[DemonRaven]

I am curious is deep learning, neural networks and machine learning the same thing or are there actual differences between the terms. Different articles on the web use all three but reading them they all pretty much seem like the same thing.

[ivan.moony]

I'm not quite an expert, but I think deep learning and neural nets are denoting the same family of algorithms, while machine learning can stand for whatever learning algorithm.

[keghn]

It is a sub section on ML. But as of lately it is dominating the whole field.

[Don Patrick]

What Ivan said. Machine learning applies to all forms of computer learning techniques, but it is often used by journalists as if it only means neural nets, the current most popular form of machine learning. Deep learning is several layers of neural nets, so it is basically more of the same tech.

[DemonRaven]

Okay thank you for explaining the difference. It makes more sense to me now.

[keghn]

My thoughts on general neural network

 NN are a bit like black box magic because they 
use "Occam's razor" logic very lightly. Not seriously at all.
 If they did, a pixel on the input side would be outputted
directly to the corresponding pixel on the output side.
 But instead there is the dark layer. A plumbing system
that can amplify or de amplify the input pixels
 and then mash them all together, then separates them back
out and then output all the processed pixel to the viewing
side.
 All of this is controlled by the weight values in the dark layer.
 So with a given image, it can be trained, by changing the
value of the weight, until it look like the input.
 It is possible with a few pixel out of many hundreds to
build a image. And all of the others pixels are de amplified out.

 It is possible to generate 10 time more date from the input pixels. And then
store them in  another layer,  a contextual layer. Then
use a little bit of everyone of them, including all the extra generated extra pixels, to generate a copy of the input on the output.


 So there is a lot of play here, and yes there
can be may way of building the original image for
a auto encoder neural network. And along with all the
near fantasy images.

 The more layer the more images that can be trained into a
NN. And programming the all weight states hold maximum
amount of images may mean some images get dithered. Like
the ones that are not that important.

 Then there is a classification NN.  When it see a object
in a image one pixel on the output side turns on. It
is also trained to do this. A program watch for which
pixel goes high and activates the pre selected label attached
to it like the word "dog".

[spydaz]

Neural networks is a sub technique of machine learning , deep learning is actually only connected to neural networks .... As it is a neural network with many layers .

Machine learning techniques , are mainly used to classify data as being either of one class of data or another .   
The KNN algorithm , clusters data in to groups which when plotted produce a scatter chart and the points which are closest to each other are considered to be similar.
Regression , is simular to a growth chart, and can be considered to be predictive .
The apriori algorythm is used to find frequent item sets ... Based on frequencys , predicts which sets will be most frequent ...
All can be also considered to be "data mining"
All can be considered to be unsupervised or supervised learning techniques .

Problems with machine learning is when you have outliers , data which is randomly skewed , this affects all techniques . So they are removed ... (Reshaping the data) . These will create bad classification or predictions . The data is split into training set , and test sets for use by the algorythms .... Problems arise again when you tune your algorythm . As you can overfit your data. Meaning it works with your reshaped data but not with unseen or new data.
Neural networks : deep learning has this problem .... Throwing mire layers at the problem does not solve it , overfits it plus takes longer to process it.
The new buzz about neural networks is hash tables .... Before it could only produce binary or numerical output , but anything can be stored as a table ie; groups of pictures representing a number 4 ... The pictures can be represented as a numerical input . The output produced can ve matched to a lookup table to produce is this image a number stored jn the table. This enables for more complex classification of pictures for instance .

Artificial intelligence goes beyond these techniques. Often considered simular but are greatly distanced. As artificial intelligence has computer programs and object oriented/ polymorphic programming algorithms. Code reduction & refactoring produces faster execution times .

Forgot;

Business intelligence/big data is machine learning but with data warehousing .... Relational or star schema or data vault . Again reshaping data ...or pooling data and using "slices of data"
Such as , sales vs frequency of sales vs location of sales = 3d query .... Multidimensional queries .... Only possible when you have a data warehouse.

[ivan.moony]

What a mess, DemonRaven, you have to admitt. Wait until you hear about back-propagation techniques of neural networks...

Huh

Edit: Anyway, NN is just about recognizing patterns in image, sound, or something else. It is not about real intelligence and conclusions, that has to be programmed by some other, yet unknown way.

[DemonRaven]

What a mess, DemonRaven, you have to admitt. Wait until you hear about back-propagation techniques of neural networks...

Huh

Edit: Anyway, NN is just about recognizing patterns in image, sound, or something else. It is not about real intelligence and conclusions, that has to be programmed by some other, yet unknown way.

Ya it is alot but I do understand the concepts better now thanks.

[keghn]

Restricted Boltzmann Machines - Ep. 6 (Deep Learning SIMPLIFIED):

[keghn]

Deep Belief Nets - Ep. 7 (Deep Learning SIMPLIFIED):

[spydaz]

The SImplest Method is always the best method
(Dartmouth College)