated with evidence that flows from their parents
    • a modeling framework that allows experts to
      • initialize a model with their beliefts
      • then udpate that belief when new evidence arrives.
  • Rules: combinations of ranges (e.g. age over 60 and name == tim) that denote regions (technically, a volume in N-d space) that selects for some desired goal
  • Neural networks
    • Rules report regions in the data, using the raw attribute names
    • SVMs add in new names (from the kernel, see below)
    • Neural nets ignore all the raw names
      • write input into layers of nodes, each of which might connect to the other
      • learns weights on the edges between nodes
    • No "grandmother cell".
  • Support vectors:
    • Given data with a boundary between regions, support vectors are the examples closest to the boundary (border guards).
    • 10,000s of examples may only have a few dozen, a few hundred, border guards
    • Support vector machines (SVM): devices to find the border guards that use the kernel trick:
      • when data can't be neatly separated, add another dimension that better divides .
        • which kernel to use? hyper-parameter optimization!
    • BTW, support vectors are not rules
      • they are an instance-based representation