Table 2 Major deep learning model architecture families and characteristics. Note that there may be multiple variants (usually with different number of layers/parameters) within each architecture family
Architecture family | Original year | Parameters | Layers | Module organization | Example application(s) |
|---|---|---|---|---|---|
AlexNet | 2012 | ~ 60 million | 8 | Convolutional, Max Pooling | Abrà moff et al. [34], Quellec et al. [66] |
VGGNet | 2014 | ~ 180 million | 19 | Convolutional, Max Pooling | Abrà moff et al. [34], Quellec et al. [66], Ting et al. [26], Gargeya et al. [47], Bellemo et al. [52] |
GoogLeNet (also Inception v1) | 2015 | ~ 7 million | 22 | Inception, Pool+Concat | Takahashi et al. [63] |
Inception (v3) | 2015 | ~ 24 million | 42 | Inception, Pool+Concat | |
ResNet | 2016 | ~ 60 million | 152 | Convolutional, Skip Connections | Bellemo et al. [52] |
Inception-ResNet (v2) | 2016 | ~ 56 million | 164 | Residual Inception |  – |
SqueezeNet | 2016 | ~ 1.2 million(before pruning) | 14 | 1 × 1 Convolutional, Squeeze & Expand Layers |  – |
ResNeXt | 2017 | ~ 25 million | 50 | Convolutional (Grouped) |  – |
DenseNet | 2017 | ~ 20 million | 201 | Dense, Transition |  – |