If you’re into technology and are looking for a job that would involve data science, you will most likely have heard about machine learning. The term has an air of mystery surrounding it - many people are baffled by the concept itself. However, if you’re looking for how to become an AI engineer or a business intelligence developer, you are probably quite familiar with machine learning and everything that surrounds it. If you want to score that job, however, you’ll have to prepare for a job interview. And what better way to prepare than revising machine learning interview questions?
In this tutorial, we’ll look over some of the most popular interview questions on machine learning. We’ll cover both the basic and the advanced stuff, so grab your thinking caps and let’s jump right on in.
Table of Contents
- 1. The Main Aspects of Machine Learning
- 1.1. Question 1: Describe ‘machine learning’.
- 1.2. Question 2: What is ‘deep learning’?
- 1.3. Question 3: What’s the difference between ‘type 1’ and ‘type 2’ errors?
- 1.4. Question 4: What is ‘data augmentation’?
- 1.5. Question 5: Why are is ‘naive Bayes’ called that?
- 1.6. Question 6: Which ones are better - deep networks or shallow ones?
- 1.7. Question 7: What’s ‘Fourier transform’?
- 1.8. Question 8: What’s a ‘convolutional network’?
- 1.9. Question 9: What should we know about the correlation between ‘True Positive Rate’ and ‘Recall’?
- 1.10. Question 10: What’s a ‘backpropagation’?
- 1.11. Question 11: What happens if we only use a ‘validation set’, without applying a ‘test set’?
- 1.12. Question 12: What is the difference between deductive and inductive machine learning?
- 1.13. Question 13: How do variance and bias play out in machine learning?
- 1.14. Question 14: What is supervised learning and how it differs from unsupervised?
- 1.15. Question 15: How do you choose an algorithm for a classification problem?
- 2. Advanced Machine Learning Interview Questions
- 2.1. Question 1: What’s the difference between the ‘generative’ and ‘discriminative’ models?
- 2.2. Question 2: Explain the differences between ‘cross-validation’ and ‘stratified cross-validation’.
- 2.3. Question 3: In what situations should you use a ‘Lasso’ and a ‘Ridge’ regressions?
- 2.4. Question 4: What’s ‘F1’?
- 2.5. Question 5: In most cases, which one of the two has a higher score - ensembles or individual models?
- 2.6. Question 6: What’s the difference between ‘correlation’ and ‘covariance’?
- 2.7. Question 7: Describe an ‘imbalanced dataset’.
- 2.8. Question 8: What is ‘data normalization’?
- 2.9. Question 9: Could you capture the correlation between categorical and continuous variables?
- 2.10. Question 10: What is the activation function used for?
- 3. Conclusions
The Main Aspects of Machine Learning
The best way to go about it is to start from the very basic machine learning engineer interview questions. These are the ones that you can expect to receive at the beginning of your interview. This way, employers want to see if you’re able of critical thinking and can form your own, cohesive thoughts. That’s why a lot of these questions will be based on definitions, comparisons, explanations and so on.
Latest Udacity Coupon Found:
EXCLUSIVE: 75% OFF
Udacity Black Friday Deal
Begin your spring with new skills - hurry up & choose your dream Udacity course with a huge 65% discount! This exclusive Udacity coupon code is active for a limited time.
Question 1: Describe ‘machine learning’.
The vast majority of your employers are probably going to ask you something similar to this as a first question. This is done for a couple of reasons.
First of all, your interviewers can’t proceed with other general machine learning interview questions before they see if you have an idea of what ‘machine learning’ even is in the first place. Furthermore, the way that you answer will show just how well you can think of your definitions - or, in other words, how well can you explain a difficult topic in an easily understandable way. If you just spit out a twenty-liner that you spent the whole night memorizing from some random science journal, it will probably give you less credibility than if you were to think of a way to explain it yourself.
So… What is machine learning?
Probably the easiest and most understandable way to describe machine learning is to call it a specific philosophy of AI development. It is a field of science that is concerned with how to make machines that would learn from the information provided to them, without being programmed to do so beforehand.
Question 2: What is ‘deep learning’?
Since deep learning is so closely intertwined with machine learning, you might even get cross deep and machine learning interview questions.
Deep learning is a branch of machine learning. This branch of science is concerned with making the machine’s neural networks resemble a human brain as closely as possible.
Question 3: What’s the difference between ‘type 1’ and ‘type 2’ errors?
Type 1 errors claim that something has happened when, in actuality, it was impossible for it to happen. Type 2 errors do the opposite - claim that nothing happened when it did.
For example, here’s a good method to help you remember the difference between the two types of errors: just imagine that a type 1 error is when you tell your dog that he is a cat, while a type two error is when you tell the same dog that dogs can’t bark.
Question 4: What is ‘data augmentation’?
One of the simpler machine learning interview questions, data augmentation is a way of modifying and creating new data out of the old one. The way that this is done is by leaving the target as it is or simply changing it into something that is already known.
Question 5: Why are is ‘naive Bayes’ called that?
Naive Bayes is called naive because of the way that it thinks. It assumes that every element in a data set is the same when it comes to their importance. Needless to say, that is seldom the case in a day-to-day scenario.
Question 6: Which ones are better - deep networks or shallow ones?
You could classify this as one of the comparison machine learning interview questions, for you have to know quite a bit about both of the networks and also be able to compare them to find the clear difference.
Deep networks are usually considered to be a better alternative. This is simply because they are comprised of more layers, most of which are hidden - this helps deep networks extract and build better features.
Question 7: What’s ‘Fourier transform’?
The ‘Fourier transform’ method is used to transform simple, generic functions into what are known as superfunctions. If this is one of the machine learning interview questions that you would like to expand a bit more on, you could compare it with a situation where you are given a car to take it apart and see all of the different components and parts that it’s made out of.
Question 8: What’s a ‘convolutional network’?
Normal, simple networks use connected layers to perform their processes. In turn, convolutional networks are those that, instead of using connected layers, use convolutional ones.
The main reason why people prefer to use convolutional networks over the standard, connected-layer ones is that convolutional networks have a much smaller amount of parameters attributed to them.
Question 9: What should we know about the correlation between ‘True Positive Rate’ and ‘Recall’?
Although this sounds like one of the more advanced machine learning interview questions, the answer is pretty simple. Both of these metrics are identically equal. We can see this by looking at their formula: TP/TP + FN.
Question 10: What’s a ‘backpropagation’?
A fancy-sounding term in of its own, backpropagation is simply a method of training for multilayered neural networks. We would train the network with this method by taking the ‘error’ from the very end of it and placing it inside every single weight within the network. This way, the machine has the opportunity to apply its computation effectively.
Question 11: What happens if we only use a ‘validation set’, without applying a ‘test set’?
In machine learning interview questions, this one might come off as a tougher one.
If you only applied a validation set, it would not provide an accurate estimate of all of the measurements of the model that you’re trying to test. This is because the ‘test set’ is used to test how the model would perform on examples that it hadn’t encountered up to that point in time. Thus, if you remove the test set, you automatically undermine the possibly valid test results, so to speak.
Question 12: What is the difference between deductive and inductive machine learning?
The main difference is how they begin. Inductive machine learning begins with examples from which to conclude. Deductive machine learning begins with conclusions, then learns by deducing what wrong or what is right about that conclusion.
Question 13: How do variance and bias play out in machine learning?
They are both errors. Variance is an error that is the result of too much complexity in the machine learning algorithm. Bias is an error that is due to flawed assumptions in the learning algorithm. Do not mix these up since you will need to remember them in following machine learning interview questions.
Question 14: What is supervised learning and how it differs from unsupervised?
The supervised machine learning is a process in that outputs are fed back into a computer for the software to learn from and get more accurate results the next time. Unsupervised machine learning means that a computer will learn without initial training which is an alternative to supervised machine learning in which ‘machine’ receives initial training to start.
Question 15: How do you choose an algorithm for a classification problem?
In this case, the answer depends on the degree of accuracy needed and the size of the training set. If the training set is small, it is recommended to choose a low variance/high bias classifier. If the situation is the opposite, the training set is large, then you should go for a high variance and low bias classifier.
Advanced Machine Learning Interview Questions
Now that you have some sort of idea of what machine learning is in general and what type of basic machine and deep learning interview questions you can expect during the job interview, we can move on to the more advanced stuff.
Don’t be fooled, however. Your employers most likely won’t ask you to build a self-sufficient AI system or write a three hundred page-long book about all of the different ways you can study deep learning. In this context, “advanced” simply means that the questions are going to be a tad bit tougher - you might be asked to provide further explanation for your answers, give examples, etc. So don’t worry, relax and let’s jump straight into it.
Question 1: What’s the difference between the ‘generative’ and ‘discriminative’ models?
Although it might sound like one of the trick machine learning interview questions, your employers most likely just want to know how these models deal with data.
A generative model, as the name implies, is going to put in the effort and learn the different categories of data that it’s provided. As opposed to that, a discriminative model will just study the differences between the various data categories.
Developers and engineers usually prefer to use the discriminative model, for it tends to handle its tasks faster and more efficiently.
Question 2: Explain the differences between ‘cross-validation’ and ‘stratified cross-validation’.
The simple cross-validation is used to randomly separate data between the period of training and the validation set. Stratified cross-validation does the same thing, but without the random variable - it does track and preserve the ration of training vs validation testing. This is one of those machine learning interview questions that might be easy to mix up, so watch out for that!
Question 3: In what situations should you use a ‘Lasso’ and a ‘Ridge’ regressions?
This falls into the category of advanced machine learning engineer interview questions mainly because you do need some in-depth knowledge concerning both types of regressions to provide a valid answer.
The Lasso regression can perform both functions of selecting variables and shrinking parameters, while a Ridge regression can only be used for the latter. On that thought, you would most likely use Lasso regression when you have just a few variables and a big effect. In turn, a Ridge regression should be used when there are plenty of small variables.
This is a good example of those machine learning interview questions that you could expand upon with your answer, not just give a generic one-liner.
Question 4: What’s ‘F1’?
No, it’s not a key on your keyboard that you can just press for the answer.
The F1 score is a measurement of just how well is your model doing. Anything close to the ‘1’ mark is great, anything below the ‘0,5’ mark should be worked upon.
Question 5: In most cases, which one of the two has a higher score - ensembles or individual models?
Ensembles are usually ones that provide a bigger score. It is because they are simply combinations of various models, made to predict a one, particular outcome. The more models there are, the more errors they can sort through - the better the end prediction score will be.
Question 6: What’s the difference between ‘correlation’ and ‘covariance’?
This will be one of the advanced machine learning interview questions only if you don’t know how these two correlate (no pun intended).
If you do know, however, then the answer is pretty simple: covariance becomes a correlation once it is standardized.
Question 7: Describe an ‘imbalanced dataset’.
An imbalanced dataset is a set that, after testing, brings back the results that more than half of the entire information is stationed in just one class.
How can you avoid this? Well, there are a couple of simple solutions - either perform the test again using a different algorithm or try to test an even larger amount of data so that the results would even out.
Question 8: What is ‘data normalization’?
Remember when we talked about ‘backpropagation’ in the previous machine learning interview questions? Well, data normalization is used to minimize the redundancy of data within the process of backpropagation. It allows the user to rescale different values as he sees fit, thus eliminating possible redundancy issues.
Question 9: Could you capture the correlation between categorical and continuous variables?
Well, you could, but you would have to use what is known as the Analysis of Covariance (ANCOVA) method. Using it, you could capture the correlation.
Question 10: What is the activation function used for?
This function allows you to diversify your network by introducing non-linear methods of learning. What this will do is that it will help your machine to learn how to process difficult processes more easily.
In this tutorial, we took a look at the interview questions on machine learning. We’ve started with the basics, and later covered some of the more advanced machine learning interview questions and answers that you might receive during your job interview.
Whether you’re looking for a job as an IT specialist or a machine learning AI expert, do your best to revise and remember the basics of ML. Sure, we just briefly touched on the tip of the iceberg, but if you learn these questions and their answers by heart, you should at least develop a general idea of what you can expect from the interview.