Machine Learning Theory Tübingen University, WS 2016/2017 Lecture 11

Transcription

1 Machie Learig Theory Tübige Uiversity, WS 06/07 Lecture Tolstikhi Ilya Abstract We will itroduce the otio of reproducig kerels ad associated Reproducig Kerel Hilbert Spaces (RKHS). We will cosider couple of easy examples to get some ituitio. Next we will motivate a importace of RKHS for machie learig by cosiderig represeter theorem, which we will also prove. Fially, we will cosider several scearios where represeter theorem actually becomes very useful. Blue colour will be used to highlight parts appearig i the upcomig homework assigmets. Reproducig kerels ad RKHS Cosider ay iput space X. We will call a fuctio k : X X R a kerel or a reproducig kerel if it is symmetric k(x, y) = k(y, x) for all x, y X ad positive defiite, which meas N, α,..., α R, x,..., x X, α i α j k(x i, x j ) 0. j= It ca be show that k defies a uique Hilbert space of real-valued fuctios o X, such that:. Fuctios k(x, ): X R for all X X belog to ;. f(x) = f, k(x, ) Hk for ay f ad X X. (the reproducig property) Throughout this lecture we will write, Hk to deote the ier product of ad Hk the orm iduced by, Hk. The space is commoly kow as Reproducig Kerel Hilbert Space (RKHS). Notice that, because is a vector space, all the fuctios of the form α i k(x i, ) i also belog to for ay fiite sequece of real coefficiets α, α,... ad poits X, X,... from X. A Hilbert space is a vector space with a ier product, which is complete with respect to the orm iduced by the ier product.

2 Feature map Aother way to look at this costructio is to say that all the poits X of the iput space X are beig mapped to the elemets k(x, ) of the Hilbert space. Moreover, for ay two poits X, X X the ier product betwee their images is equal to k(x, ), k(x, ) Hk = k(x, X ). This observatio leads to very useful implicatios. It turs out that, o matter what the iput space X is (R d, a set of strigs, a set of graphs, pg pictures,... ), oce we come up with a kerel fuctio k defied over X we simultaeously get a way to embed the whole X ito a Hilbert space. This embeddig is very useful, sice the RKHS has a very ice geometry: it is a vector space with a ier product, which meas we ca add its elemets with each other ad compute distaces betwee them somethig which was ot ecessarily possible for elemets of X (thik of a set of graphs). Next we cosider two simple examples of kerels k ad correspodig RKHS:. Liear kerel Cosider X = R d ad defie k(x, y) := x, y R d. First of all, let s check that this is ideed a kerel. It is obviously symmetric. Also ote that j= α i α j k(x i, x j ) = j= α i α j x i, x j R d = α i x i 0. R d Thus, k is ideed a kerel. It is ow easy to see that all the homogeeous liear fuctios of the form f(x) = w, x R d, w R d () belog to the RKHS. As well as all their fiite liear combiatios. Actually, it ca be show that does ot cotai aythig but the fuctios of the form (). I this case it is obvious that is of a fiite dimesioality d. The ier product i betwee its two elemets w, R d ad v, R d (which are two liear fuctios) is defied by w, R d, v, R d = w, v R d.. Polyomial kerel of a secod degree Cosider X = R ad k(x, y) := ( x, y R + ). Expadig the brackets we see: k(x, y) = x y + x y + x x y y + x y + x y +. First we eed to check that it is ideed a kerel. It is symmetric. To check the positive defiiteess ote that if we defie a mappig ψ : X R 6 by we may write ψ(x) = (x, x, x x, x, x, ) k(x, y) = ψ(x), ψ(y) R 6, x, y X. I other words, we showed that k ca be expressed as a liear kerel after mappig X ito R 6 usig ψ. We already showed i the previous example that liear kerel is ideed positive defiite. Iterestigly otice that the image of ψ is oly a subset of R 6, i.e. there are poits z R 6 such that z ca ot be expressed as ψ(x) for ay x X. Let us show that cotais all the polyomials up to degree, i.e. fuctios of the form: f(x) = v x + v x + v 3 x x + v 4 x + v 5 x + v 6, x X, v R 6. ()

3 First, we kow that all the fuctios of the form k(x, ) belog to for sure, i.e. all the fuctios of the form f(x) = w x + w x + w w x x + w x + w x +, x, w X. (3) These are polyomials with moomials of order up to two. However, we see that coefficiets of moomials are iterdepedet, ad they are all defied by settig oly two coefficiets w ad w. This is quite differet from (), where we are free to choose ay coefficiets of moomials. However, recall that RKHS is a vector space, thus it cotais all the liear combiatios of its elemets. Now, do we get all the fuctios of the form () if we take all the liear combiatios of the fuctios of the form (3)? It turs out that if we take the liear spa of the vectors of the form {(w, w, w w, w, w, ): w, w R} R 6 we will get the whole R 6 (HW). This shows that ideed cotais all the polyomials up to degree. It ca be also show that o other fuctios are cotaied i. Two examples above showed that RKHS ca be of a fiite dimesio, which may or may ot be larger tha the dimesioality of X. At this poit it is importat to say that actually RKHS ca be eve ifiite dimesioal. This is the case, for istace, for the so-called Gaussia kerel k(x, y) = e (x y) /σ. Represeter theorem Why are RKHS ad kerels so importat for machie learig? I all the previous lectures we studied problems of biary classificatio ad also shortly metioed regressio problems. But what type of predictors did we actually see? It turs out that the mai focus was o liear predictors. These fuctios (classifiers) are a good start, but of course they are ot too flexible. We also saw a example of oliear methods, such as KNN. Note, however, that KNN ca t be cosidered as a learig algorithm which chooses a predictor ĥ from a fixed set of predictors H. Fially, we saw the AdaBoost algorithm, which outputs a complex compositio of base classifiers. This compositio is of course ot a liear classifier (eve if the base classifiers were liear). Kerels ad RKHS provide a very coveiet way to defie classes H cosistig of oliear fuctios. As we saw, it is eough to specify oe kerel fuctio k to implicitly get the whole RKHS. Now, assume we would like to choose our predictors from. How do we do that? Next result shows that ofte this problem ca be solved quite efficietly. Theorem (Represeter theorem). Assume k is a kerel defied over ay X ad is a correspodig RKHS. Take ay poits X,..., X X. Cosider the followig optimizatio problem: ( mi l i f(xi ) ) + Q( f Hk ), f (4) where l i : R R, i =,..., are ay fuctios ad Q: R + R is a odecreasig. The there exist α,..., α R such that f = α i k(x i, ) solves (4). 3

4 Proof. Assume there is f solvig (4). Because is a Hilbert space we may write f = β i k(x i, ) + u, where u, ad u, k(x i, ) Hk = 0 for all i =,...,. We used the fact that ay vector (fuctio) i a Hilbert space ca be uiquely expressed as a sum of its orthogoal projectio oto the liear subspace ad a complemet, which is orthogoal to that subspace. It is also easy to check that f = β i k(x i, ) + u H k ad thus where we deoted f Hk f X Hk, f X := β i k(x i, ). Because Q is odecreasig we coclude that Q( f Hk ) Q( f X Hk ). Now ote that because of the reproducig property ( l i f (X i ) ) ( = l i f ) ( ) ( ) (, k(x i, ) Hk = li f X + u, k(x i, ) Hk = l i f X, k(x i, ) Hk = l i fx (X i ) ). I other words we shoed that ( l i f (X i ) ) = ( l i fx (X i ) ). Thus, the value of the objective fuctioal (4) at f X is ot larger tha for f, which shows that f X also solves the optimizatio problem. I order to motivate represeter theorem we will first cosider two cocrete examples of Problem 4. Biary classificatio Ca we use the real-valued fuctios from for a biary classificatio with Y = {, +}? Of course! We just eed to take the sig of f, which gives us a biary-valued fuctio. Cosider a traiig sample S = {(X i, Y i )} with X i X for ay iput space X ad Y i Y. Take ay kerel k o X. Fially, set l i (z) := {Y i z 0}. I this case ( l i f(xi ) ) = {Y i f(x i ) 0} is just a empirical biary loss associated with a classifier sgf(x). Settig Q(z) = 0 we see that (4) correspods to the empirical risk miimizatio of a biary loss over. 4

5 Squared loss regressio We may also use elemets of for predictig real-valued outputs. Set Y = R ad l i (z) = (Y i z). I this case ( l i f(xi ) ) = ( Yi f(x i ) ) is just a empirical squared loss ad thus, settig Q(z) = 0 we get the empirical squared loss miimizatio over. What is the importace of Theorem? A surprisig message is the followig. Origially, (4) is a optimizatio with respect to elemets of, which are high-dimesioal objects ad potetially eve ifiite-dimesioal. I other words, solvig (4) requires choosig m real umbers if is m-dimesioal (with m potetially huge) or choosig a fuctio, which ca ot be described by ay fiite umber of parameters if is ifiite-dimesioal. Still, Theorem tells us that i ay case this problem may be reduced to choosig oly real-valued parameters. This gives a huge boost i efficiecy if dim( ), ad especially if is ifiite-dimesioal. Usig represeter theorem ad reproducig property we may restate the Problem 4 i the followig form: mi l i α j k(x i, X j ) + Q α,...,α R α j k(x j, ) j= j= Hk = mi l i α j k(x i, X j ) + Q α i α j k(x i, X j ). α,...,α R j= j= We see that this optimizatio problem depeds o X i ad k oly through the kerel matrix K X R with (i, j)-th elemet beig k(x i, X j ). 5