Though they really true to false, inaction Department has also not there.
Business road, the road of life
the essence of science and technology should be these five words: truth and seeking
I often wonder, do artificial intelligence technology companies in the end how low the threshold. Theorists can not gymnastics, gymnastics circles no theory. The vast number of papers, the vast number of concepts being promoted various technology companies. Enterprises seek to maximize the benefits of, really let things become pragmatic worthless? ? ? ? ?
"On the deep learning (image orientation) of the underlying business logic from the perspective of the practical operation" is one of my series of essays, today to talk about some of the basic concepts of learning in depth, whether deep learning or artificial intelligence, are very experimental field, we in books, academic papers on the theory see, are likely to be overturned, so today is to share this stage of human consensus.
The concept of (a) the concept of several data and several variance
1. MINIST Dataset
- Collected a lot of people handwritten 8, 9; then we have to enter a digital computer, computer to understand what this is, this is a machine learning, most entry field of artificial intelligence data classification.
2. CIFAR10 data set
For the FIG., There are 10 class of objects, some of which are relatively similar objects, some not easily distinguishable categories, such as deer and more like a horse, more like cats and dogs, etc. There are not easily distinguishable categories, a total of there are 10 categories.
3. ImageNet data set
Here the picture is more complex classification problems
4. The within-class variance (intra-class variance) and the between-class variance (inter-class variance)
The above three data sets, from inside the computer to identify any of the above picture is what's perspective, how do we identify the degree of difficulty of these data sets do?
Such as MINIST this handwritten digital data set, why do we say that it is relatively simple? For example 1, have written have crooked twist, but it is 1, other numbers are also the same; but like CIFAR10 this data set, the inside of the cat, cats have different perspectives, different varieties, different background, other than to write a number on a black background is hard to determine.
Generally it believed that the same category, the greater the differences within the same class of, such as cats and cats, the problem is more difficult, which is within class variance; for example, to distinguish between the cat and the house is the variance between two classes.
(B) the depth of learning what can be applied in the enterprise
jacky (Zhuyuan Lu) is summarized with data-driven
Let me talk about the depth of learning to solve the problem roughly path:
- The computer to do such a thing: the number of pixels mapped to a category label, we need to find a fitting function data, this process is the process of deep learning.
Similar to the classification of deep learning, computer theory actually has been for decades, the earliest time, it is based on some manual of rules, for example: a person than 1.75 meters tall, weighing more than 100 pounds, then we think the above description is a man, a manual of rules like this; but now, some of the more complex issues, the use of artificial methods to write rules have no way to find such a function, because such functions have been too complicated, and now mainstream way of artificial intelligence is data-driven, data-driven machine consciousness is to say I let you learn from the data, find the law from the data, we can give one million images on a computer, let it come and go training, we want the computer fit yourself out function from the data we want to solve the problem of classification of our social reality.
The basic method (c) the depth of learning - classifier
A basic architecture classifier, we enter a Zhanger Gou, to see that he is not a high-quality single dog:
- Our first step is to get this input is Zhang Ergou;
- Step 2 is a very important step is feature extraction, we like the image regardless of Ye Hao, Zhang Ergou or, finally do judgment, we extract the features of Zhang Ergou three, one is he is not high, a that he was not handsome, he was not a rich. We call this process is called feature extraction . If we take the high, handsome, rich these three features together, a name called the feature vector . We find this feature of the process is called feature project .
Examples of the above data-driven approach is not used, the method is manually set, but we can better see the data extraction process and make such a determination based on the characteristics.
Zhang Ergou a judgment we do not need such a deep learning means, why we can use machine learning in image classification, even deep learning this work? Because of the nature of the difficulties picture classification: Feature extraction is a very, very difficult job. There are some issues clearer features, but there is a difficulty image features is that it's very difficult to design features.
Machine vision development for decades, traditionally no way to learn by machine or depth of learning-based, through manual design, there are many ways to feature extraction: as HOG, SIFT, SURF, and achieved some results, but the bottleneck is also evident , is the face of a more complex problem, is powerless.
The most critical depth study of the cattle X, characterized in that the learning can be performed, the self-learning feature of the training data. Feature to find a good or bad determines the upper limit of the classifier and classification methods, neural networks, whether good, or is this traditional method of random forests or, just to get close to this limit.
