How to take matrices as input to neural network?

I have 5000 matrices of size 20x20. How to take these as input to my neural network? Is making a column vector of size 400x1 the only way? Is there some way I can have 20 input features of size 1x20?

[KSSV]

It depends on your target... what is your target?

[Pritam Ghoshal]

the target is a 20x1 vector

[KSSV]

5000*20*1 right?

[Pritam Ghoshal]

yes

[KSSV]

Then you can associate one output to each 1x20 feature right?

[Pritam Ghoshal]

For each 20x20 matrix my output will be a 20x1 vector. I have 5000 such matrices.

[KSSV]

Yes...you can try for each feature of matrix 1x20 on output ....so each row corresponds to one number of target.

Answer 1

[Gagan Agarwal]

Hi Pritam

Yes it is possible and you can try for each feature of matrix to 1x20 on output and that too through various netwroks depending on the nature of your data and what you're trying to achieve.

For instance if:

• spatial relationships are crucial, consider CNNs.
• sequential relationships matter, look at RNNs or their variants.
• you're interested in treating each row/column as a separate entity without focusing on spatial or sequential relationships, the multiple input models or separate input channels approach could work.