Since the kernel size of the first and second conv net is 3x3, the feature map's demension should be 6@30x30 and 16@13x13 C1: feature maps 6@28x28 -> C1: feature maps 6@30x30 C2: feature maps 6@10x10 -> C2: feature maps 6@13x13 or just simply use conv net with 5x5 kernel size :)