Machine Learning

Candidate Companion
Applied Scientists work and solve a broad collection of practical problems that dramatically improve customer
experience, reduce costs, and drive speed and automation. Amazon has the eagerness of a fresh startup to absorb
Machine Learning solutions, and has the scale of a mature firm to help support their development at the same
time.
Amazon has a rich data environment for Applied Scientists to develop new models and algorithms, and use those
to have an impact on the lives of millions of customers. Applied Science is highly experimental in nature, and
needs to be supported through strong theoretical analysis and associated process innovations. At Amazon, Applied
Scientists work closely with Software Engineers to put algorithms into practice. We encourage rigorous customer-
impacting work that involves careful consideration of modeling assumptions, a thorough review of ML literature,
experimentation using state-of-the-art methods, and error-free scalable implementations.

Annexure
• Answering the Machine Learning questions
• Why Amazon
• Technical topics to review
• Depth and Breadth of Knowledge
• Problem Solving and Coding
• Leadership Principles
• Scope of work

Answering the Machine Learning questions

When interviewing for Science roles at Amazon, the course of
the interviews will be interactive; the interviewers will ask you
lots of questions related to the Machine Learning and you are
encouraged to ask the interviewer any necessary questions.

If you are suggesting a technology to solve a problem, please

make sure you understand how that technology works; it is
more important that you understand how your solution solves
the problem than specific technology solutions. We expect you
to think out loud and take hints from the interviewer. You may
want to use a whiteboard, if you do not have access at home
just pen and paper will do fine.

Why Amazon
Please reflect on how you think a career with Amazon would
be mutually beneficial and be prepared to speak to it. Although
“Why Amazon?” is a standard type of question, it’s not a check-
the-box type of formality for us.
We genuinely want to understand how working together with
you would be great, so we get a better sense of who you are.
Our interviewers also appreciate an opportunity to share their
thoughts and experiences, so take a moment to prepare a
couple of questions for the interviewer.
Technical topics to review
Domain Expertise: We are curious about your area of expertise, whether it is Automatic Speech Recognition (ASR),
Natural Language Understanding/ Processing (NLU/ NLP), Computer Vision (CV), Deep- learning (DL), Machine
Learning or Statistical modeling. You will be expected to demonstrate a broader and deep knowledge of your
research area and its literature, deep understanding of the field’s classical methods and your prior work, pros/
cons of modeling approaches, data sources, and practical experience in applying those research ideas to modeling
problems.

Machine Learning Problem Solving/ Application: Of course, we expect you to understand the basic machine
learning methods and algorithms. It is important that you revisit your favorite Machine Learning text books to go
through it. However, it is also important to apply those methods to real world problems. For example, given a
problem definition, you should be able to formulate it as a machine learning problem and propose a solution,
including ideas for data sources, annotation, modeling approach, evaluation, and be able to discuss potential
pitfalls and tradeoffs.

Depth and Breadth of Knowledge

We expect you to have a combined Depth and Breadth of knowledge in Machine Learning. You will be asked
questions regarding topics which are included but not limiting to categories summarized below:

Supervised Learning: Linear & Logistic regression, Naive Bayes classifier, Bagging & Boosting, K-nearest
neighbors, Trees, Neural Networks, Support Vector Machines, Random Forests, Gradient Boosted trees, kernel
methods, Stochastic Gradient Descent (SGD), Sequence Modeling, Bayesian linear regression, Gaussian
Processes, Concepts of overfitting and under fitting, Regularization and evaluation metrics for classification and
regression problems

Unsupervised Learning: Clustering algorithms, k-Means clustering, Anomaly detection, Markov methods,
DBSCAN, Self-organizing maps, Deep Belief Nets, Expectation Maximization (EM), Gaussian Mixture Models
(GMM) and evaluation metrics for clustering problems

Probabilistic graphical models: Bayesian Network, Markov Networks, Variational inference, Markov chain,
Monte Carlo methods, Latent Dirichlet Allocation (LDA), inference methods such as Belief Propagation, Gibbs
Sampling

Dimensionality reduction: Auto encoders, t-SNE, Principal Component Analysis (PCA), Singular Value
Decomposition (SVD), Spectral Clustering and Matrix Factorization

Sequential models: Hidden Markov model (HMM), Conditional random fields (CRF), Recurrent Neural Network
(RNN), Natural Language processing applications such as Named Entity Recognition (NER) and Parts of Speech
(POS) tagging

Reinforcement Learning: State–action–reward–state–action (SARSA), explore-exploit techniques, multi-armed

bandits (epsilon greedy, UCB, Thompson Sampling), Q-learning, and Deep Q-Networks (DQNs). Applied to
domains such as retail, Speech, NLP, Vision, robotics etc.

Deep Neural Networks / Deep-learning: Feed forward Neural Networks, Convolutional Neural Networks,
Backpropagation, Recurrent Neural Networks (RNNs), and Long Short Term Memory (LSTM) networks. GAN,
Attention, Dropout, Vanishing gradient, Activation Functions

Natural Language processing: Statistical Language Modelling, Latent Dirichlet allocation (LDA), Named Entity
Recognition (NER), Word Embedding, Word2Vec, Sentiment Analysis, BERT, ULMFiT

Image and Computer Vision: Object Detection, Image recognition, Pattern recognition, FaceNet, CNN, YOLO
Training and Optimization:
Adaptive gradient approaches, Regularisation and overfitting, loss functions, Bayesian v/s maximum likelihood
estimation, dealing with lass imbalance, K-fold cross validation, bias and variance

Evaluation metrics:
Accuracy, Precision, Recall, Area under ROC, R-squared, Mean average precision (MAP), Mean reciprocal rank,
Equal Error rate, A/B testing fundamentals

Problem Solving and Coding

One of the most rewarding aspects of working in Applied Science is seeing one’s research make a positive impact
on customers. As a Scientist you should be able to implement modelling ideas that can be trained/ tested on large-
scale data. This requires a strong background in data structures and algorithms and ability to write code.

Amazon Applied Scientists primarily code in Python and use a

wide-set of Amazon internal tools to deploy the model
files/code into production systems. However, in some cases
where the latency is a primary constraint they collaborate with
SDE’s within/ outside the team to take these models into
production using Java/C/C++.

Expect to be asked to write syntactically correct code of your

choice (no pseudo code). If you feel a bit rusty coding without
an IDE or coding in a specific language, it’s probably a good idea
to dust off the programming books and tutorials to get
comfortable coding on a whiteboard or on paper.

Interviewer’s expectation:
Key points to keep in mind The interviewer’s expectation is for you to solve the problem
with minimal hints. Interviewer will be interested in your
• You should be able to convert your
thinking process and hearing your explanations before you
thoughts into coding and cover all edge start solving the question on the spot, you could start with a
cases brute force approach and later try to optimize it. Think out loud
• Understanding time and space Complexity and be vocal in your communication with your interviewer.
• Code should be bug free, readable and Also, remember to ask clarifying questions.
modular
• Exposure to dynamic programming and Use Suitable Data Structures:
Coding convention Wherever applicable, pick up suitable Data Structures to solve
the problem. For e.g.: Using Stack for Parentheses Validation
solves the problem easier rather than using Heap. During
preparation, try to learn as many data structures as possible.

Use Suitable Algorithms/ Techniques:

Array is just a simple data structure, but there are plenty of techniques/ algorithms available to solve array
problems optimally. Example: Kadane's Algorithm, Sliding Window Approach, Binary Search, etc. Learn as many
algorithms as possible.

Deciding final approach:

After coming up with different approaches, you should be able to trade-off between different approaches and
choose optimal one among those. Once you check with interviewer on final approach and start coding, you should
be able to convert your approach into working code. You can use any preferable programming language of your
choice but always try to write it in production quality, code hygiene really matters. Once you are done writing the
code make sure you do dry run it yourself and if there are any bugs you should be able to identify them by yourself.
Leadership Principles
We use our Leadership Principles every
day, whether we're discussing ideas for
new projects or deciding on the best
approach to solving a problem. It is just
one of the things that makes Amazon
peculiar.

Our interviews are rooted in behavioral-

based questions which ask about past
situations or challenges you’ve faced and
how you handled them, using Leadership
Principles to guide the discussion. We
avoid brain teasers (e.g., “How many
windows are in Manhattan?”) as part of
the interview process. We’ve researched
this approach and have found that those
types of questions are unreliable when it
comes to predicting a candidate’s success
at Amazon.

The STAR method is a structured manner

of responding to a behavioral-based
interview question by discussing the
specific situation, task, action, and result
of what you're describing. Consider your
own successes and failures in relation to
the Leadership Principles. Have specific
examples that showcase your expertise,
and demonstrate how you’ve taken risks,
succeeded, failed and grown in the
process. Keep in mind, some of Amazon’s
most successful programs have risen
from the ashes of failed projects. Failure
is a necessary part of innovation. It’s not
optional. We understand that and believe
in failing early and iterating until we get it right.

Here are some examples of behavioral-based questions:

• Tell me about a time when you were faced with a problem that had a number of possible solutions what
was the problem and how did you determine the course of action? What was the outcome of that choice?
• When did you take a risk, make a mistake, or fail? How did you respond, and how did you grow from that
experience?
• Describe a time you took the lead on a project?
• What did you do when you needed to motivate a group of individuals or promote collaboration on a
particular project? How have you leveraged data to develop a strategy?

Keep in mind, Amazon is a data-driven company. When you answer questions, your focus should be on the
question asked, ensure your answer is well-structured and provide examples using metrics or data if
applicable. Reference recent situations whenever possible
Scope of work
Our work includes, but are not limited to Recommendation engines, eCommerce fraud detection, Large-scale
optimization, Automated pricing, Demand forecasting, Predicting ad click probabilities, Ranking product search
results, Matching and Classifying products, Search optimizations, Information extraction, Sentiment analysis,
Alexa ASR, Robotics, Natural language understanding, Question answering, Delivery Routes and Conversational
systems.

Some references that might want to deep dive on

• Interviewing at Amazon: https://www.amazon.jobs/en/landing_pages/interviewing-at-amazon
• Virtual Interview Prep: https://www.amazon.jobs/en/landing_pages/virtual-interview-prep
• Data Structures: http://en.wikipedia.org/wiki/List_of_data_structures
• Time complexity: http://en.wikipedia.org/wiki/Time_complexity
• Coding practice problems: geeksforgeeks and leetcode

Amazon Press
• Fortune: Bezos Article I’ve made billions of dollars of failures’
• Day One: The Amazon Blog
• How Amazon Hires
• Amazon Science page
• Publications
• Conferences-and-events
• Select-awards-and-recognition

We appreciate your interest in Amazon