Three levels of AI engineers _ depth Learning

Three levels of AI engineers

The arrival of each wave of waves means that there is no one to occupy the blue sea, it also means a lot of new growing giants, but also what it means. A large number of technical personnel needs, the demand for market development, as well as the high salaries of practitioners and numerous opportunities.

The most common thing we do is to look after the last wave of the aftermath of the far away, sigh that their untimely, but did not realize that the next wave has come to our feet.

Yes, we're talking about AI.

People in the It circle, should have a intuitive understanding. At present, no one in the domestic well-known Internet enterprises to establish their own artificial intelligence technology team, in order to use AI technology to enhance the product experience and intelligent degree.

But at the same time, a variety of obscure nouns also scare off a lot of non-trained developers. What is called convolution neural network. What is called convex optimization. Do you have to go back and reread the high number, line generation, probability. So a lot of formulas, feel completely don't understand AH. I've heard that no famous school doctor is born.

As a developer, the AI domain seems to me to be divided into so many levels of academic researchers, a long time ago.

Their job is to explain theoretically all aspects of machine learning, to try to find out "why this design model/parameter works better", and to provide better models for other practitioners, and even to advance the theoretical research step forward. Able to do this step, can be said to be very rare, talent is not around the mountains, opportunities and efforts are indispensable. Algorithm Improvement person

They may not be able to answer the "my method why work", perhaps without the great achievements of Hinton,lecun, but can play the existing models better with experience and some fantastic ideas, or suggest some improved models. These people are usually the backbone of each machine-learning giant or a growing unicorn, and it's not a problem for them to use any model, and there are usually a few fixed choices depending on where they are. At this level, insight and idea are important things, the different tools of the difference, the impact is really not so big. May make a result early or late for a few days or weeks, but it is not likely to affect the "no results." Industry-Implementing people

These people are basically not involved in the algorithm domain too deep, that is, to understand the implementation of each algorithm, the structure of each model. They are more based on the paper to reproduce the good results, or use other people to reproduce the results, and try to apply it in industry.

For most it people, it's good enough that the third category, the industrial implementation level, has already had the opportunity to take part in the big time, and that alone has defeated 99% of the nation's people (the squint-eyed smile).

Not only is the ordinary program ape so said, the art of the program Ape and ... Well, the tall program apes say the same thing.

I said, hehe.

There's only one answer: Just do it (go to it, teen)

Become an artificial intelligence engineer, in my opinion, want to machine study, deep study grasps good, can go into the fight. In addition, the theory must be combined with the actual project: Because as a programmer, read 10 times the book than run the program, rather than spend a lot of time to Shuben, as a hand to complete their own program and run it. When we write the code, we will know where we are not clear enough to learn. 02 Introduction to Machine learning

Let's start by saying what machine learning should learn.

Learn anything, do not build a platform in the floating sand (please raise your hand if you are familiar with this sentence), there are some basic knowledge or need to master. For example, in the field of computer vision, according to our team training experience, in order to be able to independently carry out the development of machine learning, it is best to complete the first few courses: Induction machine learning

Familiar with the classical algorithms, models and tasks in the field of machine learning, and learn to build and configure the machine learning environment, and learn to solve a practical problem with linear regression. Logistic regression analysis, neural network, SVM

Grasp the data set exploration, understand the classification task algorithm (logistic regression, neural network, SVM) principle, learn to use each classification algorithm to classify specific tasks under the Scikit-learn framework. Decision tree model and integrated learning algorithm

Loss function: Information gain, Gini coefficient
Division: Exhaustive search, approximate search
Regular: L2/l1
Prevention of fitting: pre-pruning and pruning; bagging principle; boosting principle;
Popular GBDT tools: Xgboost and LIGHTGBM clustering, dimensionality reduction, matrix decomposition

Principal component Analysis (PCA), independent component Analysis (ICA), nonnegative matrix factorization (NFM), implicit factor model (LFM), Kmeans clustering and mixed Gaussian model GMM (EM algorithm), Attractor propagation clustering algorithm (Affinity propagation Clustering algorithm) feature engineering, Model Fusion & Recommendation System implementation

Learn common data preprocessing methods and feature coding methods, learning characteristics of the general principles of engineering, combining various characteristics of engineering and machine learning algorithms to implement the recommendation system.

The above course will probably cost you more than 1 months of all your spare time. But believe me, it's worth it.

If you don't get a month or two of your spare time, well, I'll tell you one of the most basic requirements, to meet this requirement, you can be a machine learning primer: matrix multiplication Matrix multiplication

Don't laugh, seriously, in this frame of highly encapsulated age, the gradient does not need to calculate itself, the loss does not need to seek their own, the reverse conduction is handled properly, in the case of superficial understanding, you even need to know so many concepts can begin to write the first program:

It is through a series of matrix operations (or some other similar operations) to map the input space to the output space. The value of the matrix that participates in the operation is called the weight, and it is necessary to find the optimal value by iterative iteration.

The current weight value is far from the optimal value, expressed by a number, this value is called loss, the function that calculates this value is called the loss function.

The current weight value should be adjusted or lowered, this value by the loss function to determine the derivation, the derivation of the function is called gradient.

The method of updating weights by loss and gradient is called reverse conduction.

The iterative approach is called gradient descent.

Although the program written out must be unknown, but in fact, 20 years ago, when I first wrote Hello World in C + + is also a face, I believe that each can devote themselves to machine learning and development work of the Ape, are great perseverance and courage, nature will not lack the motivation and determination to continue to learn. 03 Getting Started with depth learning

Let's say what we should learn in depth study.

Deep learning focuses on convolution neural networks and cyclic neural networks, using a large number of real data sets, combining practical scenarios and cases to introduce the application scope and effect of depth learning techniques. Introduction of neural network and configuration of deep learning environment

Familiar with the common terminology in the field of neural network, install and configure the Depth learning Framework TensorFlow, learn to use TensorFlow to solve a practical problem. Neural Network Foundation and convolution neural network principle

Using different structure of neural network structure to verify the effect of network structure, to understand the related concepts and basic knowledge of convolution neural network, and to understand the characteristics of CNN local correlation and weight sharing through actual cases. The actual combat of convolution neural network

Image classification and Detection task: Learning image Classification task and detection task of the current main model algorithm, and through two combat case study in the TensorFlow framework to train the CNN model. An example of image segmentation based on convolution neural network

Master Segmentation Task Introduction, Deconvolution (DECONV/TRANSPOSE-CONV), FCN Loop Neural network principle

RNN Fundamentals

Threshold Cycle Unit (GRU)

Long Short term Memory Unit (LSTM)

Word Vector extraction: Word2vec

Encoder-Decoder structure

Attention mechanism models: Attention model

Picture callout (Image captioning)

Photo FAQ (Visual question answering) 04 Advanced

Congratulations, you've become a member of the AI Engineer Group.

You can then collect some of your own data and train your own identification engine, or try to optimize the model, feel the pain of the so-called resnet, or simply try to inception these more advanced networks to brush CIFAR , or you can try to learn from NLP or strengthen your learning direction. In short, these things are far less difficult than they seem.

Of course, no matter the road, learning, progress and self-motivation are inevitable courses. A new field, the vigorous vitality of the inevitable also means that the emerging results. The completion of the three courses I mentioned above can only let a person from the layman into the circle of people, have entered this field, to catch up with this wave of basic qualifications, as to whether it is to become a sea-goers or directly by the waves swallowed, or that sentence, not Labor is bound to have no income. Hard work does not necessarily make a fruition, but not to study hard, it is doomed to nothing.

Finally, wish you all the best in the field of AI.

Article reproduced from: http://blog.csdn.net/dQCFKyQDXYm3F8rB0/article/details/78302687