VGG-16

VGG-16, a modification of AlexNet, is a Convolutional Neural Network known for its simplicity and depth. Developed by Simonyan and Zisserman in 2015, it represented a significant step in streamlining network architecture.

Key Features

• Simplified Architecture: The most remarkable thing is the simplicity. The researchers decided to uniformly just use 3x3 convolutional layers with stride 1 and same padding, coupled with 2x2 max pooling layers with a stride of 2.
• Depth: Comprises 16 layers with weights.
• Parameter Count: Approximately 138 million, mostly in the fully connected layers.
• Filter Increment: The number of filters doubles after each max pooling layer, from 64 to 128, then to 256, and finally to 512.

Architecture Design

The design of VGG-16 is notable for its uniformity and simplicity:

• Convolutional Layers: All using 3x3 filters with a stride of 1 and SAME padding.
• Pooling Layers: Consistently 2x2 with a stride of 2, responsible for reducing spatial dimensions.
• Filter Increase: Sequential increase in the number of filters, peaking at 512.

Memory Requirements

• Forward Propagation: Requires around 96MB of memory per image.
• Memory Distribution: Most of the memory usage is concentrated in the earlier layers.

VGG-16 vs. VGG-19

• VGG-19: An even deeper variant with 19 layers.
• Preference: VGG-16 is generally preferred over VGG-19 due to similar performance with reduced complexity.

This network is large even by modern standards. It has around 138 million parameters (however, most of the parameters are in the hidden layers).

It has a total memory of 96MB per image for only forward propagation! But, most memory is in the earlier layers.

Impact and Attraction

• Uniformity: The regular structure of VGG-16, with its consistent use of convolutional and pooling layers, contributed to its popularity.
• Dimensional Changes: The methodical reduction in spatial dimensions (height and width) while increasing the depth (number of channels) is a hallmark of its design.
• Influence on CNN Rules: The VGG-16 paper contributed to establishing guidelines for CNN architecture in image recognition tasks.

VGG-16 stands out for its simplicity and uniform architecture, making it a valuable model for understanding deep CNNs. Its approach to layer design and parameter allocation has been influential in the field of computer vision. However, it has a lot of parameters. It is attractive in the regular way $n_H$ and $n_W$ go down and $n_c$ goes up. The paper also tries to make some rules regarding the use of CNNs.

Simonyan & Zisserman 2015. "Very deep convolutional networks for large-scale image recognition"