Introduction A taste of forcing Higher order arithmetic tuned The forcing transformation The forcing machine Conclusion

Alexandre Miquel

Description

Niveau: Supérieur
Introduction A taste of forcing Higher-order arithmetic (tuned) The forcing transformation The forcing machine Conclusion A computational analysis of the proof transformation by forcing Alexandre Miquel June 22th, 2010 – TypiCal seminar – Ecole polytechnique

• ¬¬-translation cps


• classical realizability


• study underlying program


• program transform


• realizability generalizes standard


• transform stack


• forcing

Sujets

• École polytechnique
• Miquel
• Forcing

Informations

• Publié par: profil-nechor-2012
• Nombre de lectures: 24
• Langue: English

Extrait

ht 1 ne.sd..iw/wwpt/:matton/,solnfr/∼[e-mail] diw.ww//p:tt 2 h fr [e-mail]/.frl∼a/,ot ide.snf./rf∼abhc s.fr 3 http://www.

Matthieu Solnon (ENS) 1 Sylvain Arlot (CNRS) 2 Francis Bach (INRIA) 3 ´ Equipe Sierra ´ Laboratoired’Informatiquedel’EcoleNormaleSup´erieure (CNRS/ENS/INRIA UMR 8548)

Multi-task Regression using Minimal Penalties

rf.sncianfr/,@echbas.

Intro

duction

aux

me´tho

de

multi-tˆaches

Intro

duction

aux

me´tho

de

multi-taˆches

I tnnieremnead´dpepasnsontssai´ecetsetnereensehcaˆ nttaormp ff´dispLeesnt

Cadrethe´orique

I Design :npoints (X i ) i n =1 ∈ X n . I Fest un RKHS. I fonctionsdere´rgseisno:pfonctions (f j ) j p =1 ∈ F p . I Observations :Y i j =f j (X i ) +ε i j ,∀i(ε j i ) p j=1 ∼ N(0,Σ). I Design fixe : Le but est d’estimer (f j (X i )) i,j . I  ( minimiserPerte quadratique :np) −1 P i,j (f j (X i )−g j (X i )) 2 .

.

Cadreth´eorique

I Design :npoints (X i ) i n =1 ∈ X n . I Fest un RKHS. I tcnosnoi´redergeiossn:fpfonctions (f j ) p ∈ F p . j=1 I Observations :Y i j =f j (X i ) +ε i j ,∀i(ε i j ) p j=1 ∼ N(0,Σ). I  LeDesign fixe : but est d’estimer (f j (X i )) i,j . I  ( minimiserPerte quadratique :np) −1 P i,j (f j (X i )−g j (X i )) . 2

Important Lespnoptnssecaehseˆtrentiff´edndtipe´eannds.te´nsaseceriasneme

Proc´eduresimple-taˆche

L’estimateurridgeenr´egression.

I ehcaˆteuqahcruPojminimiser n −1 X (Y j i −g(X i )) 2 +λ j kgk 2 F . i

I ertrapue`mactlendioL´easλ j rAolatdni´eedanst´e´etude´a Bach (2011). j I naecaviriutudrb:ealipnclaerimstborPirpeme`lε. Utilisationdespe´nalite´sminimales(Birge´andMassart,2007; Arlot and Massart, 2009).

G´e´lisationdelape´nalisationsimple-tˆache nera

´ Equivalenta`pimplmessbl`epro.tsanndepe´dnisehcaˆt-e

Minimiser sur (g j ) p j=1 ∈ F p (np) −1 X (Y i j −g j (X i )) 2 + X λ j kg j k 2 F i,j j

Ge´ne´ralisationdelape´nalisationsimple-taˆche

´ Equivalenta`ps.ntseniaˆhcnead´dpeemepsrobl`le-tsimp M=    λ 1 . . . λ p   

Minimiser sur (g j ) p j=1 ∈ F p (np) −1 X (Y i j −g j (X i )) 2 + X λ j kg j k 2 F i,j j

Ge´ne´ralisationdelap´enalisationsimple-tˆache

´ Equivalenta`p-tˆachesind´epenadtn.sborpme`lisseelpm M=    λ 1 . . . λ p   

Minimiser sur (g j ) j p =1 ∈ F p (np) −1 X (Y i j −g j (X i )) 2 + X λ j kg j k 2 F i,j j | {z } P j,k M j,k hg j ,g k i:=kgk 2 M

G´ene´ralisationdelap´enalisationsimple-taˆche

Ge´neralisationmulti-taˆches,lesf j stim´eesptnoesulsen ´ ind´ependamment.

M∈ S p ++ (RvgaEe`irlamisi)()5002,.lateuoine

Minimiser sur (g j ) j p =1 ∈ F p (np) −1 X (Y j i −g j (X i )) 2 + X M j,k hg j ,g k i i,j j ,k | {z } :=kgk 2 M

Gene´ralisationdelape´nalisationsimple-taˆche ´

Ge´ne´ralisationmulti-tˆaches,lesf j plusestim´eesenostn epe am ind´ nd ment.

M∈ S p ++ (R(s)ilimreai.,aletounigeEv`a)5002

Minimiser sur (g j ) p j=1 ∈ F p (np) −1 X (Y i j −g j (X i )) 2 + X M j,k hg j ,g k i i,j j ,k | {z } :=kgk 2 M

P j,k M j,k hg j g k i:=kgk 2 M rusSinutRnHK´dfieF p . , b =⇒tamitse´einrleureaif M =A M y