Figure 4.

Primary configurations for incorporating the MV-DTF layer in neural network architectures. (a) MV-DTF layer for multi-view feature extraction on single-task learning, where $g:{\mathbb{R}}^{I_1}\times \cdots \times {\mathbb{R}}^{I_M}\to {\mathbb{R}}^{J_1\times \cdots \times J_P}$, and $h:{\mathbb{R}}^{J_1\times \cdots \times J_P}\to {\mathbb{R}}^{O_t}$. (b) MV-DTF layer for multilinear regression on single-task learning, where ${\zeta }_m:{\mathbb{R}}^{I_m}\to {\mathbb{R}}^{H_m}$, and $g:{\mathbb{R}}^{H_1}\times \cdots \times {\mathbb{R}}^{H_M}\to {\mathbb{R}}^{O_t}$. (c) MV-DTF for multi-view feature extraction on multitask learning, where $g:{\mathbb{R}}^{I_1}\times \cdots \times {\mathbb{R}}^{I_M}\to {\mathbb{R}}^{J_1\times \cdots \times J_P}$, and $h_t:{\mathbb{R}}^{J_1\times \cdots \times J_P}\to {\mathbb{R}}^{O_t}$.