Corollary 15.7 (last part): with λ = √(2ρ²/(B²m)), E_S[L^{0−1}_D(A(S))] ≤ E_S[L^hinge_D(A(S))] ≤ min_{‖w‖≤B} L^hinge_D(w) + √(8ρ²B²/m)
ProvedUnderstandingML.soft_svm_norm_boundgeneralization-boundmarginsoft-svmsupport-vector-machines
Corollary 15.7 (last part). For every , if we set then
Formally: the Soft-SVM parameter depends on ; "" is "for every with ".
Preamble
import Definitions.Def_UnderstandingML_SVM open MeasureTheory open scoped InnerProductSpace
Formal statement
namespace UnderstandingML
/-- **Corollary 15.7** (p. 208), last part. For every `B > 0`, if Soft-SVM (15.6) is run with
`λ = √(2ρ²/(B²m))` then
`E_S[L^{0−1}_D(A(S))] ≤ E_S[L^hinge_D(A(S))] ≤ min_{w : ‖w‖ ≤ B} L^hinge_D(w) + √(8ρ²B²/m)`.
The parameter depends on `m`; "`min`" is "for every `w` with `‖w‖ ≤ B`". -/
theorem soft_svm_norm_bound {d : ℕ} (D : Measure (Vec d × ℝ)) [IsProbabilityMeasure D]
{ρ B : ℝ} (hρ : 0 < ρ) (hB : 0 < B) (hD : ∀ᵐ z ∂D, ‖z.1‖ ≤ ρ ∧ (z.2 = 1 ∨ z.2 = -1))
(A : Learner (Vec d × ℝ) (Vec d))
(hA : ∀ (m : ℕ) (S : Fin m → Vec d × ℝ),
IsRLM hingeLoss (Real.sqrt (2 * ρ ^ 2 / (B ^ 2 * m))) S (A m S))
(hAmeas : ∀ m, Measurable (A m)) (m : ℕ) (hm : 0 < m) :
∫ S, risk zeroOneLoss D (A m S) ∂(iidLaw D m) ≤
∫ S, risk hingeLoss D (A m S) ∂(iidLaw D m) ∧
∀ w : Vec d, ‖w‖ ≤ B →
∫ S, risk hingeLoss D (A m S) ∂(iidLaw D m) ≤
risk hingeLoss D w + Real.sqrt (8 * ρ ^ 2 * B ^ 2 / m) := by sorry
end UnderstandingML
Source
Shalev-Shwartz and Ben-David, Understanding Machine Learning: From Theory to Algorithms, Cambridge University Press 2014, doi:10.1017/CBO9781107298019, §15.2.1 p. 208, Corollary 15.7 (from Corollary 13.9)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.