Hands-On Machine Learning for Algorithmic Trading

GoogLeNet – fewer parameters through Inception

Christian Szegedy and his team at Google began working on more efficient CNN implementations that reduce the computational costs to facilitate practical applications at scale. The resulting GoogLeNet won the ILSVRC 2014 with only 4 million parameters, due to the Inception module, compared to AlexNet's 60 million and VGG's 140 million.

The Inception module builds on the network-in-network concept that uses 1 x 1 convolutions to compress a deep stack of convolutional filters, and reduce the cost of computation. The module uses parallel 1 x 1, 3 x 3, and 5 x 5 filters, but combines the latter two with 1 x 1 convolutions to reduce the dimensionality of the filters passed in by the previous layer. ...