Suppose Q1 linear
(1) whether or logistic regression linear regression feature amount when too much, the calculated load will be very large. The 50x50 pixel feature 2500, if there will be combinations of two 2500 2 /2 (wherein nearly 3,000,000). Ordinary linear regression and logistic regression model can not effectively deal with so many features, this time need to use a neural network.
Q2 neurons and brain
After a piece of the brain can learn, learn other functions, such as a piece of tactile feel, but after receiving the visual training, can feel the visual.
Q3 model represents 1
(1) have axons and dendrites of neurons, dendrites many receiving electrical signals, a axons transmit electrical signals.
(2) According to neuron model, to create a logistic regression model:
(3) Multi neurons, multilayer, called input layer, hidden layer and output layer:
Q4 model represents 2
(1) represents a vector is more efficient than coding cycle:
These are just a training example, if the entire training set is calculated, it needs to be transposed X, such that in a same instance.
(2) more powerful than the neural network regression and the linear regression logic, characterized in that the former will continue to be high-class.
Wherein Q5 and intuitive understanding 1
(1) In essence, the neural network can obtain its own set of features by learning.
(2) a logical AND operation
Wherein θ is the [-30,20,20] T .
(3) the logical OR
Wherein θ is the [-10,20,20] T .
(4) Non-operation
Sample II and intuitive understanding Q6
Constructed on a complicated operation by a simple operation (when the same is 0 with 1 or take 1)
In this way we can begin to construct more complex functions, can be more powerful eigenvalues. This is the neural network powerful place.
Q7 multi-class classification
To identify pedestrians, cars, motorcycles, trucks four categories in a picture, this is the neural network can set up four outputs, each with a 1 or 0 for whether there is some sort of can.