Why Using array 3*3 in convolutional?
There is not much context to your question, but I assume you mean the filter size. Prof Ng is just using that as an example in the parts where he goes through the process of stepping through and showing how convolutional filters work. Note that he’s simplifying a bit here by only assuming one “channel” in the input. Almost always there are more.
3 x 3 x n_c is one of the common choices for filter sizes. As Prof Ng explains in the lectures, it is more common to use odd filter sizes, so it’s usually 3 x 3 or 5 x 5 or 7 x 7. The reason is that the dimensions work out better when the filter dimensions are congruent to 1 mod 2.
If I have missed the point of your question, please follow up with a more specific question.