How to learn artificial intelligence by system

Three main points of artificial intelligence: English, theory, engineering.

English is very important, all the classical textbooks are English

1 translation is always not fluent, not as easy to understand the original

2) Lost in translation

The most important scientific research in all English

1 translation will never catch up with the speed of the new knowledge generation

2 The ability to innovate AI talent will also publish their own results in English

The community is the English environment

1) Github

2) StackOverflow

The AI Leader's work environment does not open in English

Theory

Linear algebra (recommended the Matrix Cookbook)

1 The main way to understand nonlinear systems is still through local linearization

2) Machine Learning algorithm involves a large number of matrix operations

Probability theory (elementary linear algebra and introductory probability theory should suffice)

1 Statistical Machine Learning Foundation is probability theory

2 Understanding the statistical properties of complex nonlinear systems is essential for analyzing deep learning algorithms

Computer algorithm (recommended: The ART of Computer programming)

1 AI is not only deep Learning

Large-scale data preprocessing, extraction, etc.

Online service

Embedded system, Resource Bandwidth Limited

2) algorithm optimization

Training 3 days Complete VS10 days complete VS30 day complete directly affect scientific research or product launch

Whether the GPU is running full, IO is a bottleneck

e.g., approximate Softmax

3) The concept of numerical computation

Convergence, speed of convergence

Machine learning theory (classical bible:pattern recognition and Machine Learning)

1) Although deep learning unified Lake, but the classical machine learning theory still need to know a probably.

Linear regression/classification (SVM, Lasso, Kernel, etc)

Clustering (K-means, etc)

dimensionality reduction (PCA, etc)

Probabilitic modeling (mixture model, EM)

Adaboost,etc

2) The insight of the classical machine learning has been repeated in deep learning

Restricted Boltzman Machine

Denoising NN

Machine learning theory can be deep and shallow

1 practical need not too deep theoretical foundation

2 Theoretical Foundation is indispensable in scientific research

Intuition is more important than theoretical deduction, but good intuition comes from solid theoretical foundation.

Engineering

It's important.

1 The theoretical framework of DNN is basically determined (unless quantum computer production, otherwise short-term change is not small)

2 TensorFlow come out, no more derivative number

3) model design requires a large number of experimental verification

4 Google as a large company, adhere to all the core system of independent research and development, with a sound code Review system.