Machine Learning Engineer · Phone Screen · Multiple Types

Apple — Machine Learning Engineer ❌ Failed

Level: Senior-Level

Round: Phone Screen · Type: Multiple Types · Difficulty: 6/10 · Duration: 60 min · Interviewer: Friendly

Topics: Machine Learning, PyTorch, U-Net, Overfitting, Underfitting, Super Resolution

Location: San Francisco Bay Area

Interview date: 2025-12-15

Got offer: False

Summary

This was a phone screen for a Machine Learning Engineer position. The interviewer asked ML-related questions and a debugging question related to U-Net in PyTorch.

Details

In my phone screen, the interviewer, asked me the following questions:

• Small dataset vs large dataset
• Overfitting vs underfitting and their solutions
• When developing a model for a task, how do you approach it? Do you start with a small model and a small dataset?
• Super resolution with MSE loss and perceptual loss. Which causes blurry images?
• U-Net explanation
• What happens if we don’t have that skip connection in UNets?

The coding question was to debug a U-Net implementation. The required changes were:

1. Input tensor shape
2. A boolean variable in the decode block
3. Kernel size of the final block
4. num_classes of the U-Net

I needed to modify a few numbers to fix the issues. To prepare for this question, I should have reviewed the U-Net paper, especially the architecture section. The input size is 572, not the standard 512 or 256. The final output shape is different from the input, resulting in 2x388x388. I made a mistake by assuming the input and output shapes were the same for most image and video applications.