Maximum number of input neurons to the neural network

Hi,

I have EEG data to classify, for 2 classes. The data dimesnion is 30000x512. Where 512 is the number of samples (50:50 for each class). I tried feeding the raw EEG data to neural network and tuned it's parameter with 3 hidden laeyrs [10 10 10], and it works really well in terms of the classification accuracy.

I was wondering if this is acceptable to do.? Or it's necessary that i have to feed reduced number of input neurons in the form of extracted features.?

I will highly appreciate if anyone can guide me through.

[Tomer Nahshon]

Hello Naina,

I doubt what you are doing is beneficial in the sense of handling the data.

You haven't split your data into train/test set or train/validation/test set.

In order to estimate the performence of your algorithm on data that it wasn't trained on you need to split your data.

Since you have only 512 samples (small dataset in the sense of deep learning or almost any other machine learning algorithm) I would go on cross validation

https://www.mathworks.com/help/bioinfo/ref/crossvalind.html

After you train your model, try testing the predictions using the test data and see your result.

I have a feeling you are overfitting to your training set.

[Joana]

Hi Tomer,

Thanks for the detailed reply.

I skipped the details of data division. I'm using 10-fold cross validation actually. I tried 70:10:20 for training:validation:testing as well.

The results for 10-cross validation are slightly better than the manual data division into 70:30. And i also tested the trained model on completely new dataset and still it gives the accuracy>65%. so i assume the model is learning pretty well.?

Mainly, I was just curious that if it's right to train with input neurons equal to 30000.? and then feed it to [10 10 10] hidden layers, with 1 output neuron.

Answer 1

[Mahesh Taparia]

Hi Naina,

You are having the dataset which consists of 512 samples with 30000 dimensions. You have less data.

But the way you are training is not optimum. You can use PCA (Principal Component Analysis) to reduce the number of dimensions which will reduce model complexity. For more details about PCA, you can refer to this link.