Parameter dimensions in Practice Lab

[flawn]

Hey everyone,

I don’t understand why the dimensions of e.g. the matrix are flipped in this lab. For example: Usually, if I have 400 input features, and 25 hidden units in the first hidden layer, my weight matrix has the dimensions of 25 x 400 (25 rows, and 400 columns - needed to matrix multiply this matrix with the input features properly)

Can anybody explain me why this is how it is?

[Alireza_Saei]

Hi @flawn

I remember I was confused about this, too. The answer is: This depends on how you implement you neural network logic and both approaches are correct.

[TMosh]

The orientation of the training data and the weight matrices is not rigorously defined. You will find many different design choices.