Skip to content

CO3519 Artificial Intelligence
CO3519 Lecture 12 - Basics of Deep Learning

Lecture Documents

CO3519 Lecture 13.pdf

Learning Objectives

  • Introduce ConvNet
  • Learning how CNN works considering strides
  • How to construct a complete CNN as a end-to-end model

Contents

  • Convolutional Neural Network
  • Building blocks of CNN
  • Convolutional Operation
  • Constructing CNN

Brief about CNN

  • Convolutional neural networks are a special kind of multi-layer neural network.
  • CNN is a feed-forward network that an extract the topological properties from an image.
  • Like almost every other neural network, they are trained with a version of back-propagation algorithm.
  • Convolutional neural networks are designed to recognise visual patterns directly from pixel images with minimal pre-processing.
  • Can recognise patterns with extreme variability i.e. handwritten characters

CNN Building Blocks

CO3519 L13 - CNN Building Blocks.png

Convolutional Layer

CO3519 L13 - Convolution Layer 1.png
CO3519 L13 - Convolution Layer 2.png
CO3519 L13 - Convolution Layer 3.png
CO3519 L13 - Convolution Layer 4.png
CO3519 L13 - Convolution Layer 5.png
CO3519 L13 - Convolution Layer 6.png
CO3519 L13 - Convolution Layer 7.png
CO3519 L13 - Convolution Layer 8.png
CO3519 L13 - Image Process.png

Strides Example

CO3519 L13 - Strides 1.png
CO3519 L13 - Strides 2.png
CO3519 L13 - Strides 3.png
CO3519 L13 - Strides 4.png
CO3519 L13 - Strides 5.png
CO3519 L13 - Strides 5.5 (Mistake Ordering).png
CO3519 L13 - Strides 6.png
CO3519 L13 - Strides 7.png
CO3519 L13 - Strides 8.png
CO3519 L13 - Strides 9.png

Convolution Calculation Example

CO3519 L13 - Convolution Calc 1.png
CO3519 L13 - Convolution Calc 2.png
CO3519 L13 - Convolution Calc 3.png
CO3519 L13 - Convolution Calc 4.png
CO3519 L13 - Convolution Calc 5.png
CO3519 L13 - Convolution Calc 6.png
CO3519 L13 - Convolution Calc 7.png

Error

Skips 2 steps. Only moving in steps of 1. Slide error.

CO3519 L13 - Convolution Calc 8.png
CO3519 L13 - Convolution Calc 9.png
CO3519 L13 - Convolution Calc 10.png
CO3519 L13 - Convolution Calc 11.png
CO3519 L13 - Convolution Calc 12.png
CO3519 L13 - Convolution Calc 13.png
CO3519 L13 - Convolution Calc 14.png
CO3519 L13 - Convolution Calc 15.png
CO3519 L13 - Convolution Calc 16.png

Reasons for Feature Map Creation

CO3519 L13 - Feature Map Reasons.png

Pooling Layer

A subsampling layer commonly referred to as the pooling layer which reduces the spatial size of feature maps helping achieve translation invariance and reduces computational complexity.

Ensures that the same result will be obtained even when the image features have small translations.

Important

Important operation for object classification and detection.

Types of Pooling

Max Pooling - Takes the maximum value from a given window. Helps preserve the most important features.
Average Pooling - Computes the average value in the window, resulting in a smoother representation.
Global Pooling - Reduces an entire feature map to a single value, often used before fully connected layers.

Pooling Layer Example (Feature Map to Pooling Layer)

CO3519 L13 - Pool Layer 1.png
CO3519 L13 - Pool Layer 2.png
CO3519 L13 - Pool Layer 3.png
CO3519 L13 - Pool Layer 4.png
CO3519 L13 - Pool Layer 5.png

Constructing CNNs

  • Construct a CNN by placing a number of convolution and pooling layers in cascade:
  • Specify input size.
  • Specify filter size at each layer, etc.
  • Add a number of fully connected layers after convolution and pooling layers
  • Add a classifier at the end of the network.

CO3519 L13 - Constructing Full CNN.png

Next Week

  • Extension of CNN.

CO3519 Lecture 14 - ConvNets II