Why does my neural network perform better with more input neurons than features/variables ? Now if I use the exact input neurons to features/variables the neural network performs much worse. For example, I have 6 dimensions of data that are 200 in length (or 200 samples). Within that data there are groups of 6 data sets - is this why my NN with 36 neuron inputs performs better than 6 neuron inputs? The dimensions should suggest I just need a NN with 6 input neurons. The hidden layers for the 6 inputs are 12 and 6 and for the 36 inputs, are 72 and 36 respectively.
It's been a long time since I last used NNs and so many thanks for answers or pointers on this?
Don't worry about this question unlike my PhD and Post Doc years I am dancing in and out of ML and just realised the following code:
%net = newff(minmax(P),[S1 S2 S3 S4],{'tansig' 'tansig' 'tansig' 'purelin'},'traingdx','learnk','mse'); ......... has the input layer set to minmax(P) and not S1 ........basically I have a very bloated DNN ! I was wondering why cIouldn't find anything on this subject ! I am not going mad !
Hello Dr. James Marcus Griffin
Thanks for the question and the clarification answer.
Best regards
Hi Dr.
More neurons in the same layer mean more teachers are extracted in each layer. But that comes with more computation complexity and maybe overfitting so you must balance high accuracy and overfitting. My suggestion is that you must think in two ways one to increase the data and the second to update the structure of the model
- Badges
- Science topic