Pinterest·Machine Learning Engineer·Technical Phone Screen
Jun 2026
Pinterest ML Engineer interview that went deep on foundational theory. The session was basically a graduate-level oral exam on bias-variance tradeoff and backprop dynamics, which I was not fully expecting from a product company.
- As you increase a model's capacity while keeping the training set fixed, what happens to bias and variance? Is a higher-capacity model generally higher-variance or higher-bias, and why? How does this affect generalization?
- In a deep fully-connected network trained with backprop, which layers suffer most from the vanishing gradient problem, and what causes it mechanically?
- What properties of the activation function and network architecture make vanishing gradients worse, and what do practitioners actually do to mitigate it?
- How would you detect in practice whether a deployed model is bias-limited versus variance-limited, and what would you do differently in each case?
- Very large over-parameterized models often generalize well despite what classical bias-variance theory would predict. How do you reconcile that with the U-shaped test error curve?
- What is the exploding gradient problem, when does it tend to dominate over vanishing gradients, and how do gradient clipping and normalization address it?
- If each layer contributes a gradient factor of roughly 0.25 (the maximum sigmoid derivative), how many layers does it take before an early-layer gradient is around one-millionth of a late-layer gradient? What does that say about trainable depth with sigmoid activations?
“I knew the textbook answer but stumbled explaining the 'why' intuitively.”