Why deep learning code does not work well?

Hi, I have trained, validated and tested my neural network with nprtool, using trainscg and croos-entropy.

The inputs are all in a single matrix and even the targets.

My problem occurs when I give the net more than 11264 columns as input and target (in my case I add 1024 columns every time, step by step), because the confusion matrix and the ROC curve give low performances. In fact, when I give until 10240 columns as input and target, the net has a precision of 98/99% at most but when the dimension increases, the precision drops to 91%....

I don't know how, sincerly... Con you help me?

Answer 1

[Aiswarya Subramanian]

Can you explain the structure of input matrix once again? What does "in my case I add 1024 columns every time, step by step" mean?

Also, I am understanding that by 'columns', you mean features. If then, it is possible that performance decreases by increasing the number of input features when there is high variance in your model. If your model is overfit to the training data, it’s possible you’ve used too many features and reducing the number of inputs will make the model more flexible to test or future datasets.

[wolss]

I give as input signals (in frequency domain) that are matrices 1024x512 (then I transpose them). I reunite them into a single matrix called dataset_fft that is 512x (no. of signals)*1024.

I hope I've explained...