Why are Transformers replacing CNNs?

Why does a Transformer classify this cat as a cat… while a ResNet calls it a macaw? In this video we break down one of the biggest shifts in computer vision: why Transformers replaced Convolutional Neural Networks (CNNs) — even though CNNs were designed for images and Transformers for language. We’ll compare convolution vs self-attention, explore CNNs’ inductive biases (locality, translation invariance, hierarchical features), and see why self-attention is strictly more expressive than convolution. You’ll also learn how attention can exactly implement convolutional kernels using relative pos…