Neurons:A single nerve cell has 2 states: 1. Excitement 2. Inhibition. By means of synaptic transfer media, the cell is excited when the medium (signal) exceeds a certain value. Frank Rosenblatt simulates synaptic/threshold/cell-body excitement with the invention of a machine called a perceptron, the mathematical expression of which is as follows: F (x) is the cell body, 1 is excited; for 0, the intensity b for the analog synapse is suppressed, and the threshold X for the activation is a series of synaptic pairs The effect of cell body is easy to see, it can be classified by two yuan with Perceptron. At the same time, can be seen through the adjustment of weights, learning two yuan classification.
learning algorithms for perceptual machines:
- First set the ownership value to 0 so that the input value is preset to have no effect on the result. Also, the offset is preset to 0.
- We use the parameter maxiter. This parameter is the only hyper-parameter (hyper-parameter) in the entire algorithm. This hyper-parameter means that when we can't find the exact answer, we have to optimize the weights and offsets a few times.
- In line 5th of the algorithm Perceptrontrain, we calculated the current excitation value A with the addition formula mentioned earlier. Since this simple perceptron algorithm can only produce two-dimensional results, we only check if the symbol for the excitation value A is the same as the actual value (line 6th). If the same number, this means that the algorithm for the current data input values found the appropriate weights and offsets, the algorithm ends.
- If the excitation value is different from the actual value, the perceptron algorithm is going to self-regulate the process (7th, 8 lines). The self-tuning process is divided into two steps:
The new excitation value A is affected by the new weights W and the new bias B This is a more intuitive form, the more general derivation process is as follows: The mathematical concept of inner product: The Learning strategy of perceptual classification is that the training data set is linear and can be divided, The goal of Perceptual machine learning is to obtain a completely correct separation between the positive and negative instance points of the training set.
separating the super plane
Let's take a look at the geometric interpretation of the perceptual machine:
linear equations:Define a loss of experience function and minimize the loss. The loss function can select the total number of false points, or the distance from the wrong point to the super plane. In order to ensure continuous, the distance from the wrong point to the super plane is chosen as the experience loss function. Note y = 1 or-1 the general form of the Perceptron learning algorithm: When n is 1 o'clock, degenerate to the above special learning form to correct the geometrical meaning of the error:
- Look at the top right of the y=+1, correct is positive, but calculated negative, that the angle of W and X is too large, to turn W to X, because at this time y=+1 so W+yx is w+x, equivalent to the middle of the picture of the vector;
- Look at the picture of Y=-1, correct is negative, but calculated is positive, the angle of W and X is too small, to the direction of the W away from the X, at this time w+yx=w-x, so the result is shifted away from the direction of X.
Perception Machine-1
