Introduction to neural networks

AI, machine learning , The relationship between deep learning

The process of machine learning

1. Data acquisition
2. Feature Engineering
3. Build a model
4. Evaluation and application

The difference between machine learning and deep learning

Artificial intelligence is more artificial in machine learning , Manual selection data , Manual selection features , Manual selection algorithm , Get the data manually , It gives people the feeling that they just realize the mathematical formula . And in deep learning , Give the data to deep learning , It will automatically find out what parameters , The combination is appropriate , Give people a sense of intelligence . Solved the part of feature engineering .
Algorithms are important in deep learning , But the most important thing is feature extraction .
Deep learning is widely used , It's all about , For example, self driving cars. , Face recognition , In medicine, for example, cancer cells are detected , Detect whether the cells have mutation , The way genes are combined , Prediction of protein structure, etc ！

The past and present lives of neural networks

A.D. 2009 year , There is a foreign scientist lifeifei , Specialized in computer vision , At that time, very few people did computer vision , There is no intelligent detection , Face recognition . At that time, lifeifei thought that this major was quite popular , Because there are too few people doing it , How can we arouse people's sympathy ？ Just think of a question , Because there is no common goal , There is no common data , So it is troublesome for us to communicate , I wonder if I can create a dataset ？ after 6 year , They called on many scientists in Colleges and universities to do a public welfare event , Is to collect images , Label image . In general, deep learning or machine learning are supervised problems , The neural network we are going to talk about also has the problem of supervision , Such as face detection , First of all, I must tell the machine where my real face is in the picture , Don't underestimate this matter , To use the face one by one 68 Mark the points , He is driving his face crazy . It's not a simple thing . stay 89 In, she called on scientists to do , It took a year to complete a data set , It's called IMAGENET, Contains 22000 Species ,1400 Ten thousand pictures , The amount of data is huge , It contains almost everything you can think of . We can't play with notebooks , At the very least 8 individual W Starting server . What if you want to play , You can use smaller datasets , such as CIFAR-10 After the data set was created, a competition was held , stay 09,10,11 Not many people took part in the , The results of the first place and the tenth place are only a little worse , until 12 year , There is ALEX, The neural network in deep learning won the championship by several percentage points beyond the second place , The second algorithm is the traditional artificial intelligence algorithm . thus , Scientists have discovered the outstanding effect of neural networks in computer vision , Be recognized , It's developing rapidly . until 2017, Neural network recognition exceeds human recognition ability , This competition just exited the stage .
Of course , When the data scale is relatively small, there is no difference between deep learning and traditional artificial intelligence algorithms, or even no deep learning , When the amount of data reaches more than hundreds of thousands, the advantages of in-depth learning are highlighted .

Computer vision

What does an image look like in a computer ？
A picture is represented as a three-dimensional array , The value of each pixel is from 0 To 255

The routine of machine learning

1. Collect data and give labels
2. Training a classifier
3. test , assessment

Linear function （ Scoring function ）

Mapping from input to output

Like a picture of a cat （32323） after f（x,W） Get a score for each category , For example, the score of the cat , Dog score , Car score .
there W Weight parameter means weight parameter
b Represents the offset parameter , Play a fine-tuning role .
If you divide the cat a little , There's a total of 3702 Pixels , Each pixel has a different effect on the result , Some play a facilitating role , Some have an inhibitory effect , All the weight parameters are introduced here W, The weight of each pixel is different , therefore 3702 Pixels correspond to 3702 Two weight parameters .

there W It stands for the score 10 Species , Each species has 3072 A weight ,x For a picture , Yes 30721 Pixels .
b by 101 On behalf of this 10 Fine tune both values , Be careful ： Fine tune each category , It is troublesome to list them here one by one .

Final , Multiple groups of weight parameters constitute the decision boundary

Loss function

The parameters in the matrix are random . Big data means high weight , The parameters here identify a cat as a dog . Surely not , But how bad is this parameter ？ There must be a definition , The loss function is introduced ：


Sj Represents other categories ,Syi Represents the correct category ,1 The expectation bias on behalf of oneself refers to
Li The more close to 0 The better this group of weight parameters .