The Layer 1 Overreach Trap

Senior Computer Vision Engineer interview at Tesla, and the interviewer asks:

“Your team is building a defect detector for high-resolution 4K manufacturing images. An engineer configures the very first convolutional layer to use massive 31x31 filters, arguing that Layer 1 needs to ‘see the whole defect at once’ to be accurate. Do you approve this PR?”

Don’t say: “Yes, if the defect is physically large on the sensor, the network needs a massive receptive field immediately to capture the global context of the anomaly.”