readme.md

Neural Networks (Delta Learning Algorithm)

1. Learning

Weights for a linear threshold unit can be learnt using the same methods used to set the parameters of a linear discriminant function:

• Perceptron Learning
• Minimum Squared Error Learning (Widrow-Hoff)
• Delta Learning Rule

1.1 Delta Learning Rule

The Delta Learning algorithm is a supervised learning approach. The weights are adjusted in proportion to the difference between:

• The desired output, t
• The actual output, y

Delta Learning can do either sequential or batch update:

• Sequential update: w←w+η( t−y ) x^t
• Batch update: w←w+η ∑_p( t p−y p ) x p

2. Delta Learning Rule pseudocode implemented using Gradient Descent:

Implementing using Gradient Descent is equivalent to the Sequential Delta Learning Algorithm.

• Set value of hyper-parameter (η)
• Initialise w to arbitrary solution
• For each sample, (x_k, t_k) in the dataset in turn:
  • Update weights: w←w+η ( t_k−H ( wx_k ) ) x^t_k

3. Heaviside Function

The Heaviside Function often written as H(x), is a non continuous function whose value is zero for a negative input and one for a posotive input. In the implementation we are using the following definition of the Heaviside Function.

3. Implementation Explanation

• heaviside_function_run(w, x)

  Is the function for the Heaviside Function, the fucntions inputs are:

  • w : List of weights
  • x : List of feature vectors

• sequential_delta_learning_run(X, n, w, t, epoch)

  Is the function for the Sequential Delta Learning Algorithm, the functions inputs are:

  • X : List of lists of feature vectors
  • n : Learning rate
  • w : List of weights
  • t : List of classes
  • epoch : Integer value