Acing the Machine Learning Interview

A whiteboard during a machine learning interview.In my decade of managing applied machine learning teams I’ve interviewed maybe a hundred people. Over that time, I’ve come to rely on two main questions. I’m going to tell you what they are.
First, a bit of philosophy. There are lots of things we could talk about in an interview:
  • What do you like?
  • What did you do in your last project?
  • Can you tell a good story about yourself?
  • Have you read lots of papers about machine learning?
  • Can you program?
  • Do you know statistics?
All of that is great, and of course candidates must know those things to get a job, but what I also want to know is: what can you do when you have a blank screen in front of you and an open-ended machine learning task to complete?
That isn’t easy to figure out in an interview, but I try. The approach I take is to talk through an end-to-end problem. For example:
Let’s walk through an example of intelligence creation: a blink detector. Maybe your application is authenticating users by recognizing their irises, so you need wait till their eyes are open to identify them. Or maybe you are building a new dating app where users wink at the profiles of the users they’d like to meet. How would you build it?
There are so many interesting things to discuss, so many ways to approach this question, and I still learn from the conversations I have. A good answer has discussion on the following topics:
  • Understanding the Environment
  • Defining Success
  • Getting Data
  • Getting Ready to Evaluate
  • Simple Features and Heuristics
  • Machine Learning
  • Understanding the Tradeoffs
  • Assessing and Iterating

Understanding the environment

The first step in every applied intelligence-creation project is to understand what you are trying to do. Detect a blink, right? I mean, what part of “detect a blink” is confusing? Well, nothing. But there are some additional things you’ll need to know to succeed. Candidates might ask things like:
  • What kind of sensor will the eye images come from? Will the image source be standardized or will different users have different cameras?
  • What form will the input take? A single image? A short video clip? An ongoing live feed of video?
  • Where will the product be used? On desktop computers? Laptops? Indoors? Outdoors?
  • How will the system use the blink output? Should the output of the intelligence be a classification (that is, a flag that is true if the eye is closed and false if it is opened)? Should the output be a probability (1.0 if the eye is closed, and 0.0 if the eye is opened)? Or should the output be something else?
  • What type of resources can the blink detector use? How much RAM and CPU are available for the model? What are the latency requirements?
That’s a lot of questions before even getting started, and the answers are important to making good decisions about how to proceed.

Defining Success

To succeed, the blink detector will need to be accurate. But how accurate? This depends on what it will be used for. I want to know if a candidate can consider the experience that their model will drive and discuss how various levels of accuracy will change the way users perceive the overall system.
Questions include:
  • How many mistakes will a user see per day?
  • How many successful interactions will they have per unsuccessful interaction?
  • What will the mistakes cost the user?
I look for a discussion of options for how accuracy and experience will interact, how users will perceive the mistakes, and how will they be able to work around them.

Getting Data

Data is critical to creating intelligence. If you want to do machine learning right out of the gate, you’ll need lots of training data. I hope a candidate can discuss two distinct ways to think about getting data:
Getting data to bootstrap the intelligence:
  • Search the web and download images of people’s faces that are a good match for the sensor the blink- detector will be using (resolution, distance to the eye, and so on). Then pay people to separate the images into ones where the eye is opened and ones where it is closed.
  • Take a camera (that is a good match to the one the system will need to run on) to a few hundred people, have them look into the camera and close and open their eyes according to some script that gets you the data you need.
  • Something else?
How to get data from users as they use the system:
A well-functioning Intelligent System will produce its own training data as users use it. But this isn’t always easy to get right. In the blink-detector case some options include:
  • Tie data collection to the performance task: For example, in the iris-login system, when the user successfully logs in with the iris system, that is an example of a frame that works well for iris login. When the user is unable to log in with their iris (and has to type their password instead), that is a good example of a frame that should be weeded out by the intelligence.
  • Creating a data collection experience: For example, maybe a setup experience that has users open and close their eyes so the system can calibrate (and capture training data in the process). Or maybe there is a tutorial in the game that makes users open and close their eyes at specific times and verify their eyes are in the right state with a mouse-click (and capture training data).

Getting Ready to Evaluate

A candidate should have a very good understanding of evaluating models, including:
1. Setting aside data for evaluation:
Make sure there is enough set aside, and the data you set aside is reasonably independent of the data you’ll use to create the intelligence. In the blink-detector case you might like to partition by user (all the images from the same person are either used to create intelligence or to evaluate it), and you might like to create sub-population evaluation sets for: users with glasses, ethnicity, gender, and age.
2. Creating a framework to run the evaluation:
That is, a framework to take an “intelligence” and executes it on the test data exactly as it will be executed at runtime. Exactly. The. Same.
3. Generating reports on intelligence quality that can be used to know:
  • How accurate the intelligence is.
  • If it is making the right types of mistakes or the wrong ones.
  • If there is any sub-population where the accuracy is significantly worse.
  • Some of the worst mistakes it is making.

Simple Features and Heuristics

I like to have some discussion about simple heuristics that can solve the problem, because:
  1. Making some heuristics can help you make sure the problem is actually hard (if your heuristic intelligence solves the problem you can stop right away, saving time and money).
  2. It can create a baseline to compare with more advanced techniques—if your intelligence is complex, expensive, and barely improves over a simple heuristic, you might not be on the right track.
In the case of blink-detection you might try:
  • Measuring gradients in the image in horizontal and vertical directions, because the shape of the eye changes when eyes are opened and closed.
  • Measuring the color of the pixels and comparing them to common “eye” and “skin” colors, because if you see a lot of “eye” color the eye is probably open, and if you see a lot of “skin color” the eye probably closed.
Then you might set thresholds on these measurements and make a simple combination of these detectors, like letting each of them vote “open” or “closed” and going with the majority decision.
If a candidate has computer vision experience their heuristics will be more sophisticated. If they don’t have computer vision experience their heuristics might be as bad as mine. It doesn’t matter as long as they come up with some reasonable ideas and have a good discussion about them.

Machine Learning

I look for candidates who can articulate a simple “standard” approach for the type of problem we’re discussing. And I am aware that standards change. It doesn’t matter what machine learning technique the candidate suggests, as long as they can defend their decisions and exchange ideas about the pros and cons.
And here is where I bring in the second question. I let the candidate pick their favorite machine learning algorithm and then ask them to teach me something about it.
This can mean different things for different people. They might go to the board and explain the math about how to train the model. Maybe they explain the model representation and how inference works. They could discuss what types of feature engineering works well with the approach. Maybe they explain what types of problems the approach works well on — and which it works poorly on. Or maybe they explain the parameters the training algorithm has and what the parameters do and how they know which to change based on the results of a training run.
What’s important is that they understand the tool and make me believe they can use it effectively in practice.

Understanding the Tradeoffs

I want a candidate to be able to discuss some of the realities of shipping a model to customers. This is a process of exploring constraints and trade-offs. Discussing questions like these:
  • How does the intelligence quality scale with computation in the run-time?
  • How many times will we need to plan to update the intelligence per week?
  • What is the end-to-end latency of executing the intelligence on a specific hardware setup?
  • What are the categories of worst customer-impacting mistakes the intelligence will probably make?
The answers to these questions will help decide where the intelligence should live, what support systems to build, how to tune the experiences, and more. The candidate should be able to talk about these.

Assess and Iterate

And of course, machine learning is iterative. The candidate must be able to talk about the process of iterating, saying things like:
  • You could look at lots of false positives and false negatives.
  • You could try more or different data.
  • You could try more sophisticated features.
  • You could try more complex machine learning.
  • You could try to change people’s minds about the viability of the system’s objectives.
  • You could try influencing the experience to work better with the types of mistakes you are making.
  • And then you iterate and iterate and iterate.
A junior candidate might start in the middle of this list and might only be able to talk about one or two of these topics. A senior candidate should have a good sense of all of them and be able to discuss options as I probe and add constraints. There is no right answer — good discussion is key.
And if you really want to learn how to ace the machine learning interview, you can check out the book or the audio book, which you can get for free if you start a trial account with Audible.