Convolutional Neural Networks

by ComputingNotes
Machine Learning

References:

  1. https://www.ibm.com/topics/convolutional-neural-networks
  2. https://d2l.ai/chapter_natural-language-processing-applications/sentiment-analysis-cnn.html
  3. https://developer.ibm.com/articles/introduction-to-convolutional-neural-networks/

What are CNNs?

  • Three main layers: Convolutional, Pooling and Fully-Connected Layer
  • Convolutional Layer:
    • Core building block
    • Requires input data, filter and feature map.
  • Pooling layer – downsampling, dimensionality reduction
  • FC Layer – peforms task
  • Types – AlexNet, VGGNet, GoogleNet, ResNet, LeNet-5 (classic CNN architecture)

Different steps:

  1. Image channels
    • Way to represent image in numerical format.
    • Pixel represents image. Each pixel mapped to number between 0 to 255 (representing color ranging from 0 to 255)
    • Image represented in 3D array each for Red, Green and Blue values
  2. Convolution
    • Extract key features from image
    • Filter or kernel used
    • Filter strided over input array to produce output array called feature map.
  3. Pooling
    • Adding extra layers of zeros to outer rows and columns in the input array.
  4. Flattening
    • make outputs compatible with an artificial neural network
    • multidimensional array to A matrix with one column and n rows (nx1) also called a column vector.
  5. Full Connection
    • typically towards the end of neural network architecture
    • also known as dense layers
error: