Looking from the bottom up--the new cortex data's reading part five starts with the Perceptron Learning auto-tuning

We already know from the last sectionperceptron Not a smart algorithm, On the basis of it we need to solve the problem of the assistant. Since it is the adjustment of parameters, always have a goal and means to see how these great gods do. First set goals, for life often face the problem of classification, such as: In the crowd to find out the rogue molecules, we already have 3000 7000 The file records . We want to find a way to give a new file, according to the situation of this file to identify the 10001 The individual is not a rogue. Well, it's not ethical, and the rogue seems to be too much, but let's just do it. First of all, we first do this thing in the mode of employing. Please comb the application questions first.

        1. 10000

2. each sample has 2 attributes gluttonous X ( corresponding to P1) and the degree of Lust Y(corresponding to P2) and whether it is rogue A.

        3. find a new person to see, is not a rogue.

If you're lucky, we can get the picture on the right (blue is rogue). At this time the students immediately found that, ah, the original greedy and lustful basic are rogue ah. So simply draw a line in the middle to distinguish between these two categories of people. As soon as the newcomer arrives, it is a one-point line. This example can be used to summarize the logic of basic work immediately.

What is perceptron capable of? He is not the most like the line, in accordance with the previous section of the way to pull perceptron in. Try to find a way to let Perceptron himself to draw the lines. First, a stupid, computer favorite loop, write a loop it all the possibility to try, the last line to draw out, this line? Of course, this perceptron is already "smart", and you just realized the first learning algorithm. In the next very satisfied, after the same problem it can be solved with this method.

So this "machine learning" is essentially guessing the rules, and "learning" is the process of guessing in a specific way. For 10,001, the likelihood of a rogue being misjudged is a possibility. And then look at the line that we are employing the same, in the middle position are OK, 4 lines for these 10,000 samples are positive results, since there are so many possibilities, they inevitably have differences between, can find a good point out of it. So, here comes the second requirement, that is, how to effectively evaluate the results of the speculation, and then follow the assessment results to the best left. Specifically, we use the variance to evaluate it (in fact, you can evaluate it all, depending on your specific scenario), we follow the most in the middle. This basic logic is quite consistent with the basic pattern of cognitive neuroscience. Such a basic "machine learning" framework and algorithms are available.

This mode of work is the basic mode of "machine learning". It has a tall name, "supervised learning (supervised learning)". That is, the process of adjusting the parameters of the classifier to the required performance by using a set of samples of a known class. The process of looping inside is called "training".

Looking from the bottom up--the new cortex data's reading part five starts with the Perceptron Learning auto-tuning