AI three elements essential
data
algorithm
computing power
cpu io-intensive tasks
gpu compute-intensive
tpu Google
AI "machine learning" deep learning - contains
195,019,802,010
relations
machine learning, artificial intelligence is a way to achieve
deep learning is a learning machine direction evolved
Father Turing Test Turing computer艾伦麦席森origin of the
Dartmouth Conference - - Artificial Intelligence 1956 the first year - with a machine to mimic human learning and development in other directions
Branch
of computer vision
NLP NLP
cover text mining / classification, machine translation, speech recognition,
speech-to-text speech recognition 1.
2. The text-to-speech (TTS)
problems: 1 voiceprint recognition similar to fingerprinting
2. Wake word cocktail party effect
text mining / classification
syntactic analysis, emotion detection, spam detection
bottleneck data are different, ambiguous
machine translation
is limited vocabulary problem
robot
stationary robot
mobile robot
computer vision,
natural language processing
Key ---- machine learning
machine learning to analyze the data obtained from the model and use the model to predict unknown data
Machine learning workflow
1. Obtain data
2. The basic data processing
3. Characteristics engineering - Key
4. model - a machine learning algorithm (model train)
5. model assessment
1. good - on-line service
2. bad - repeated iterations --2,3, repeat 4,5-oriented service line
Data Introduction
row of data is a sample of
a data is a feature of
some data and some data did not have a target value target value
Data type
Data type feature value + a target value (target is continuous, and discrete)
data type only two feature values, no target
Data dividing
training set - to build the model
test set - Evaluation results of the data model
dividing ratio of the column is generally 28 minutes
training set of 70% -80% -75%
test set 30% -20% -25%
Basic processing data
that is missing data values, outlier removal processing such
Engineering characteristics - focus on
the use of special background knowledge and skills in handling data, so that the feature can play a better role in the process of machine learning algorithms
significance: the machine will directly affect the effectiveness of learning
data and characteristics determine the upper limit of machine learning, and model and algorithm just approaching this limit it, determines the accuracy of the algorithm
Wherein the content comprises engineering
1. feature extracting
arbitrary data (text / image) into digital feature
2 wherein preprocessing
feature data - - through some process - conversion function - to - the feature data - converted - into a more - suitable algorithm model -
3. the feature reduction
aimed at certain defined conditions, to reduce the number of (feature) random variable, to obtain a set of "irrelevant"
Machine learning algorithm classification
Composition can be divided according to different data sets:
1. There supervised learning target characteristic values of supervised learning
objectives - continuous - Regression
goal - discrete - Classification
2. eigenvalues no unsupervised learning target
3 semi-supervised learning part of the data part of the data did not have a label tag
4. reinforcement learning alpha dog
Supervised learning
data types: a characteristic value + target
Regression
goal is - Continuous
Classification
goal is - Discrete
Unsupervised learning
only features no target value
Semi-supervised learning
characteristic value - the target
characteristic value
Reinforcement learning
goal is to get the most cumulative awards
Independent and identically distributed
independent of each sample are independent and do not affect each other, there is no relationship
with the distribution of each sample are subject to the same distribution of
independent and identically distributed sample of each independent and identically distributed
Alphago - a large number of sample data to learn - chess - supervised learning
Alphago zero - reinforcement learning
Model evaluation
according to different sets of target data
continuously regression
discrete classification
classification model evaluation
calculation accuracy, precision, recall, F1-score, AUC index
Regression models to assess
the root mean square error (Root Mean Squared Error, RMSE)
loss of information between the actual and predicted values measure
other
relative squared error of
the mean absolute error
relative absolute error
Two kinds of good results and poor
fitting - good - good fitting effect
is not good - fitting ineffective
underfitting: learning model is too rough, even the training set of sample data features had nothing to do out of school, too model simple, did not learn to features,
over-fitting: effective training set, test set ineffective, learning the training set is too full, the model is too complex, model jagged edges, unstable
Examples of machine learning models built Azure
1. Data acquisition
training set - Download
2. Basic Data Processing
3 wherein Engineering
4 model
The model assessment
to see the word document
Deep learning - the depth structure learning, learning level, the depth of the machine learning algorithm is a set of classes, a branch of machine learning is
the parent convolutional networks
depth study stratified
Neural Networks
Machine learning environment to install and use basic
installation - library
installed - jupyter notebook
enhanced version ipython is the Web version of this
editing mode enter
command mode esc
plus a line on a
lower b add a line
dd delete
m markdown mode
z
Shfit enter the code and run to the next execution Cell
Ctrl stay in the current execution Enter
-Pycharm- than an advantage in drawing and data showing direction
Execution code shfit + enter a new implementation of the unit cell