多模态数据缺失模态处理
Mitigating Inconsistencies in Multimodal Sentiment Analysis under Uncertain Missing Modalities
处理多模态缺失数据
Input
视觉模态
709-dimensional (通过OpenFace2.0得到)
文本模态
768-dimensional(通过一个12layer 768-hidden 12-head的 bert 得到)
声音模态
33-dimensional
Input sequence
这些特征被扔到多头注意力机制里去（自注意力机制）
E
m
=
M
H
A
(
K
m
,
K
m
,
K
m
)
,
K
m
∈
{
X
v
,
X
a
,
X
t
}
.
(1)
\begin{aligned}&E_m=\mathbf{MHA}(K_m,K_m,K_m),\\ &K_m\in\{X_v,X_a,X_t\}.\end{aligned}\tag{1}
Em=MHA(Km,Km,Km),Km∈{Xv,Xa,Xt}.(1)
得到 Input sequence
X
=
[
E
v
∣
∣
E
a
∣
∣
E
t
]
\mathcal{X}=[E_v||E_a||E_t]
X=[Ev∣∣Ea∣∣Et]
预训练网络 Pre-trained Network
E p r e = [ E v ∣ ∣ E a ∣ ∣ E t ] , P p r e = s o f t m a x ( F C ( E p r e ) ) . (3) \begin{aligned}\\ E_{pre}&=[E_v||E_a||E_t],\\ P_{pre}&=softmax(\mathbf{FC}(E_{pre})).\end{aligned} \tag{3} EprePpre=[Ev∣∣Ea∣∣Et],=softmax(FC(Epre)).(3)
用完整的三模态数据训练 训练完成后就冻结
编码器-解码器 网络 Encoder-Decoder Network（用于预测缺失模态）
X ⟶ ϕ F , F ⟶ ψ X ′ , \begin{aligned}\mathcal{X}\overset{\phi}{\longrightarrow}\mathcal{F},\\ \mathcal{F}\overset{\psi}{\longrightarrow}\mathcal{X}',\end{aligned} X⟶ϕF,F⟶ψX′,
其中 X \mathcal{X} X表示输入 ϕ \phi ϕ表示encoder ψ \psi ψ表示decoder
三种用于编码-解码的网络架构
架构1：AE AutoEncoder
数学表示如下
其隐藏层大小如下
[300, 256, 128, 64, 128, 256, 300]
h
i
=
{
X
,
i
=
0
R
e
L
U
(
F
C
(
h
i
−
1
)
)
,
0
<
i
≤
7
,
(5)
\begin{aligned}h_i=\left\{\begin{array}{ll}\mathcal{X},&i=0\\ ReLU(\mathbf{FC}(h_{i-1})),&0<i\le7\end{array}\right.\quad,\end{aligned} \tag{5}
hi={X,ReLU(FC(hi−1)),i=00<i≤7,(5)
其中encoder输出为
h
4
h_4
h4 decoder输出为
h
7
h_7
h7
架构2：MMIN Missing Modality Imagination Network
D M M I N = X + ∑ i = 1 5 X i ′ , E M M I N = F C ( [ F 1 ∣ ∣ F 2 ∣ ∣ . . . ∣ ∣ F 5 ] ) , (6) \begin{aligned}D^{MMIN}&=\mathcal{X}+\sum_{i=1}^5\mathcal{X}_i',\\ E^{MMIN}&=\mathbf{FC}([\mathcal{F}_1||\mathcal{F}_2||...||\mathcal{F}_5]),\end{aligned}\tag{6} DMMINEMMIN=X+i=1∑5Xi′,=FC([F1∣∣F2∣∣...∣∣F5]),(6)
架构3：TF Transformer-based encoder-decoder model
E T F = F N ( M H A ( X , X , X ) ) , D T F = F F N ( M H A ( F , F , F ) ) , F F N ( x ) = R e L U ( W 1 x + b 1 ) W 2 + b 2 ) , \begin{aligned}&E^{TF}=\textbf{F}\mathbf{N}(\textbf{M}\mathbf{H}\mathbf{A}(\mathcal{X},\mathcal{X},\mathcal{X})),\\ &D^{TF}=\mathbf{F}\mathbf{F}\mathbf{N}(\textbf{M}\mathbf{H}\mathbf{A}(\mathcal{F},\mathcal{F},\mathcal{F})),\\ &\textbf{F}\mathbf{F}\mathbf{N}(x)=ReLU(W_1x+b_1)W_2+b_2),\end{aligned} ETF=FN(MHA(X,X,X)),DTF=FFN(MHA(F,F,F)),FFN(x)=ReLU(W1x+b1)W2+b2),
(这里的 F \mathcal{F} F是怎么来的？)
1821
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
