Graduate student at a non-top university.
I applied in November and the recruiter reached out to me immediately the day after.
I initially applied for an NLP role but they were only hiring for Computer Vision intern roles.

1st Round: Phone chat with recruiter.
2nd Round: Take home assessment.
I was given the Computer Vision assessment.

• The challenge is to build a CNN model with less than 2mil trainable parameters that regresses the width, height, center coordinates and rotation angle of one object in the image.
• The task was to achieve the highest IoU score you can (open ended)
  • Intersection of Union (essentially how much the predicted bounding box and the actual bounding box overlap)
• I did several experiments with several architectures and custom loss functions and compiled my results into a pdf document and sent it over.

I passed the take home assessment and was scheduled for a 45 minutes tech phone screen.

3rd Round: Phone screen

• The first question was to use numpy to create a distribution of differences of uniform values. For example, create a random sample of 10 numbers between 0 and 1, then take the difference of the 4th and 5th numbers, save the difference, repeat N times. Then, visualize this list (distribution)

• Repeat the above step but for the 5th and 6th numbers and visualize the new distribution.

• Compare the two distributions visually, are they the same? What would you change to help you make the decision of whether they are the same or not (visually)?

  • Increase N.

• I was also asked how I'd compare the two distributions numerically (are there tests we can use to check if the distributions are the same)

  • mean and standard deviation of both lists reach equality (i.e. mean1 = mean2 and std1 = std2) the larger N samples we take.

• Note the two distributions turned out to be right-tailed so this kind of threw me off a bit because I was somehow sure they were normal.

  • Also note that the reason the distributions turned out to be right-tailed is because I sorted the sample list before taking the difference. I don't remember if I was asked to get the difference of the 4th and 5th largest numbers or just the 4th and 5th numbers.

• We spent a good 20 minutes on this and then moved on to a deep dive on prior ML internship experience. I wasn't able to articulate any depth in my understanding of the tech i used duirng my ML internship.

The 4th round would have been the on-site (3 hours remote interview) but I didn't make it this far so not sure what they ask about.

Hope this helps someone better prepare for the intenrship.