Table 2. FC-VAE Model Architecture.
For FC-VAEs based on Product of Experts, the number of layers in each expert (modality-specific encoder-decoder) differ across each FC-VAE variant. Taking FC-VAE (L1) as an example, for each expert, the encoder contains one fully connected layer in addition to one that outputs mean and log-variance parameters, and the decoder consists of two fully connected layers.
| FC-VAE Model Architecture | |||||||||
|---|---|---|---|---|---|---|---|---|---|
|
| |||||||||
| Layers* | Encoder Layers | Decoder Layers | |||||||
|
| |||||||||
| 1 | FC | FC(μ,σ) | FC | FC | |||||
| For each expert | 2 | FC | FC | FC(μ,σ) | FC | FC | FC | ||
| 3 | FC | FC | FC | FC(μ,σ) | FC | FC | FC | FC | |
*
Number of encoding layers (excluding FC(μ,σ))