Hands-On Machine Learning for Algorithmic Trading

VGGNet – going for smaller filters

The runner-up in ILSVRC 2014 was the network from Karen Simonyan and Andrew Zisserman from Oxford University. It demonstrated the effectiveness of much smaller 3 x 3 convolutional filters combined in sequence, and reinforced the importance of depth for strong performance. It contains 16 convolutional and fully-connected layers that only perform 3 x 3 convolutions and 2 x 2 pooling:

VGG16 has 140 million parameters that increase the computational costs of training and inference, as well as memory requirements. However, most parameters are in the fully connected layers that were since discovered to be not essential, ...