Where did you guys learn Neural Network? I would like to make simple programs with FANN, but I need some materials to learn and I tought I could ask this for the best people in /g/
- Hardly any good PDF available
- Really long
+ It's the authoritative book on AI
+ Best notation
+ Pretty short and to the point
+ Requires leat math/probability knowledge
+/- Its python
- Total basics
- Notation is a bit shit
Murphy's "Machine Learning: A Probabilistic Perspective" is pretty okay but you'll need to know your maths pretty well.
Barber's "Bayesian Reasoning and Machine Learning" is a bit less math and nice as well. Plus it's an officially free book and has matlab code. It has the NN part kind of 'mixed in' with the rest though, so not great if you ONLY want to learn nn (which I wouldnt advise but well)
- There is no accepted notation around the field so there is like a compatibility problem between any two books/websites/wikipedia etc
- R&N's is the industry standard
This is not the 90's.
There are better alternatives to neural networks.
The concept of neural networks is not that complicated but the problem is that you cannot really explain what the model represents.
This makes the whole model worthless.
You can learn and implement a basic nnet in Andrew Ng's intro to machine learning course on coursera, which is free and available anytime.
NNETs are certainly not useless and are used in industry all the time for image recognition and similar tasks. FB, google, etc. all use deep nets.
i once feeled clever being able to do neural nets, but once you realize no one can explain WHY it actually works, you finally understand it's just about implementing a basic algorithm, with some math you don't even understand anyway, (code monkey really)
The thing is that the brain is far too complex to simulate using traditional methods. Neural networks is the only way we know that even comes close. There is likely a reason the brain evolved using the architecture it does.
Give one example where ANN is a good method, classification? decision making? we have better better methods to do those things.
If we should mimic the behaviors of the brain, we should focus on signal processing rather than ANN.
>replicating the behaviors rather than mimicking the construction.
But the construction of the brain ENABLES the behaviors to happen, You can make normal computers replicate behaviors because you know what you want it to do, but if you just provide the groundwork you get something unique
It's not going to do "something cool" just by itself.
It's not magic, it's just mathematics.
You can use neural networks for specific tasks like pattern recognition or separating data into groups.
as I said.
Welcome to the 90's.
Say I contract you to build me a system that could manage peoples taxes with a neural network.
In the beginning everything would be controlled by humans and over time, the system would be powerful enough to handle everything.
Would that be a good system?
Would that be a good way to design such a system?
It is not.
That is my point.
In the 90's people thought everything could be solved by ANN.
Because human brains can solve any problem, ANN would be the solution to replace humans without having to specify anything more than the data and the desired outcome.
Nobody thinks that way anymore.
What about situations where it could be applied?
Classification? depending on the data, there is simpler methods, there is better methods and there is faster methods. Random forest is a lot better if the data is complex.
Decision making? Maybe, but again making a model based on knowledge of the task is usually better. Monte Carlo would be more efficient for most games.