This one was particularly intriguing because it isn’t something that most organisations inquire about. This query came from a company that specialises in computer vision, as you might anticipate. There are two components to this solution. Convolutions, for starters, conserve, encode, and exploit the spatial information from the image. We wouldn’t have any relative spatial information if we solely employed FC layers. Second, because each convolution kernel operates as its own filter/feature detector, Convolutional Neural Networks (CNNs) contain some built-in translation invariance.