newlind() network and adapt() training function

Is it possible to add another layer of neurons in newlind()? I'm using adapt() function for training the time series problem without any delay.

If its not possible to add! What other network i can use? Feedforward is throwing error as

--Error using + Matrix dimensions must agree.

Error in nn7.grad2 (line 95) gA{i} = gA{i} + LWderivP' * gLWZ{k,i};

[Greg Heath]

Why aren't you using one of the timeseries functions???

    timedelaynet, narnet and narxnet

In addition to searching ANSWERS and the NEWSGROUP, you can check out the documentation. For example

 help narxnet

 doc narxnet

Hope this helps.

Thank you for formally accepting my answer

Greg

[Pkm]

@Greg, I'm not using delayed neural networks because

I want to train X{Input}->>T{Target} only.

If its delayed it would not happen so.

Can i still use it?

[Greg Heath]

If there is no delay, use FITNET.

[Pkm]

@Dr. Heath,

Thanks alot for the help. I'm using fitnet with adapt for training. But the network just don't get trained. :( I'm further looking into it.

Thank you again

[Greg Heath]

Why are you using adapt? The BEST approach is to FIRST try to use as many DEFAULTS as possible. After all, a group of expert developers carefully chose those values for good reasons.

Hope this helps.

Greg

[Pkm]

@Greg, as i told . I have large number of input dimension(960). So if i use train(), which is default for fitnet, it shows

"array exceeds maximum dimension".

So i won't be able to use:(

True Story!

Answer 1

[Greg Heath]

Use the fitnet defaults.

adapt is not a default.

Greg

[Pkm]

@Dr.Heath I have 1*3200 cell array for each cell array 960*1 inputs . Same applies to targets. So using defaults will throw Error as my array size exceeds the limit. I have no choice then to use adapt().

My 960*1 corresponds to 20ms .wav signal. So i can't change that.

[Greg Heath]

Yes you can:

By using dimensionality reduction. Do you think that all 960 variables are independent of the others?

The most common technique is PRINCIPAL COMPONENT ANALYSIS (PCA) which uses a smaller number of principal components that are linear combinations of the original inputs.

Another technique is PARTIAL LEAST SQUARES (PLS)which is used infrequently because it is less known ... presumably because it involves transforming both outputs and inputs.