关于“Deep Adversarial Metric Learning”

论文阅读笔记

  1. 论文题目:Deep Adversarial Metric Learning
  2. 论文创新点:DAML利用大量 易辨识的负例( easy negatives)生成 难辨识的负例(hard negatives);现存的度量学习(metric learning)方法仅利用数量少的hard negatives而忽略数量多的easy negatives。如下图所示,关于“Deep Adversarial Metric Learning”_第1张图片
  3. 具体内容:
      DAML的网络结构如下:关于“Deep Adversarial Metric Learning”_第2张图片
       DAML的目的是通过优化设计好的目标函数 θ f a = arg ⁡ min ⁡ θ f J m ( θ f ; x i , x i + , x ~ i − , f ) θ ^a _f = \mathop {\arg\min} \limits_{θ_f} \mathrm{J} _m(θ_f ; x_i , x ^+ _i , \tilde{x} ^− _i , f) θfa=θfargminJm(θf;xi,xi+,x~i,f)来获得参数 θ f θ_f θf,这里 x ~ i − \tilde{x}^− _i x~i表示产生的负例(negative sample), x ~ i − = G ( θ g ; x i − , x i , x i + ) \tilde{x} ^− _i = G(θ_g; x ^− _i , x_i , x ^+ _i ) x~i=G(θg;xi,xi,xi+)

       生成器(generator)的目标函数为:
    min ⁡ θ g J g e n = J h a r d + λ 1 J r e g + λ 2 J a d v = ∑ i = 1 N ( ∣ ∣ x ~ i − − x i ∣ ∣ 2 2 + λ 1 ∣ ∣ x ~ i − − x i − ∣ ∣ 2 2 + λ 2 [ D ( x ~ i − , x i ) 2 − D ( x i + , x i ) 2 − α ] + ) \mathop {\min} \limits_{θ_g} \mathrm{J}_{gen} = \mathrm{J}_{hard} + λ_1\mathrm{J}_{reg} + λ_2\mathrm{J}_{adv} = \sum \limits ^{N} _{i=1} (||\tilde{x} ^− _i − x_i ||^2 _2 + λ_1||\tilde{x} ^− _i − x ^- _i ||^2 _2 + λ_2[D(\tilde{x} ^− _i, x_i) ^2 − D(x ^+ _i , x_i) ^2 − α]_+) θgminJgen=Jhard+λ1Jreg+λ2Jadv=i=1N(x~ixi22+λ1x~ixi22+λ2[D(x~i,xi)2D(xi+,xi)2α]+)

      对抗性度量学习(Deep Adversarial Metric Learning)的框架可以应用于有监督度量学习的各种目标函数,即用以下目标函数同时训练 难辨识的负例生成器(hard negative generator)和距离度量:
    min ⁡ θ g , θ f J = J g e n + λ J m \mathop {\min} \limits_{θg,θf} \mathrm{J} = \mathrm{J}_{gen} + λ\mathrm{J}_m θg,θfminJ=Jgen+λJm
       DAML (cont):对于contrastive embeddings, J m = ∑ i = 1 N i D ( x i + , x i ) 2 + ∑ j = 1 N j [ α − D ( x ~ j − , x j ) 2 ] + \mathrm{J}_m = \sum \limits ^{N_i} _{i=1} D(x ^+ _i , x_i) ^2 + \sum \limits ^{N_j} _{j=1} [α − D(\tilde{x} ^− _j , x_j ) ^2 ] _+ Jm=i=1NiD(xi+,xi)2+j=1Nj[αD(x~j,xj)2]+

       DAML (tri):对于triplet embeddings,
    J m = ∑ i = 1 N [ D ( x i + , x i ) 2 − D ( x ~ i − , x i ) 2 + α ] + \mathrm{J}_m = \sum \limits ^{N} _{i=1} [D(x ^+ _i , x_i) ^2 − D(\tilde{x} ^− _i , x_i ) ^2 + α] _+ Jm=i=1N[D(xi+,xi)2D(x~i,xi)2+α]+
       DAML (lifted):对于lifted structure,
    J m = 1 2 N i ∑ i = 1 N i max ⁡ ( 0 , J i + , i ) \mathrm{J}_m = \dfrac{1}{2N_i} \sum \limits ^{N_i} _{i=1} \mathop{\max}(0, \mathrm{J}_{i^+, i}) Jm=2Ni1i=1Nimax(0,Ji+,i)
    J i + , i = max ⁡ ( max ⁡ α − D ~ ( x i + ) ,   max ⁡ α − D ~ ( x i ) ) + D ( x i + , x i ) \mathrm{J}_{i^+, i}= \mathop{\max}( \mathop{\max} α − \tilde{D}(x ^+ _i ),\ \mathop{\max} α − \tilde{D}(x_i)) + D(x ^+ _i , x_i) Ji+,i=max(maxαD~(xi+), maxαD~(xi))+D(xi+,xi)
            这里 D ~ ( X ) \tilde{D}(X) D~(X)表示负例对(negative pairs)到X的距离。

       DAML (N-pair):对于 N-pair loss,
    J m = 1 C ∑ c = 1 C log ⁡ ( 1 + ∑ c ′ ≠ c exp ⁡ ( D ( x c , x ~ c ′ + ) − D ( x c , x c + ) ) ) \mathrm{J}_m = \dfrac{1}{C} \sum \limits ^{C} _{c=1} \mathop{\log}(1 + \sum \limits _{c' \ne c} \mathop{\exp}(D(x_c, \tilde{x} ^+ _{c'}) − D(x_c, x ^+ _c ))) Jm=C1c=1Clog(1+c̸=cexp(D(xc,x~c+)D(xc,xc+)))这里 D ( x i , x j ) = f i T f j D(x_i , x_j ) = f ^T_i f_j D(xi,xj)=fiTfj是在N-pair loss度量相似性。

你可能感兴趣的:(关于“Deep Adversarial Metric Learning”)