Daniel BourkeTechnical Staff, Artificial Analysis · teaching AI/ML at zerotomastery.io, building nutrify.app

How to code neural networks without math*

Name: How to code neural networks without math*
Uploaded: 2021-10-17T07:22:24+00:00
Duration: 43 s
Channel: Daniel Bourke
Description: In this short monologue, the speaker explains his personal way of thinking about neural networks: as simple stacks of layers. He illustrates the idea…

17 Oct 2021 · 1 min

Watch on YouTube

0:43

In this short monologue, the speaker explains his personal way of thinking about neural networks: as simple stacks of layers. He illustrates the idea with an LSTM and then says transformers are essentially the same, describing a transformer as a large dictionary that creates embeddings which are then multiplied by matrices. While he acknowledges that understanding the underlying mathematics—sigmoid functions, matrix multiplication, linear layers—is fundamentally important, he prefers to learn and build deep‑learning models by writing code and stacking components rather than deriving formulas. The talk emphasizes a pragmatic, code‑first approach to deep learning, using concrete analogies to demystify complex architectures.

Chapters

look at this. This is how I think of neural networks. It's basically just a stack, especially like this is an LSTM, but transformers are the same. Transformers are like literally like a, a big dictionary. They create an embedding and you do matrix multiplication on that embedding and there's your transformer model. But this is how I think of coding neural networks. I never think about the math behind it. I just stack the things on top. Of course, I've looked into the formulas of a sigmoid and matrix multiplication, all that stuff, and linear layers. But I write neural networks like this. So yes, math is fundamentally important to it all, but for me, I prefer to write code and learn about deep learning like this.