How to learn artificial intelligence?

Step 1: Review linear algebra. (I forgot a lot about the line generation of the scum-_ ||)

I was too lazy to read a book and directly used the famous MIT open class: Linear Algebra, which is easy to understand and effective, and the SVD and Hilbert space that will be used in the future are introduced;

Advertisement: I summarized a set of notes while watching Github-Zlotus/Notes-Linear-Algebra: Notes on Linear Algebra.

Step 2: Introduce the machine learning algorithm.

because I was lazy, I directly used the famous open course of Stanford University: Machine Learning Course, the video of Professor Andrew Ng's old version of cs229, which was very detailed (the goal of the algorithm->; Mathematical deduction->; Pseudo code). The only disadvantage of this tutorial is that it doesn't introduce the neural network of the recent fire, but it is actually an advantage, which makes me understand that all algorithms have their own application fields, and not all problems need to be solved by neural network;

To say a little more, this course introduces in detail the general linear model, Gaussian series model, SVM theory and its implementation, clustering algorithm and various related applications of EM algorithm, PCA/ICA, learning theory and Markov series model. The class notes are in: CS 229: Machine Learning (Course Handouts), which are also very detailed.

Advertisement: I summarized a set of notes while watching Github-Zlotus/Notes-LSJu-Machine-Learning: Machine Learning Notes

Step 3: Try to implement the algorithm with code.

still because I am lazy, I continue to use the famous machine learning | Coursera directly, or the course taught by Professor Andrew Ng, but this is a very simple version of cs229, which almost teaches how to quickly realize a model in matlab (this tutorial contains the basic concepts and implementation of neural networks). The disadvantage of this course is that the difficulty is relatively low and the derivation process is very simple, but this is also its advantage-let me concentrate on transforming theory into code.

advertisement: refer to github-z lotus/coursera _ machine _ learning _ exercises: machine learning by Andrew ng from coursera for homework

Step 4: Implement a fully functional model by yourself-in progress.

or because I was lazy, I found the course video cs231n Winter 216-YouTube, a class taught by Professor Li Feifei, and lectured by Andrej Karpathy and Justin Johnson, mainly introducing the application of convolutional neural network in the field of image recognition/machine vision (the previous neural network code was not written enough? This course is so exciting that you are writing from scratch everywhere. The homework of this course is even more intimate. It is arranged directly with Jupyter Notebook, and you can run it locally and check your own mistakes. Mainly use Python and Python series of scientific computing libraries (Scipy/Numpy/Matplotlib). The translation of class notes can refer to the intelligent unit-Zhihu column, which is mainly translated by Du Ke, a friend. It is very well written ~

To say a little more, this course is more distracting for programmers, because it is not biased towards algorithms and models like the homework realized by matlab in the previous step. The model realized by Python in this course pays attention to software engineering at the same time. Including common forward/backward encapsulation of layer, self-defined combination of layer, how to compose layer into network, how to integrate batch-normalization and dropout functions in network, how to do gradient check under complex model, etc. In the last assignment, there are the manual implementation of RNN and its gay friends LSTM, the compilation of CNN visualization function which is helpful for debugging, Google's DeepDream and so on. (You can basically understand all kinds of popular picture style transformation programs after you finish your homework, such as Cysmith/Neuro-Style-TF) In addition, the homework of this course is very highly praised by computational graph, and I don't know if it is my illusion ... It should be noted that lecturer A.K's speech speed is extremely fast. Fortunately, YouTube has the function of automatically generating commentary, and the accuracy is not bad, so it can be viewed as subtitles.

advertisement: refer to github-z lotus/cs 231n: cs 231n convolutional neural networks for visual recognition (winter 216) for homework (I added some deduction calculus to my notebook ~ it can be used as reference: d).