From Design to Deep Learning
I spent my first years out of school thinking about grids, kerning, and the way a page guides someone's eye before they've read a single word. None of it looked like a machine learning career on paper. It turned out to be exactly the training I needed.
The unexpected overlap
Design is the discipline of making structure visible. Machine learning, especially representation learning, is the discipline of discovering structure that isn't visible yet. The tools are different — Figma versus PyTorch — but the underlying question is the same one I'd been asking for years: what is the real shape of this thing, and how do I show it clearly?
When I started reading papers on embeddings and attention, I wasn't drawn in by the math first. I was drawn in by the pictures — t-SNE plots, attention heatmaps, loss landscapes. I wanted to understand what those shapes meant, and the math was the only way to get there honestly.
What actually transferred
A few habits from design carried over directly:
- Reduction as a skill. Good design cuts until only the essential remains. Good model architecture does the same — every added layer or parameter needs to earn its place.
- Comfort with iteration. Design work rarely lands on the first pass. Neither does a training run. I was already used to treating a failed attempt as information, not defeat.
- An instinct for the audience. Every interface has a person on the other end. Every model has a downstream decision. Both disciplines punish you for forgetting that.
What I had to unlearn
The biggest adjustment was patience with abstraction. Design feedback is immediate — you can see whether a layout works. Model behavior is often invisible until you've built the right diagnostic to see it. Learning to sit with that uncertainty, and to build tools that make it visible again, has been the real work of the last few years.
If you're making a similar jump, my honest advice is: don't discard the instincts you already have. They're probably more relevant than they look.