In this video we are solving a popular deep learning interview question asked at companies like Google, Anthropic, Amazon, Apple and OpenAI: Implement a 2D Convolutional Layer from scratch using PyTorch.
This is one of those problems where most candidates know what convolution does at a high level, like slide a kernel across an image and sum the element-wise products but when asked to actually implement it in real interview, things get tricky fast.
We'll walk through exactly how to approach this in a real interview — understanding how padding adds zeros around the input, how stride controls the step size, deriving the output dimension formula, writing clean code, handling edge cases like non-square matrices and stride values that don't divide evenly, and tracing through the full example step by step. Stay tuned to see how to nail this classic question!
If you enjoyed the video and got some value from it, then consider dropping a subscription to my channel and helping me reach my goal of 1000 subscribers!
