Click on the top " AI proper way ", select the "star" Public No.
Heavy dry goods, the first time served Transfers | Almost Human
Currently information about machine learning in an endless stream, including both books, course videos, there are a lot of open source projects algorithm model. But for beginners, learning to read notes is perhaps one of the most easy and quick method to get started.
This article is to introduce a 80-page long study notes, series aims to summarize the basic concepts of machine learning (such as gradient descent, back-propagation, etc.), different machine learning algorithms and popular models, as well as a number of authors in practice learned skills and experience.
If you are just a person entry field of machine learning, this study notes may be able to help you take a lot less detours; if you are not a student, these notes can also be a quick reference for you when you forget some models or algorithms. If necessary, you can use Ctrl + F search concept they want to know.
Notes link:
https://createmomo.github.io/2018/01/23/Super-Machine-Learning-Revision-Notes/#tableofcontents
The following notes divided into six parts:
Activation function
Gradient descent
parameter
Regularization
model
Practical tips
In the first part of the "activation function", the authors provide a Sigmoid, tanh, Relu, Leaky Relu four commonly used machine learning activation function.
The second part of the "gradient descent" is divided into calculation chart, back-propagation, L2 regularized gradients, gradient and gradient disappear section 12 small explosions:
In order to help the reader understand the author gave some examples, a lot of content and visual presentation:
Gradient descent
In addition, the authors of some of the symbols used in the code are explained in detail, very friendly for the novice:
The third part notes are in machine learning parameters, it can be divided into learning parameters and super parameters, initialization parameters, parameter adjustment over several excellent sections.
To prevent novice detours, at the beginning of the "initialization parameters" section on cautioned: Actually, TensorFlow machine learning framework has provided a robust parameter initialization function. There are a lot of similar reminder in your notes.
The fourth part is the regularization notes, comprising regularization L2, Ll regularization, Dropout, early stopping four sections.
The fifth part is the top priority of the whole note, a detailed description of logistic regression, multi-class classification (Softmax back), transfer learning, multi-task learning, convolutional neural network (CNN), series model, Transformer and other eight BERT class machine learning models. Further, the following model is divided into eight categories for various subclasses explain, as shown below:
Relatively simple explanation of the first four categories of machine learning model.
The most detailed explanation convolutional neural network (CNN), including Filter / Kernel, LeNet-5, AlexNet, ResNet, target detection, face verification and nervous style migration.
Series models, including the common Recurrent Neural Networks (RNN), Gated Recurrent Unit (GRU), LSTM, two-way RNN, depth RNN example, embedded word, example translation model to sequence and so on.
Transformer and BERT model.
Note the last part gives some "practical tips", including training / development / test data sets, data does not match the profile, enter the 6 aspects of normalization and error analysis. Some tips from Deep Learning AI and other online courses, some of it is the author's own summary obtained.
On the other notes
In addition to this machine study notes, before finishing off the author probabilistic graphical models, BiLSTM top layer of CRF and other related notes. Inventory as follows:
On finishing probabilistic graphical models reviewed his notes.
CRF-related notes on the layer of finishing BiLSTM.
Recommended Reading
(Click on the title to jump to read)