Convolution and parameter sharing are why a CNN needs a tiny fraction of the weights a dense net would, and why it generalizes across position. The conceptual answer, the feature-hierarchy picture, and the honest take on when a fine-tuned CNN still beats reaching for a giant vision-language model.
Updated Aug 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.
Convolution and parameter sharing are why a CNN needs a tiny fraction of the weights a dense net would, and why it generalizes across position. The conceptual answer, the feature-hierarchy picture, and the honest take on when a fine-tuned CNN still beats reaching for a giant vision-language model.
The screen is whether the candidate explains parameter sharing as the core efficiency and inductive-bias win, not just 'filters slide over the image.' The reserved follow-up for an FDE is the pragmatic one: 'a customer wants to classify 5000 product photos, what do you actually ship': the strong answer is fine-tune a pretrained CNN/ViT, not train from scratch and not necessarily a multimodal LLM, and knowing why is the applied signal.
No comments yet — be the first to share your approach.
