AlexNet Architecture: Convolutional Layers, GPU Training, and Dropout

Summary

This note summarizes the AlexNet deep convolutional neural network described in the ImageNet classification paper. It outlines the model’s use of multiple convolutional channels, five convolutional stages and three fully connected layers, and explains how GPU computation supported training at scale. It also highlights dropout in the fully connected layers as a measure intended to reduce overfitting. An example traces the first convolution from a color image through an 11-by-11 kernel with stride four, producing feature maps split across GPUs.

The source reports training on ImageNet with 1.2 million images across 1,000 categories and describes the network as having about 60 million parameters and 650,000 neurons. These details offer a compact architectural overview, rather than a full account of the paper’s experiments, evaluation metrics, or later developments. The note’s input-size description and some layer-shape details are simplified, so it is best treated as an introductory summary rather than a precise implementation specification.

Key ideas

• AlexNet applies multi-channel convolutions to large-scale color-image classification.
• The summarized architecture contains five convolutional stages and three fully connected layers.
• GPU computation is presented as a way to accelerate training.
• Dropout is used in fully connected layers to help reduce overfitting.
• The note gives only a partial architectural overview and does not detail experimental evaluation.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.