Understanding Machine Learning: From Theory to Algorithms

(Jeff_L) #1
29.6 Exercises 409

29.6 Exercises



  1. Letd,k >0. Show that there exists a binary hypothesisHbinof VC dimension
    dsuch that Ndim(HbinOvA,k) =d.

  2. Prove Lemma 29.6.

  3. Prove Natarajan’s lemma.
    Hint:Fix somex 0 ∈ X. Fori,j∈[k], denote byHijall the functionsf:
    X {x 0 } →[k] that can be extended to a function inHboth by defining
    f(x 0 ) =iand by definingf(x 0 ) =j. Show that|H|≤|HX{x 0 }|+



i 6 =j|Hij|
and use induction.


  1. Adapt the proof of the binary fundamental theorem and Natarajan’s lemma
    to prove that, for some universal constantC >0 and for every hypothesis
    class of Natarajan dimensiond, the agnostic sample complexity ofHis


mH(,δ)≤C

dlog

(kd


)

+ log(1/δ)
^2

.


  1. Prove that, for some universal constantC >0 and for every hypothesis class
    of Natarajan dimensiond, the agnostic sample complexity ofHis


mH(,δ)≥C

d+ log(1/δ)
^2.
Hint:Deduce it from the binary fundamental theorem.


  1. LetHbe the binary hypothesis class of (nonhomogenous) halfspaces inRd.
    The goal of this exercise is to prove that Ndim(HOvA,k)≥(d−1)·(k−1).

    1. LetHdiscretebe the class of all functionsf: [k−1]×[d−1]→ { 0 , 1 }for
      which there exists somei 0 such that, for everyj∈[d−1]
      ∀i < i 0 ,f(i,j) = 1 while∀i > i 0 ,f(i,j) = 0.
      Show that Ndim(HOvAdiscrete,k) = (d−1)·(k−1).

    2. Show thatHdiscretecan be realized byH. That is, show that there exists
      a mappingψ: [k−1]×[d−1]→Rdsuch that




Hdiscrete⊂{h◦ψ : h∈H}.
Hint:You can takeψ(i,j) to be the vector whosejth coordinate is 1, whose
last coordinate isiand the rest are zeros.


  1. Conclude that Ndim(HOvA,k)≥(d−1)·(k−1).

Free download pdf