EMP is a god, now I understand NN
- Thread starter eliquid
- Start date
OK I promise to write a more detailed, easy to understand explanation and examples of what the applications are if there is interest.
So... Any takers?
::emp::
So... Any takers?
::emp::
You can use a neural net for function learning. Suppose you have a list of numbers being fed into a function, and you have another list of numbers that represent the output of the function for each input.
You could take these numbers by hand and plot them in a 2-D space, and if you had enough, you could start to see the curve forming on the plane, and possibly guess if the function is linear, exponential, ...
If the function outputs across multiple dimensions and forms more complicated surfaces, like terrain, it gets harder to predict what an output might be. So you can use something like a neural net that takes in multiple inputs and gives outputs based on what it has learned about the surrounding area.
In the past, I have used a neural net to do OCR on CAPTCHAs. Essentially, if you know what an 'A' is supposed to look like, across 1000s of fonts, you can take an 'A' that has been warped by one or more functions, and learn the inverses of those functions to take the character back to a recognizable state.
I'm probably still going to use one for a NLP app I'm doing thesis work on.
Something else that is cool to check out are genetic algorithms. The black boxes they use on Wall Street usually incorporate either a NN or GA.
You could take these numbers by hand and plot them in a 2-D space, and if you had enough, you could start to see the curve forming on the plane, and possibly guess if the function is linear, exponential, ...
If the function outputs across multiple dimensions and forms more complicated surfaces, like terrain, it gets harder to predict what an output might be. So you can use something like a neural net that takes in multiple inputs and gives outputs based on what it has learned about the surrounding area.
In the past, I have used a neural net to do OCR on CAPTCHAs. Essentially, if you know what an 'A' is supposed to look like, across 1000s of fonts, you can take an 'A' that has been warped by one or more functions, and learn the inverses of those functions to take the character back to a recognizable state.
I'm probably still going to use one for a NLP app I'm doing thesis work on.
Something else that is cool to check out are genetic algorithms. The black boxes they use on Wall Street usually incorporate either a NN or GA.
I use C++ and python. There are plenty of Java implementations out there; since these things should ideally be run on dedicated machines, you could probably get away with using Java.
So C++ For the heavy lifting and python for the presentation layer? Man i havent used C++ in a long time!
Some further watching:
CS 229: Machine Learning
https://www.youtube.com/watch?v=UzxYlbK2c7E&list=ECA89DCFA6ADACE599 (intentionally broken link, gives you the full course)
CS 229: Machine Learning
https://www.youtube.com/watch?v=UzxYlbK2c7E&list=ECA89DCFA6ADACE599 (intentionally broken link, gives you the full course)
For any visual stuff OpenCV is good a solid algorithm library with interfaces in many common languages: ABOUT | OpenCV
I was reading a local science magazine and found mention of neural networks, so just posting what I read there.
Neural Networks
As the names suggest it's branchof AI & biology based.
Neural Networks is about training your data on the basis of existing examples. The knowledge gained from training data sets is used on fresh data sets. This technique is used in image recognition.
Here is a practical analogy to understand this. Suppose you ask for an advice from your parents, teachers and elder brother or sister. They all give you their opinions, suggestions and all. Whose advice would you consider the most? Your priority order may be like – parents (parents and grand parents are first teachers of the child), teachers and elder sister or brother. You are attaching some weight (wis) or factor of importance to their views (xis) and y is your final conclusion. This leads to the backpropagation model that finds many commercial and industrial applications.
In the actual model, training sets are used to obtain these weights. Then once the satisfactory values are obtained, we use them on actual/real data to find y.
Neural Networks
As the names suggest it's branchof AI & biology based.
Neural Networks is about training your data on the basis of existing examples. The knowledge gained from training data sets is used on fresh data sets. This technique is used in image recognition.
Here is a practical analogy to understand this. Suppose you ask for an advice from your parents, teachers and elder brother or sister. They all give you their opinions, suggestions and all. Whose advice would you consider the most? Your priority order may be like – parents (parents and grand parents are first teachers of the child), teachers and elder sister or brother. You are attaching some weight (wis) or factor of importance to their views (xis) and y is your final conclusion. This leads to the backpropagation model that finds many commercial and industrial applications.
In the actual model, training sets are used to obtain these weights. Then once the satisfactory values are obtained, we use them on actual/real data to find y.