How does convolution work with multiple filters?

Question

Neural networks

That is, in the standard convolutional network for MNIST, two layers of convolution are used. In the first it is 32 two-dimensional and in the second there are already 64 of the same size. Then, at the first convolution from a two-dimensional array, we get a three-dimensional array of features. But how do the filters of the second convolution layer work, because our array has a depth of 32, and the filter array is 64?

Answer 1

Convolutional filters are 3D tensors, not 2D ones (if the channel is only 1, then the tensor is kernel_size x kernel_size x 1).
Their number determines the number of channels at the output of the convolutional layer.