Sara Beery teaches this lecture from MIT's 6.7960 Deep Learning course, focused on convolutional neural networks as the natural architecture for data arranged on a grid, such as images. She builds up the case for convolution from first principles, covering translation equivariance, local connectivity, weight sharing, and pooling, and explains why these properties make CNNs efficient compared to fully connected networks when inputs have spatial structure. The lecture works through the mechanics of convolutional layers, receptive fields, and common architectural patterns used in vision models. Running 84 minutes, it is a core session in MIT's graduate deep learning sequence, assuming familiarity with prior lectures on neural network fundamentals and building toward later material on other architecture families.