upstart | Machine Learning Engineer | Interview Experience
Interview Date: Not specified
Result: Not specified
Difficulty: Not specified
Interview Process
The application was submitted in early November, and the HR reached out during the Thanksgiving holiday. A simple screening call was conducted with HR, which included three questions:
- Properties of the mean and variance of a standard normal distribution.
- Distribution of the expected number of coins falling into a cup when flipping n coins.
- Loss function of logistic regression.
Following the screening, two technical interviews were scheduled about a week later.
The first technical interview focused on probability and probability simulation. The candidate was asked to theoretically calculate the probability that particles would still exist after 10 days, given that there were 100 particles with a half-life of one day. Then, the candidate had to write code in CoderPad to simulate this process. A follow-up question asked how to determine the decay rate if it was uncertain, given that n particles survived on the m-th day. The candidate was prompted to use bootstrap sampling to obtain the distribution of n and calculate the confidence interval.
The second technical interview was a machine learning case study. The candidate was provided with a notebook and a classic California house price dataset, where the training and testing sets had data shifting due to different income brackets. The task was to build a model from scratch to predict house prices, analyze the results, and optimize the model. The interviewer’s focus was on how to construct a system to balance variance and bias, given the data shifting. The candidate suggested creating a pipeline using a high variance model for interpolation and a high bias model for extrapolation, followed by an evaluation pipeline to test predictions across income brackets. However, the candidate spent too much time on feature engineering and ultimately ran out of time for model analysis and optimization.
Technical Questions
- Normal Distribution Properties
- Coin Flipping Distribution
- Logistic Regression Loss Function
Tips & Insights
The interview experience felt somewhat unconventional, emphasizing practical skills over theoretical knowledge. The HR was efficient, sending a rejection email just two hours after the interview. The candidate hoped that sharing this experience would be helpful to others.