Please enable JavaScript.

Coggle requires JavaScript to display documents.

Deep Learning (Introduction (A machine learning subfield of learning…

- - - - Role
        
        Feature extraction
        
        Reduce the dimension/spatial size of representation data
        
        Reduce the amount of parameters and computation in the network
        
        Hence, control overfitting
      - How?
        
        Takes the largest value from one patch/region of an image
        & place it in new matrix
      - Finding
        A region that with high correlation to
        each feature are preserved
    - - Role
        
        Perform classification based on the
        features extracted by the previous layers
      - How?
        
        Using softmax activation function
        
        Act like normal neural network
    - - Convolution
        Mathematical combination of two functions
        to produce a third function
      - How?
        
        Use kernel/filter (include bias)
        to product feature map
        
        After that ReLU activation function is applied
      - Role
        Feature extraction
- - - - Distributed hidden state allows to
        store information about past
      - Allow to update hidden state
    - - Hidden state at time step t
        \(h_t = f(Ux_t + Wh_{t-1}) \)
        where,
        U & W = weight
        \(x_t \) = input at time t
        \(h_{t-1} \) = previous hidden state
        f can be ReLU/softmax
      - Output at time t
        \(o_t = softmax(V s_t) \)
        where,
        V = weight
        \(s_t \) = hidden state at time t
      - Input at time t
        \(x_t \)