Deep learning (deep structured learning or hierarchical learning) is a set of algorithms in machine learning that attempt to model high-level abstractions in data by using model architectures composed of multiple non-linear transformations

To start up your work, please refer

http://gradworks.umi.com/36/04/3604470.html

http://arxiv.org/pdf/1310.6343.pdf

http://jmlr.org/proceedings/papers/v27/baldi12a/baldi12a.pdf

http://jmlr.org/proceedings/papers/v32/arora14.pdf

Thanks

Deep learning (deep structured learning or hierarchical learning) is a set of algorithms in machine learning that attempt to model high-level abstractions in data by using model architectures composed of multiple non-linear transformations

To start up your work, please refer

http://gradworks.umi.com/36/04/3604470.html

http://arxiv.org/pdf/1310.6343.pdf

http://jmlr.org/proceedings/papers/v27/baldi12a/baldi12a.pdf

http://jmlr.org/proceedings/papers/v32/arora14.pdf

Thanks

Dear Ismail, 

Could you please state more clearly on what you mean by "non-mathematical NP complete problem". In addition, please also explain why you considered the problem as NP complete. I am curious on this. 

Deep learning as Hari has said is hierarchical learning, as such it takes a divide and conquer strategy to it's problems. It just breaks them down in a number of levels and builds a hierarchical model of them. It relies on large data sets to fill the lower levels of the model, and the idea is that the structure of the problem must not have too deep a pattern of hierarchical organization or the problem will not be completed before the heirarchical maxima is reached. If it maximizes the hierarchical structure too early, the result will be ambiguous answers.

 Dear Mr Mohamad Ivan Fanany,

I am in the problem of sensor placement in an environement. Suppose you have an MxN meter room, and you wanted to put k number of sensors in that room. The goal is to place as optimal sensors as possible in that particular room, for which the coverage of the sensors must be maximized in that room. The sensors for example can be a beacon tower for mobile robot positioning.

Of course we could formulate a cost function for this problem, and solve it by using any techniques of optimization.

But i am very interested if somehow we could solve it by deep learning. Some thought though.

Cheers!

Dear Abdul,

Your problem is actually quite mathematical. If you consider your floor as a graph and the sensing range of each sensors in term of a collection of nodes, you could probably formalize it in mathematical terms. Then you could solve it, as you said, through standard optimization methods...

The deep learning, or any machine learning, approach is interesting. I'd be curious to see if you had any hint with this... If you feed enough verified solutions to the network, would it become able to generate solutions based on, let say, some random seed or partial solution...

The main problem I see is the formalization of the input to the network. For each different "size" of input you would need to train a new network. Unless you work with an algorithm that is able to process sequences, and then you formalize the input of your system as a temporal sequence...

In any case, you would still have just an approximation of a solution and no explicit guarantee that this solution is viable.

Hi Stephane, 

Thanks for your comments.

Yes, i did solve this problem using brute-search and cost functions, which is typical solution in optimization problem.

But, it is unfortunate to tell that i solve it using deep learning as the inputs (i did treat the grid location as inputs) were too many, as you may expect, no solutions from the neural (too much processing time, it make my machine hang).

Anyhow, i move on. and still i hope to use DL in my works later on.

^cheers.