Necessary Changes in the Philosophy and Practice of Probability - - PowerPoint PPT Presentation

necessary changes in the philosophy and practice of
SMART_READER_LITE
LIVE PREVIEW

Necessary Changes in the Philosophy and Practice of Probability - - PowerPoint PPT Presentation

Necessary Changes in the Philosophy and Practice of Probability & Statistics William M. Briggs Statistician to the Stars! . matt@wmbriggs.com What is probability? All men are mortal Socrates is man Socrates is mortal


slide-1
SLIDE 1

Necessary Changes in the Philosophy and Practice of Probability & Statistics

William M. Briggs

————– Statistician to the Stars! . matt@wmbriggs.com

slide-2
SLIDE 2

What is probability? All men are mortal Socrates is man Socrates is mortal

slide-3
SLIDE 3

Just half of Martians are mor- tal Socrates is Martian Socrates is mortal

slide-4
SLIDE 4

Pr(A) does not exist! Pr(A|evidence) might exist A does not “have” a distribution Distributions do not exist

slide-5
SLIDE 5

If Pr(A) = limiting relative frequency, then no probability can ever be known. If Pr(A|evidence) is subjective, then Pr(x = 7|x + y = 12) = 1 if I say so.

slide-6
SLIDE 6

Interocitors can take states s1, . . . , sp This is an interocitor This interocitor is in state sj Pr(sj|Interocitors can...) = 1/p :: No sym- metry!

slide-7
SLIDE 7

Jack said he saw a whole bunch of guys There were 12 guys Pr(12|Jack said...) = not too unlikely.

slide-8
SLIDE 8

Because Cause :: form + material + mechanism + direction :: essence + power Pr(Y|cause or determine) = 1 y = tan(θ) · x − g(2v2

0 cos2 θ)−1 · x2

Pr(y|xgvoθ) ∈ {0, 1}

slide-9
SLIDE 9

Chance or randomness are not ontic, thus

  • powerless. No probability model is causal (in-

cluding QM). Every potential must be made actual by something actual (including QM). We have Pr(Y|X), where X is that informa- tion we think or assume is probative of Y— meaning we think X is related to the causal path of Y. If not, pain.

slide-10
SLIDE 10

Hypothesis testing? We cannot derive from Pr(Y|X) = p that Y. Probability is not deci- sion! P-value = Pr(larger ad hoc stat|MΘ, x, θs = 0), which is no way related to Pr(θs = 0|x, MΘ). Pr(larger ad hoc stat|MΘ, x, θs = 0) may be lower!

slide-11
SLIDE 11

Models Bayes is not important: probabaility is. A parameterized model M relates X to Y probabilistically, e.g. µ = β0 + β1x where µ is central parameter of normal used to charac- terize uncertainty in some y. “Priors” a real distraction: start finite!

slide-12
SLIDE 12

With rare exceptions, parameters are of no interest to man nor beast. ˆ Y = f(X, ˆ θ(Mθ)) ignores uncertainty, and makes a decision. Pr(θ|data, Mθ) only about unobservable pa- rameters.

slide-13
SLIDE 13

We want this: Pr(Y|new X, data, M), where the data are

  • ld values of Y and X, and M are the argu-

ments that led to a (parameterized) model; the parameters having been integrated out. This—and only this—captures the full un- certainty, given M. Prediction!

slide-14
SLIDE 14

Every model—neural net, statistical, machine learning, artificial intelligence, anything—can fit into the Pr(Y|XDM) schema. What dif- ferentiates them is usually a matter of ad hoc complexity and form—and a building in

  • f decision.
slide-15
SLIDE 15

Demystifying “learning” ANNs, GANs, Deep this-and-thats, etc. = parameterized non-linear regressions Learning = estimating parameters Extracting features = f(input data)

slide-16
SLIDE 16

There is no such thing as unsupervised learn- ing. Every algorithm does exactly what it is de- signed to do, and therefore gives correct results— conditional on the algorithm. Not all probability is quantitative, and not all algorithms live in machines.

slide-17
SLIDE 17

Monte Carlo — The place to lose your money, and your way. Jaynes: “It appears to be a quite general principle that, whenever there is a random- ized way of doing something, then there is a nonrandomized way that delivers better per- formance but requires more thought.”

slide-18
SLIDE 18
slide-19
SLIDE 19
slide-20
SLIDE 20

Image D with possible signal + background Pr(dij|MB) ∼ Poisson(λB) Pr(dij|MS+B) ∼ Poisson(λS + λB) Pr(dij|MS+B, MB) = pP(λB)+(1−p)P(λS + λB) Pr(MS+B|dij)

Guglielmetti et al., 2002, Mon. Not. R. Astron. Soc

slide-21
SLIDE 21

Roe et al.

slide-22
SLIDE 22

Skill: Obs S B Mod S 3 5 B 5 87 Super machine neural deep-learning boosting forest machine boasts 90% accuracy! Skill and calibration curves, not ROC.

slide-23
SLIDE 23

Model-based vs. verification-based uncertainty; verify “features”. All uncertainty carried through to the bitter end. In the absence of knowledge of cause, all probabilistic models will classify imperfectly.

slide-24
SLIDE 24

“This is not not a statis- tics text, it is not a treatise

  • n philosophy of science or
  • logic. This work is like noth-

ing I have seen before, an excellent combination of the above, indeed the ‘the soul

  • f modeling, probability ...’,

presented with passion and accessible to everybody.” “It is a deep philosophical treatment of probability writ- ten in a plain language and without the interference of unnecessary math.”