Scale AI New Grad ML Research Phone Screen: Beta Distribution

The earlier take-home was Complete a Small CNN Oriented-Box Regression Pipeline. They modified a PyTorch ResNet, stayed under a parameter limit, and used a loss that treated angles separated by pi as the same box direction.

The 45-minute phone screen started with Adjacent Order-Statistic Gap Distributions, using the gap between the fifth and sixth ordered values from ten uniform samples. The follow-ups covered possible distributions for that gap, fitting a beta distribution with maximum likelihood, optimizing its parameters, and deriving gradient descent. They also used a SciPy Kolmogorov–Smirnov test to check the fit.

They passed the phone screen. The final interviews hadn’t happened yet.

You can practice Beta Maximum Likelihood Fit on FastPrep.