Andrew Ng's machine learning course notes - the legendary story of the godfather of deep learning

The interview with Geoffrey Hilton in the third lesson of Andrew Ng’s course was very exciting.

The richness of the content goes beyond the machine learning content itself.

Thanks to my friend Haobo for the recommendation.

Backward learning mechanism - similar to the related content in the brain-inspired lecture

unsupervised learning

Discussion with Symbolization - An idea is a large vector with causal power.

Suggestions:

If you have a good gut feeling and most people say the idea is stupid, it's a good idea, stick with it.
Go into business or study for a PhD?
Don't stop programming

Why do you always use mind mapping? Because it's like our brain stores things. This neuron-like structure allows our brains to make better direct connections (?) - I guess. Try drawing a note, something graphic, which may help you remember it better.

Links:

Courses:

http://open.163.com/movie/2017/8/9/9/MCS5AQH7J_MCS5PHV99.html

https://www.deeplearning.ai/

Notes: https://pan.baidu.com/s/1Vp3FZpppjEOVIq-sh-Q2og

machine learning

1 Geoffrey Hilton

1.1 1.1.1 High school mathematics -1966 Hologram -Brain

1.1.2 Biology, Physics -only one

1.1.3 Philosophy - Can't tell right from wrong, Psychology - Too simple

1.1.4 AI with Longuet Higgins -givingup

1.1.5 Got Phd, cound not get jobSolon Scholarship

1.1.6 Don Norman and David Rumelhart1982

1.1.7 1982 There is also a previous paper on back propagation into Naturestraut - British Can Learn Dictionary

Features

1.1.8 How brain works vs psychology

Concept like amap

like a graphcalstructure

Graph like-tree structure and human face

eigenvector

1990 Text getreal words.

1996 Computer is slow

boltzmann machine - the most beautiful thing loop - neural network constrained first time result can be used as new data

S-confidence network

Learning deepband

calculus of variations

Approximately-Estep

Variational Bayesian Methods-1993

Dropout ReLU

2014 ReLU inGoogle

Back prop andBrain

1987 Loop algorithm input - hidden bypass input

Skipe - Spike Timing Algorithm?

2007 A group of B machines can reproduce spike timing

fast weights

1972 Neurons for circulation

Jimma ba

capsule algorithm

There are important keys in the transfer

multidimensional individual

There is a special individual

a bunch of neurons

different capsules

Different properties

regular agreement

Normal only has a bunch of

Maybe the mouth, maybe the nose

Different capsules come to Vote

Possible corresponding space

Different from Filtering

Changes in concepts - FIFO supervised learning, can have different learning

● Forward learning is different

You can also do backward learning

● Google Team in Toronto

backpropagation

Unsupervised learning - weigstein

It's good in the long run.

In the past 10 years, supervised learning has been

Unsupervised learning still matters

we haven't reached it yet

Variation method change code

Generative Adversarial Network

capsule

sparsity

slow characteristics

record first

Doing nonlinearity

Transformation matrix

pixel to coordinate

capsule

Matrix calculation

Suggestions