Explain dataset size, generalization, and U-Net skips

You are interviewing for an ML Engineer role in an image/video team. Answer the following conceptual questions clearly and concisely.

1) Small vs. large datasets

• What are the typical advantages/disadvantages of training on a small dataset vs a large dataset?
• How do compute, labeling cost, noise, and distribution shift change your strategy?

2) Overfitting vs. underfitting

• Define overfitting and underfitting .
• Give practical signs you would observe in training/validation curves.
• Provide common fixes for each.

3) How do you approach a new modeling task?

Describe your end-to-end approach when developing a model for a new task:

• Do you start with a small model and small dataset? Why or why not?
• What baselines do you set?
• How do you iterate (data, model, loss, evaluation)?

4) Super-resolution losses and blurriness

In single-image super-resolution:

• Compare optimizing pixel-wise MSE (L2) loss vs a perceptual loss (e.g., feature-space loss using a pretrained network).
• Which objective more commonly produces blurry images, and why?

5) U-Net architecture

• Explain the U-Net architecture at a high level (encoder/decoder, downsampling/upsampling, skip connections).
• What is the purpose of the skip connections ?
• What typically happens if you remove skip connections in a U-Net used for dense prediction (segmentation), in terms of optimization and output quality?