N N Models - Artificial Intelligence

Artificial Neural Network (ANN)

• Used with structured data like tabular form
• Types
  • Single
  • Multilayer Feed Forward
  • Recurrent
• Encoding Techniques => Used to represent categorical data in a numerical format that can be used as input for training
  • One-Hot Encoding
  • Label Encoding
  • Binary Encoding
  • Manual Encoding
• Steps
  • Import libraries and Datasets
  • Data Exploration and Analysis
    • Remove Null Values
  • Data Preprocessing
    • Data Encoding
  • Data Splitting
  • Build Model => Regression, Classification
    • Initialize => Sequential
    • Add layers => Dense
      • Optimizer, Loss function
    • Compile & Train model
  • Test & Compare model
  • Predict
• Feed Forward Neural Network (FNN)
  • Input Layer > Hidden Layer > Output Layer
    • Accuracy increases with increase in Hidden layer
    • Heaviside step activation function is generally used in Output Layer
  • Considers only current input state, Memory is not used
    • No loop is formed, Data only moves forward
    • Can not deal with sequential data
  • Back Propagation Algorithm (Backward Propagation of Error)
    • Goal is to reach target value
      • Reduce the error factor
    • 
  • Feedback Network
    • Allows loop & feedback
    • Recurrent/Recursive network
    • Very complex to implement
• Sentiment analysis
  • Text Preprocessing
    • Remove the special characters
    • Convert entire text to upper or lower case
    • Split the data
    • Remove stopwords
    • Stemming / Lematization
    • Join the words
  • Vectorization
    • Convert text to number
  • Build Model

Convolution Neural Network (CNN)

• 
• 
• Accepts input in form of image
• Deep Learning algorithm that can take in an input image, assign importance (learnable weights and biases) to various aspects/objects in the image, and be able to differentiate one from the other
• Layers
  • Convolution => Learn to extract local patterns and features from the input data
    • Kernal/Filters > Feature Map
  • Pooling => Partition the input into non-overlapping regions and keep only the maximum value from each region
  • Flattening => Convert multi-dimensional input data into a one-dimensional vector
  • Fully Connected Layer (Dense layer) => Applies a linear transformation to the input followed by a non-linear activation function
• Steps
  • Import libraries and Datasets
    • Data Augmentation => ImageDataGenerator
  • Data Preprocessing
    • Normalize
  • Build Model => Regression, Classification
    • Initialize => Sequential
    • Add layers => Dense, Convolutional
    • Compile & Train
  • Test & Compare model & Predict

Recurrent Neural Network (RNN)

• 
• Memory is used that stores past input states, Considers previous state
• Output of Hidden layer serves as the input for next timestep (X_i)
  • Initial hidden state (h_o) is considered equal to "0" given as input along with Current input state (X₁) to generate current hidden state (h₁) which used to generate output (Y_i)
  • h₁ is given as input to next step as previous hidden step
• h_t = ActivationFunction(WeightPreviousHiddenStep * h_t-1 + WeightCurrentHiddenStep * X_t)
• Y_t = WeightOutputStep * h_t
• LSDM
  • Steps
    • Import libraries and Datasets
    • Data Exploration and Analysis
      • Remove Null Values
    • Data Preprocessing
      • Normalization
    • Data Splitting
    • Build Model => Regression, Classification
      • Initialize => Sequential
      • Add layers => Dense
        • Optimizer, Loss function
      • Compile & Train model
    • Test & Compare model
    • Predict