Our experts from the Big Data lab tell us.GoogLeNet is a deep convolutional neural network developed by a team of researchers from Google in 2014 for image classification. The "Le" in its name is a reference to the LeNet 5 neural network, one of the first convolutional neural networks that contributed to the development of the idea of deep learning since 1988.

The main difference between the GoogLeNet architecture and other convolutional neural networks (CNN) is the use of an additional module for initial data processing — Inception. It applies convolutions with different kernel sizes (1x1, 3x3, 5x5) in parallel, and then combines feature vectors. This makes it possible to more effectively identify local and global features of the analyzed image.

Despite the deep network architecture consisting of 22 layers, the number of GoogLeNet parameters used remains relatively small. This is achieved by using a 1x1 convolution, which essentially works as a linear filter and reduces the dimension of the next layer. Therefore, GoogLeNet is less demanding on the amount of video card memory than AlexNet and other architectures without the Inception module.

Due to its balance, GoogLeNet shows high classification accuracy on images of various sizes. In 2014, she won the ImageNet competition. Since then, more modern neural networks have been developed on its basis, also using a deep convolutional architecture and the concept of the Inception module.