Please enable JavaScript.

Coggle requires JavaScript to display documents.

Machine learning (models (sigmoid (h(x) = 1 / 1 + e^-θTX) (decision…

- - - - linear θTX = θ0 + θ1x1 + θ2x2 + .. + θnxn
      - non-linear θTX = θ0 + θ1x1^2... + θnxn^b
    - - One-vs-all
  - - - input layer
        represent input signal
      - hidden layers
      - output layers
        
        usually contains neuron
        for each class
        
        for example 10 neurons
        for digits recognition
    - - analog of biology neuron
      - has many inputs
        
        each input has its own weight
        
        can has bais input x0 = 1
      - and one output
        
        output calculates as
        activation function of
        weighted sum of inputs
      - activation function
        
        usually sigmoid
        
        ReLU function
        
        linear
- - - - add to cost function lambda * sum[1..n] (θ^2)
      - regularization parameter (lambda)
      - θ0 - is not used in regularization
- - - - Regression - predict continuous values
      - Classification - predict discrete value
    - - cost function J(θ) = 1/2m * sum[1..m] (h(xi) - yi)^2
      - process of minimize cost function
      - θj = θj - alpha delta(cost function)
        [**delta - derivative of cost function]
      - lernin rate (alpha)
      - batch - each step uses all the training examples
      - feature scaling
        we need all features in
        one range for convergence
        
        mean normalization
        xi = (xi - mean(x)) / range(x)
      - Backpropagaton :warning:
        
        algorithm for correction
        weights in neural networks
        
        perform forward propogation
        
        error of output layer is ai - yi
        
        proportionality
        split error on inputs
        
        if error for output 1 = 0.5
        and input 3 inputs with weights [1, 2, 3]
        w0 = 1 + 1/6 0.5
        w1 = 2 + 2/6 0.5
        w3 = 3 + 3.6 0.5
        
        links
        
        https://www.coursera.org/learn/machine-learning/home/week/5
      - Gradient checking :warning:
        
        manually calculate gradient of cost function
        delta J(θ) = J(θ + epsilon) - J(θ - epsilon) / 2 epsilon
        
        compare manually calculated values with gradient errors
    - - start with a simple algorithm
      - plot learning curves :warning:
      - Manually examine the error examples
  - - - data analize
      - data segmentation
      - anomaly detection