Phone Screen
Part one was ML knowledge. The interviewer basically didn't ask about basic definitions — the whole thing revolved around digging into my project. From model selection to training details to evaluation, every answer just led to another follow-up. Typical questions included: how did you do evaluation for your model? Which metrics did you use? Why did you choose them? What are the mathematical definitions of these metrics? Is there a better evaluation pipeline? If you added human labels, how would that improve things? If there were no labels at all, in the simplest case how would you judge the similarity between two answers? He also asked about the difference between transformer and RNN, what the pros and cons of each are when the input is very long, why attention can capture long-range dependencies, and why RNN can't.
Another question was how to judge whether two sets of images come from the same distribution.
Part two was ML coding. I hand-wrote K-means on the spot — it didn't need to be fully runnable, just clearly walk through the steps and get the matrix dimensions right. The interviewer said he doesn't test LeetCode, but other interviewers might.
Onsite:
Round 1 was Statistics Application. You're given two groups of user data, one North American users and one European users, and each user has multi-dimensional continuous features (like session duration, click-through rate, purchase conversion rate, etc.). The question was how to determine whether the two groups have a significant difference in their overall distribution.
Round 2 was an ML deep dive. This round again revolved around digging into a project I'd worked on. The interviewer started with data preprocessing — why I chose that particular data cleaning approach, why I used that feature engineering approach.
Then he kept digging into evaluation, asking why I chose that set of metrics, what each of them reflected, how I split the validation set, and whether I'd considered stratified sampling.
After that he asked how I would choose if I had to switch to a different model, and had me analyze the pros and cons of different models in terms of training time, inference latency, interpretability, and robustness to noise.
Finally the interviewer asked from an optimization angle — what the pros and cons of my optimizer choice were, and what would change if I switched to Adam or RMSProp. He also pushed on why I didn't choose AdamW, and in what scenarios it would be more suitable.
Round 3: coding. 1. Merge Intervals 2. Top K Frequent Elements
Round 4: chat with the hiring manager, ML trivia
Round 5: behavioral
Discussion
Loading comments…