A discussion on the classical algorithm of machine learning-artificial neural network

Update: The article migrated to here. Http://lanbing510.info/2014/11/07/Neural-Network.html, there is a corresponding PPT link.
Note: Organize the PPT from shiming teacher
Students who can't see the picture can open the link directly: Https://app.yinxiang.com/shard/s31/sh/61392246-7de4-40da-b2fb-ccfd4f087242/259205da4220fae3

Content Summary 1 Development History2 Feedforward Network (single layer perceptron, multilayer perceptron). Radial basis function Network RBF 3 feedback network (Hopfield network).Lenovo Storage Network, SOM. Boltzman and restricted Boltzmann machine rbm,dbn,cnn)

Development History




single-layer perceptron 1 Basic model

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:473px ">

content=# "style=" >

content=# "style=" >
2 Assume that the excitation function is linear. Direct calculation with least squares available

content=# "style=" ">3 assumes that the excitation function is sifmoid function and can be iterated (one-time or sample-by-update)

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:487px ">

content=# "style=" ">" only do a simple derivative expansion. Very easy derivation

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:496px ">



Multilayer Perceptron

1 Basic model

2 Example (Multilayer Perceptron MLP with one hidden layer)
Model:

content=# "style=" >
Y=h (v) =h (H (U))
Solving:

content=# "style=" "> here how to convert to 6 K (xi)?

Then the weights of the two layers are then divided into two different values:

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:478px ">

content=# "style=" >
Then the update is available, reverse propagation (BP)

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:474px ">



3 Experience

content=# "style=" ">4 Advantages and disadvantages

content=# "style=" >


RBF Neural Network

1 Models

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:506px ">

content=# "style=" ">2 solution
3 Strengths and perspectives

content=# "style=" >

content=# "style=" >

content=# "style=" >




A brief introduction to deep learning

1 Forward Neural network

content=# "style=" >
2 Development history

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:459px ">

content=# "style=" >

content=# "style=" ">3 overview

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; width:492px ">
4 Some worth paying attention to the academic industry

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:478px ">

content=# "style=" >






Belief Network & Hopfield Network & Boltzman Machine & RBM structure at a glance

1 Belief Network

content=# "style=" ">2 Hopfield Network3 Boltzman Machine

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; width:493px ">4 RBM-limited Boltzmann machine




RBM

1 Models

content=# "style=" >

content=# "style=" >

content=# "style=" >

content=# "style=" >

content=# "style=" "> Use formula. Can get

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:494px ">

content=# "style=" >

2 Solving the CD algorithm

content=# "style=" >

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:484px ">

content=# "style=" >

content=# "style=" >

content=# "style=" >





DBN

1 Models

content=# "style=" >

content=# "style=" ">2 Training for feature extraction

content=# "style=" >

content=# "style=" "> Category-oriented

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; width:475px ">

content=# "style=" >

content=# "style=" >





Dbm

Model


CNN 1 Models

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:514px ">
2 Training

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; Width:514px ">

resizesmall&width=832 "class=" En-media "style=" margin:0px; padding:0px; border:0px; max-width:100%; Height:auto; width:500px ">


References

content=# "style=" >

content=# "style=" >

A discussion on the classical algorithm of machine learning-artificial neural network