3-Convolutional Neural Networks
This is Lecture 3 of the series, where we begin moving from general deep learning frameworks to specific model families. It focuses on visual tasks, emphasizing why image problems typically do not use plain fully connected structures, and how CNNs leverage local connectivity, parameter sharing, and hierarchical feature extraction to process image data.
What This Lecture Covers
This lecture covers both fundamental architectures and classic models with hands-on practice, making it suitable for building your first-round understanding of "from principles to network design."
- Overview: starting from the limitations of deep neural networks on image problems, it introduces the basic ideas behind convolutional neural networks.
- Building a CNN from Scratch: introduces the basic structure of CNNs, training methods, and how to implement a working convolutional network using Paddle.
- Exploring Classic CNN Architectures: connects representative architectures such as AlexNet, VGG, GoogLeNet/Inception, and ResNet to help you understand the direction of model evolution.
- Course Practice: uses cat-vs-dog classification to bring the image classification task into real data and model implementation.
How to Study This
- When studying this lecture, repeatedly pay attention to how tensor shapes change. The changes in convolution layers, pooling layers, channel counts, and feature map sizes are key to understanding CNNs.
- When studying classic networks, do not just memorize names -- focus on comparing what problems each one solved, such as depth, width, computational efficiency, or gradient propagation.
- If time is limited, at least implement a minimal CNN yourself and run a classification task once. This will make it much easier to follow detection, segmentation, and other topics later.