 One way of understanding how multi-layer perception works is trying to visualize what you have there. One approach to gain some insight is to plot the weights of the first layer. After all, what is the first layer? Well, it goes from an image to a first set of neurons, which means that the weight vectors actually lie in image space, which means we can image them and look at them and see if they give us insights into what they mean. Look at the weights. What do they reveal about how this specific MLP works?