week7#

This notes is completed with assistance of ChatGPT

View Lecture 13. Convolutional Neural Networks (CNN)

View Lecture 14. Bayesian regression

View workshop8-slides

View workshop8

Convolutional Neural Networks (CNN) - Summary

  1. Overview:

    • CNNs are a type of deep learning model primarily used for image recognition and processing.

    • Inspired by the human visual system.

  2. Architecture:

    • Input Layer: Takes raw pixel values of the image.

    • Convolutional Layer: Filters slide over the image to produce feature maps. Highlights important features.

    • ReLU Layer: Introduces non-linearity using a function like Rectified Linear Unit (ReLU). Makes the network powerful.

    • Pooling/Subsampling Layer: Reduces spatial size, reduces computation, and helps in achieving translational invariance.

    • Fully Connected (FC) Layer: Connects every neuron from the previous layer, making decisions based on learned features.

  3. Key Concepts:

    • Filter/Kernels: Small, learnable weight matrices which slide over the input data (like a window) to produce a feature map.

    • Stride: Number of pixels the filter moves while sliding over the image. Affects the size of the output feature map.

    • Padding: Adding layers of zeros around the image border to ensure the spatial dimensions remain the same after convolution.

    • Pooling: Down-sampling technique. Max pooling (takes maximum value) and average pooling (takes average) are popular methods.

  4. Advantages:

    • Parameter Sharing: Reduces the number of parameters, thus reducing computation and preventing overfitting.

    • Translational Invariance: Recognizes patterns irrespective of their position in the image.

    • Hierarchical Learning: Lower layers learn basic patterns; higher layers learn complex features.

  5. Applications:

    • Image and Video Recognition

    • Image Classification

    • Medical Image Analysis

    • Self-driving Cars

    • Many others in vision-based tasks.

  6. Challenges:

    • Requires a large amount of labeled data for training.

    • Computationally intensive, especially for larger images or deeper networks.

  7. Optimization and Training:

    • Often trained using backpropagation and optimizers like SGD, Adam, or RMSprop.

    • Regularization techniques (e.g., dropout) are used to prevent overfitting.