Machine Learning Design

Patterns PDF
Valliappa Lakshmanan

Scan to Download
Machine Learning Design Patterns
Essential Patterns for Solving Common Machine
Learning Challenges
Written by Bookey
Check more about Machine Learning Design Patterns
Summary
Listen Machine Learning Design Patterns Audiobook

Scan to Download
About the book
In "Machine Learning Design Patterns," authors Valliappa
Lakshmanan, Sara Robinson, and Michael Munn compile
essential best practices and proven solutions to common
challenges faced in the machine learning landscape. Drawing
on their extensive experience as Google Cloud engineers, they
present 30 carefully crafted design patterns that address critical
aspects such as data representation, operational efficiency,
reproducibility, and fairness. Each pattern provides clear
problem definitions, a range of viable solutions, and expert
recommendations to guide engineers in selecting optimal
approaches. By engaging with this resource, readers will
enhance their ability to effectively train, evaluate, and deploy
machine learning models, ensuring robust performance and
equitable outcomes for all users.

Scan to Download
About the author
Valliappa Lakshmanan is a distinguished expert in machine
learning and artificial intelligence, known for his contributions
to both academia and industry. With a strong background in
software engineering and data science, he has played a pivotal
role in developing innovative solutions and scalable systems
that harness the power of machine learning. As a thought
leader and educator, Valliappa has authored several influential
publications and frequently shares his insights through
speaking engagements and workshops. His practical
experience, combined with a passion for helping others
navigate the complexities of machine learning, makes him a
sought-after figure in the tech community, particularly for
those looking to implement effective machine learning
strategies and design patterns.

Scan to Download
Summary Content List
Chapter 1 : The Need for Machine Learning Design Patterns

Chapter 2 : Data Representation Design Patterns

Chapter 3 : Problem Representation Design Patterns

Chapter 4 : Model Training Patterns

Chapter 5 : Design Patterns for Resilient Serving

Chapter 6 : Reproducibility Design Patterns

Chapter 7 : Responsible AI

Chapter 8 : Connected Patterns

Scan to Download
Chapter 1 Summary : The Need for
Machine Learning Design Patterns

Chapter 1: The Need for Machine Learning Design

Patterns

Overview of Design Patterns

- Design patterns in engineering encapsulate best practices

and solutions to common problems, allowing practitioners to
leverage expert knowledge.
- This book presents a catalog of machine learning design
patterns drawn from experiences with numerous machine
learning teams.

Scan to Download
Origins and Importance of Patterns

- Originating from architecture, patterns were formalized in

Christopher Alexander's "A Pattern Language" and later
adapted to software design in "Design Patterns: Elements of
Reusable Object-Oriented Software" by Gamma et al.
- In machine learning, adopting proven design patterns is
essential as the discipline matures, helping teams tackle
recurring challenges efficiently.

Purpose of the Book

- The book serves as a reference for common machine

learning problems and effective solutions.
- It outlines various patterns, their associated problems,
possible solutions, and trade-offs, along with implementation
code examples in SQL, scikit-learn, and Keras.

Using the Book

- Readers are encouraged to explore individual patterns rather

than read sequentially, starting with foundational chapters to
understand the framework.

Scan to Download
Machine Learning Terminology

- Defines key terms and concepts including models,

frameworks, data types, and the process of feature
engineering.
- Differentiates between supervised and unsupervised
learning, emphasizing that supervised learning is the primary
focus.

Data and Feature Engineering

- Describes the significance of data quality in machine

learning, highlighting the role of training, validation, and
testing datasets.
- Introduces structured vs. unstructured data and emphasizes
the importance of preprocessing through feature engineering.

Machine Learning Process

- Outlines the lifecycle of building machine learning models:

training, testing/evaluation, and serving.
- Introduces concepts like online vs. batch prediction and the
importance of predictions versus inferences.

Scan to Download
Data and Model Tooling

- Mentions various Google Cloud products that aid in

implementing machine learning solutions, emphasizing
serverless options.
- Tools discussed include BigQuery for data warehousing and
BigQuery ML for model management and predictions.

Roles in Machine Learning

- Defines key roles: data scientists, data engineers, and ML

engineers, explaining their responsibilities and collaborative
nature.

Common Challenges in Machine Learning

- Addresses key challenges such as data quality,

reproducibility, data drift, scaling, and balancing multiple
objectives.
- Emphasizes the importance of data accuracy, completeness,
consistency, and timeliness while detailing strategies for
overcoming obstacles in model training and deployment.

Scan to Download
Summary

- The chapter concludes by reaffirming the significance of

design patterns as a foundation for effective machine learning
practices and how addressing common challenges is pivotal
in the lifecycle of machine learning systems.

Scan to Download
Critical Thinking
Key Point:The Importance of Design Patterns in
Machine Learning
Critical Interpretation:The chapter emphasizes that
machine learning design patterns are critical for
practitioners, yet this reliance on established patterns
may limit innovation and adaptation to unique
challenges faced in different contexts, suggesting that
each scenario may require tailored approaches rather
than a one-size-fits-all solution. This perspective is
worth investigating further as it's debated whether
creative deviation from standard methodologies might
occasionally yield better results. As noted in works like
'The Pragmatic Programmer' by Hunt and Thomas,
flexibility and critical thinking are essential in software
development, which could also apply to machine
learning paradigms.

Scan to Download
 Chapter 2 Summary : Data
Representation Design Patterns

Section Content

Overview Discusses the necessity of effective data representation for machine learning models, which rely on
mathematical functions tailored for specific data types.

Input vs. Feature

Input: Real-world data (e.g., baby weight).
Feature: Transformed data for model operation (e.g., boolean checks).
Feature Engineering: Creating features from inputs.

Mathematical Real-world data must be transformed into boolean variables for models like decision trees.
Representation in
Models

Learnable Data
Representations Embeddings: Dense representations learned by deep neural networks that capture relationships
and convert sparse inputs into lower dimensions.
Feature Extraction: Automated learning of features using methods like embeddings.

Key Design
Patterns Hashed Feature: Uses hashing to manage categorical inputs and reduce memory/model size.
Embeddings: Learnable representations for high-cardinality data, useful in text and categorical
features.
Feature Cross: Captures interactions between multiple categorical features for complex
relationships.
Multimodal Input: Handles various input types (numerical, categorical, image, text) within a
single model.

Simple Data Discusses scaling numerical inputs and converting categorical data into numerical formats using
Representations methods like one-hot encoding.

Complex Data Explores text representation using Bag of Words and embeddings, as well as image representations
Representations emphasizing pixel arrays and convolutional layers.

Scan to Download
Section Content

Summary Outlines the importance of transforming and representing diverse input data effectively to enhance
model training and output. Precedes next chapter on model outputs and prediction representations.

Chapter 2: Data Representation Design Patterns

Overview

Machine learning models rely on mathematical functions

tailored for specific data types, necessitating effective data
representation for real-world data that doesn’t directly fit
these models. This chapter discusses various design patterns
for data representation to enhance model operations.

Input vs. Feature

-
Input
: Real-world data fed into the model (e.g., baby weight).
-
Feature
: Transformed data that models operate on (e.g., boolean
checks like baby weight < 3 kg).

Scan to Download
-
Feature Engineering
: The process of creating features from inputs.

Mathematical Representation in Models

Models like decision trees require boolean variables for

functions, meaning real-world data needs transformation for
compatibility.

Learnable Data Representations

-
Embeddings
: Compact, dense representations learned by deep neural
networks. They convert sparse inputs into lower-dimensional
formats, capturing essential information and relationships
among data points.
-
Feature Extraction
: The automated learning of features through models like
embeddings.

Key Design Patterns

Scan to Download
1.
Hashed Feature
:
- Resolves issues in categorical features by using a hashing
process that groups categorical inputs into a defined number
of buckets, allowing for better memory and model size
management.

2.
Embeddings
:
- Learnable representations for high-cardinality data.
Useful for text and categorical features, embeddings leverage
neural networks to identify relationships between inputs.
3.
Feature Cross
:
- Captures the interaction between multiple categorical
features, allowing models to learn complex relationships
more efficiently regardless of original data separability.
4.
Multimodal Input
:

Scan to Download
 - Handles different input types (numerical, categorical,
image, text) within the same model. This approach helps
models leverage diverse data sources for improved
predictions.

Simple Data Representations

- Discusses numerical input handling, scaling, and various

methods like min-max scaling, z-score normalization, and
winsorizing to enhance model training efficiency.
- Explores categorical and array of categorical variable
representations, emphasizing converting categorical data into
numerical formats using one-hot or dummy coding.

Complex Data Representations

- Addresses text representations through Bag of Words and

embeddings, alongside unique tabular features that describe
aspects of the text data.
- Examines image representation techniques, emphasizing
pixel value arrays and convolutional layers to extract detailed
features while enabling models to perform effectively.

Summary

Scan to Download
Chapter 2 outlines essential patterns for data representation
in machine learning, focusing on the importance of
transforming and effectively representing diverse types of
input data to enhance model training and output
effectiveness. The next chapter will focus on model outputs
and different approaches to representing prediction tasks.

Scan to Download
Example
Key Point:Feature Engineering is crucial for
transforming raw data into meaningful inputs for
machine learning models.
Example:Imagine you have a dataset containing
patients' medical records. You input raw data like age,
cholesterol levels, and blood pressure directly into your
model. However, without transforming these inputs into
features—such as categorizing cholesterol levels into
'high', 'normal', or 'low', or creating binary indicators for
high blood pressure—your model struggles to find
patterns effectively. The resulting features enhance the
model's understanding and predictive power. By
consciously engineering these features, you allow your
model to learn from interactions and relationships in the
data, ultimately leading to more accurate health
predictions.

Scan to Download
Chapter 3 Summary : Problem
Representation Design Patterns

Chapter 3: Problem Representation Design Patterns

This chapter examines various types of machine learning

problems and how they influence model architectures based
on input and output types. The chapter discusses specific
design patterns that are useful in framing and solving
classification and regression problems.

Key Design Patterns Overview

1.
Reframing Design Pattern

- Changes the representation of a machine learning

problem, e.g., transforming a regression problem into a
classification task.
- This can improve predictions by capturing probability
distributions rather than single metrics.
2.

Scan to Download
Multilabel Design Pattern

- Used where training examples can belong to more than

one class.
- Applies sigmoid activation for the final output instead of
softmax, creating multi-hot encoded labels.
3.
Ensemble Design Pattern

- Combines multiple models to enhance predictions.

Methods include bagging (e.g., random forests), boosting
(e.g., AdaBoost, gradient boosting), and stacking.
- These methods help tackle bias and variance issues in
models.
4.
Cascade Design Pattern

- Breaks down complex problems into a series of smaller

tasks where the output of one serves as the input to another.
- Careful design is needed, as errors in initial predictions
can propagate through the cascade.
5. Install Bookey App to Unlock Full Text and
Audio
Neutral Class Design Pattern

Scan to Download
Chapter 4 Summary : Model Training
Patterns

Chapter 4 Summary: Model Training Patterns

Typical Training Loop

- Training machine learning models typically involves a loop

called the training loop.
- Various optimization methods are applied to train ML
models, including gradient descent and its variants like
stochastic gradient descent (SGD).
- SGD is widely used for large datasets and involves training
models in mini-batches, monitoring error, and assessing
overfitting via training and validation datasets.

Keras Training Loop

- Keras simplifies the training loop with functions such as

`fit()` for training and `evaluate()` for model testing.
- Sample code outlines the model compilation, fitting, and

Scan to Download
evaluation process.

Training Design Patterns

- The chapter introduces various design patterns to modify

the traditional training loop:
1.
Useful Overfitting
: Intentionally overfitting to the training dataset when no
unseen data will be encountered.
2.
Checkpoints
: Saving the model state during training to allow recovery
from interruptions.
3.
Transfer Learning
: Using pre-trained models with frozen weights to save time
and resources on smaller datasets.
4.
Distribution Strategy
: Leveraging multiple workers for training large models to
reduce time.
5.
Hyperparameter Tuning

Scan to Download
: Automating the optimization of model hyperparameters to
enhance performance.

Design Pattern 11: Useful Overfitting

- This pattern focuses on situations where overfitting can be

beneficial, particularly in scenarios like simulating physical
systems and using deterministic models without validation
datasets.

Design Pattern 12: Checkpoints

- Checkpoints involve storing the complete state of a model

periodically to resume training after interruptions.
- They provide resilience against failures and are crucial for
effective model training.

Design Pattern 13: Transfer Learning

- Transfer learning reuses parts of pre-trained models for

tasks on specialized datasets, enhancing efficiency and
performance, especially when data is limited.

Design Pattern 14: Distribution Strategy

Scan to Download
- Distributing training workloads across multiple machines or
GPUs speeds up training for large neural networks through
data or model parallelism.
- Synchronous and asynchronous methods for training across
multiple devices are discussed.

Design Pattern 15: Hyperparameter Tuning

- This involves using optimization methods to discover the

best hyperparameter settings efficiently, utilizing approaches
like Bayesian optimization to minimize the trial runs
required.

Conclusion

- The chapter emphasizes various strategies, including

overfitting, checkpoints, and transfer learning, to enhance the
training of ML models, improve efficiency, and ensure
effective hyperparameter tuning in machine learning
workflows.

Scan to Download
Chapter 5 Summary : Design Patterns
for Resilient Serving

Chapter 5: Design Patterns for Resilient Serving

The chapter discusses essential design patterns for deploying

machine learning models in production environments,
emphasizing scalability and resilience.

Stateless Serving Function

The Stateless Serving Function design pattern enables

production ML systems to handle thousands or millions of
prediction requests per second by using stateless functions
that simplify scaling and management.

Batch Serving

Batch Serving allows models to process large volumes of

data asynchronously, leveraging distributed data processing
algorithms. This approach is ideal for non-time-sensitive use
cases.

Scan to Download
Continued Model Evaluation

This pattern focuses on monitoring the performance of

deployed models over time. It addresses challenges like
concept drift and data drift, ensuring timely retraining of
models to maintain effectiveness.

Two-Phase Predictions

In scenarios where models must be deployed at the edge

(e.g., mobile devices), Two-Phase Predictions splits tasks
into simpler on-device models and more complex
cloud-based models, balancing responsiveness with accuracy.

Keyed Predictions

The Keyed Predictions design pattern introduces

client-supplied keys alongside inputs, helping manage
outputs from batch requests and facilitating scalability by
tracking predictions more effectively.

Summary

Scan to Download
The chapter emphasizes modeling techniques aimed at
maintaining the resilience and scalability of machine learning
models in production, detailing patterns that decouple model
training and deployment, manage batch predictions, and
ensure continuous improvement through evaluation and
retraining. The focus is on building systems that can
autonomously adapt to new data and operational demands
while allowing easy integration with different environments
and clients.
The upcoming chapter will delve into reproducibility patterns
in machine learning, addressing the variability and
randomness in model training processes.

Scan to Download
 Chapter 6 Summary : Reproducibility
Design Patterns
Section Summary

Overview Reproducibility in machine learning is essential for consistent model results across different
environments and stages. This chapter covers design patterns to tackle reproducibility challenges.

Transform Design Maintains a separation between raw inputs and transformed features to ensure consistent feature creation,
Pattern preventing training-serving skew.

Repeatable Uses deterministic hash functions on well-distributed columns to consistently split datasets into training,
Splitting Design validation, and test sets, minimizing information leakage.
Pattern

Bridged Schema Allows transitions to newer schemas while leveraging older data by probabilistically or statically
Design Pattern mapping old schema entries to new definitions.

Windowed Externalizes model state and uses sliding windows to manage sequences of data, providing accurate
Inference Design predictions based on real-time data.
Pattern

Workflow Emphasizes end-to-end reproducible pipelines by containerizing workflow steps, facilitating

Pipeline Design collaboration and efficient machine learning management.
Pattern

Feature Store Centralizes feature management and reuse, decoupling feature creation from model development to
Design Pattern ensure consistent access across training and serving.

Model Versioning Allows coexistence of multiple model versions with backward compatibility, providing different REST
Design Pattern endpoints for users to access older models.

Conclusion The patterns discussed facilitate reproducibility in machine learning, addressing challenges in model
training, deployment, and evaluation, leading into the next chapter on responsible AI design patterns.

Chapter 6: Reproducibility Design Patterns

Overview

Reproducibility in machine learning is vital for ensuring that

Scan to Download
models produce consistent results across different
environments and stages, such as training, deployment, and
evaluation. This chapter discusses design patterns aimed at
addressing various reproducibility challenges in machine
learning.

Key Design Patterns

1. Transform Design Pattern

This pattern involves maintaining a clear separation between

raw inputs and the transformed features used by machine
learning models. By explicitly capturing the transformations
applied to inputs, models can ensure consistent feature
creation during both training and serving, which helps
prevent training-serving skew.

2. Repeatable Splitting Design Pattern

This pattern addresses the issue of consistently splitting

Install
datasets intoBookey App to Unlock
training, validation, FullBy
and test sets. Text
usingand
deterministic hash functionsAudio
on well-distributed columns,
such as dates, the likelihood of information leakage among

Scan to Download
Chapter 7 Summary : Responsible AI

Chapter 7: Responsible AI

In this chapter, we explore the broader impact of machine

learning models beyond the development teams, emphasizing
the importance of various stakeholders involved in the model
lifecycle—executives, end users, auditors, and regulatory
agencies.

Stakeholders in ML Models

-
Model Builders
: Data scientists and engineers involved in developing ML
models.
-
ML Engineers
: Teams responsible for deploying the models.
-
Business Decision Makers
: Evaluate the model's fit for business applications.
-

Scan to Download
End Users
: Individuals or groups that utilize model predictions.
-
Regulatory and Compliance Agencies
: Evaluate compliance from a regulatory perspective.

Patterns for Responsible AI

The chapter introduces several design patterns under

Responsible AI:
1.
Heuristic Benchmark
: This pattern allows comparison of model performance
against a simple heuristic to provide context and
understanding for stakeholders.
2.
Explainable Predictions
: Techniques to enhance user trust by elucidating how models
arrive at certain predictions, thus addressing trust issues
particularly in high-stakes fields like healthcare and finance.
3.
Fairness Lens
: Methods to ensure equitable model behavior across diverse
user groups and use cases.

Scan to Download
The need for responsibility in AI practices pertains to aspects
like fairness, interpretability, privacy, and security.
Recommended practices include employing human-centered
design, understanding dataset limitations, and continuous
monitoring post-deployment.

Design Patterns Explained

Heuristic Benchmark

This pattern aids in contextualizing model performance using

simple, intuitive benchmarks. For instance, comparing an
ML model's metrics to its operational predecessors provides
meaningful insights for stakeholders.

Example Applications
:
- MAE for bicycle rental durations.
- MAP for product catalog classification.
Trade-offs include the utility of heuristic benchmarks for
comparisons, especially when existing methods are
operational.

Scan to Download
Explainable Predictions

This design pattern focuses on increasing trust in machine

learning models through feature attribution methods. These
reveal the driving factors behind model predictions,
enhancing transparency and aiding in bias detection.

Types of Feature Attributions

:
-
Instance-Level
: Explain predictions for individual instances.
-
Global
: Assess model behavior collectively across a dataset.
Tools like SHAP and Integrated Gradients are highlighted for
effective computation of feature attributions.

Fairness Lens

This pattern addresses bias in modeling by applying

systematic evaluations of fairness through techniques,
ensuring models treat all groups equitably. The chapter
discusses mitigation strategies for identifiably problematic

Scan to Download
biases that occur throughout the data lifecycle, including:
- Data Collection Bias.
- Representation Bias.
- Labeling Bias.
Methods using the What-If Tool are discussed for analysis
and visualization of data representation.

Summary

Responsible AI embodies ensuring ethical, unbiased

interpretations of ML models by assessing the implications
for diverse stakeholders throughout the ML lifecycle. The
patterns introduced provide frameworks to enhance model
evaluation, interpretability, and fairness in machine learning
systems, emphasizing the need for accountability in
deploying AI technologies.

Scan to Download
Critical Thinking
Key Point:The Responsibility of Stakeholders in AI
Model Deployment
Critical Interpretation:A crucial aspect emphasized in
this chapter is the multi-faceted responsibility of various
stakeholders involved in the lifecycle of machine
learning models, which poses the question of
accountability. While the author asserts that involving
stakeholders such as executives, end users, and
regulatory agencies is pivotal for fostering responsible
AI practices, it's essential to recognize that such
involvement does not inherently guarantee ethical
outcomes. Critics may argue that merely including
diverse perspectives does not address fundamental
biases and systemic issues within AI, as highlighted by
researchers like Cathy O'Neil in "Weapons of Math
Destruction". Stakeholder engagement must also be
paired with rigorous assessments of bias, continuous
monitoring, and an unwavering commitment to
transparency to genuinely achieve responsible AI.

Scan to Download
Chapter 8 Summary : Connected
Patterns

Connected Patterns

In this chapter, we create a reference catalog of machine

learning design patterns that help solve recurring problems in
the design, training, and deployment of machine learning
models and pipelines. The patterns are organized according
to their application within an ML workflow, covering input
representation, model selection, modifications to training
loops, and ensuring responsible use of ML systems. This
chapter not only lists these patterns but also discusses their
interrelations and how they can be applied effectively for
common ML tasks.

Patterns Reference

The chapter summarizes various design patterns along with

their corresponding problems and suggested solutions, which
include:
-

Scan to Download
Data Representation

-
Hashed Feature
: Solves problems with categorical features by hashing
strings.
-
Embeddings
: Preserves relationships in high-cardinality features by
mapping them to lower dimensions.
-
Feature Cross
: Enhances model learning by introducing combinations of
input features.
-
Multimodal Input
: Combines multiple data representations.
-
Problem Representation

-
Reframing
: Changes the output representation of ML problems.
-

Scan to Download
Multilabel
: Handles multiple labels for a single example.
-
Ensembles
: Combines models to address bias-variance trade-offs.
-
Cascade
: Integrates multi-step problems to enhance maintainability.
-
Patterns That Modify Model Training

-
Overfitting
: Intentionally overfits to learn specific models.
-
Checkpoints
: Saves model state during long training processes.
-
Transfer Learning
: Reuses parts of previously trained models.
-
Hyperparameter Tuning
: Optimizes model parameters.
-

Scan to Download
Resilience

-
Stateless Serving Function
: Allows scalable model predictions.
-
Batch Serving
: Efficiently processes many predictions at once.
-
Continued Model Evaluation
: Monitors model performance over time.
-
Reproducibility

-
Transform
: Ensures consistent input transformation across training and
serving.
-
Repeatable Splitting
: Guarantees data splits are reliable.
-
Bridged Schema
: Adapts data schemas for continuous retraining.

Scan to Download
 -
Windowed Inference
: Combines feature aggregation without training-serving
skew.
-
Responsible AI

-
Heuristic Benchmark
: Provides simpler evaluation metrics for model assessments.
-
Explainable Predictions
: Enhances understanding of model decisions.
-
Fairness Lens
: Evaluates models for bias.

Pattern Interactions

Patterns often interact with each other. For example, the

Hashed Feature
and
Embeddings
patterns can work synergistically for high-cardinality inputs.

Scan to Download
The chapter emphasizes the utility of combining patterns to
enhance the effectiveness of ML solutions.

Patterns Within ML Projects

The chapter outlines the ML life cycle that consists of

discovery, development, and deployment phases. Each phase
builds upon the success of the previous one:
-
Discovery
: Defines business use cases, evaluates project viability, and
explores data quality.
-
Development
: Involves setting up data pipelines, feature engineering,
model training, and evaluation.
-
Deployment
: Focuses on operationalizing the model, establishing
MLOps, and implementing continuous monitoring for model
performance.

AI Readiness

Scan to Download
Companies may exist in different stages of AI readiness,
characterized as tactical (manual development), strategic
(using pipelines), and transformational (fully automated
processes). Each phase represents advancements in
capability, automation, and integration into business
operations.

Common Patterns by Use Case and Data Type

Various design patterns can be particularly useful depending

on specific ML use cases such as natural language
understanding, computer vision, predictive analytics,
recommendation systems, and fraud detection. Each use case
has tailored patterns that enhance performance and efficacy
in achieving desired outcomes.

Scan to Download
Best Quotes from Machine Learning
Design Patterns by Valliappa
Lakshmanan with Page Numbers
View on Bookey Website and Generate Beautiful Quote Images

Chapter 1 | Quotes From Pages -34

1.In engineering disciplines, design patterns capture
best practices and solutions to commonly
occurring problems.
2.Each pattern describes a problem which occurs over and
over again in our environment, and then describes the core
of the solution to that problem, in such a way that you can
use this solution a million times over, without ever doing it
the same way twice.
3.Building production machine learning models is
increasingly becoming an engineering discipline, taking
advantage of ML methods that have been proven in
research settings and applying them to business problems.
4.One benefit of our jobs in the customer-facing part of
Google Cloud is that it brings us in contact with a wide

Scan to Download
 variety of machine learning and data science teams and
individual developers from around the world.
5.We will have succeeded if this book gives you and your
team a vocabulary when talking about concepts that you
already incorporate intuitively into your ML projects.
6.Common challenges in machine learning tend to revolve
around data quality, reproducibility, data drift, scale, and
having to satisfy multiple objectives.
Chapter 2 | Quotes From Pages 35-94
1.The process of creating features to represent the
input data is called feature engineering, and so we
can think of feature engineering as a way of
selecting the data representation.
2.The learned data representation consists of baby weight as
the input variable, the less than operator, and the threshold
of 3 kg.
3.The learning algorithm needs to extract the most salient
information from the input and represent it in a more
concise way in the feature.

Scan to Download
4.A decision tree where each node can represent only one
input variable reduces to a stepwise linear function,
whereas an oblique decision tree... reduces to a piecewise
linear function.
5.By scaling all features to lie between [–1, 1], we ensure that
there is not much of a difference in the relative magnitudes
of different features.
6.A feature cross is a synthetic feature formed by
concatenating two or more categorical features in order to
capture the interaction between them.
7.The Embeddings design pattern addresses the problem of
representing high-cardinality data densely in a lower
dimension by passing the input data through an embedding
layer that has trainable weights.
8.By treating day of the week as a categorical feature... we
discover insights that would otherwise remain hidden in a
purely numeric representation.
9.While it is acceptable to throw away invalid input, it is not
acceptable to throw away valid data.

Scan to Download
10.In a hybrid model, we can take the best of both worlds by
extracting features from raw text and using them in
conjunction with embeddings.
Chapter 3 | Quotes From Pages 95-154
1.The first step of building any machine learning
solution is framing the problem.
2.Changing the context and reframing the task of a problem
can help when building a machine learning solution.
3.By modeling a classification as a discrete probability
distribution, the model is able to capture the bimodal
structure of the predictions.
4.The Cascade design pattern addresses situations where a
machine learning problem can be profitably broken into a
series of ML problems.
5.There is rarely just one way to frame a problem, and it is
helpful to be aware of any trade-offs or alternatives of a
given implementation.
6.In practice, it can become hard to keep a Cascade workflow
straight.

Scan to Download
7.The Neutral Class design pattern is one to keep in mind at
the beginning of a machine learning problem.
8.Choosing the right tool for the problem is crucial in
machine learning.
9.For imbalanced datasets, metrics other than accuracy are
preferred.
10.Combining different techniques can lead to optimal
results in handling imbalanced datasets.

Scan to Download
Chapter 4 | Quotes From Pages -216
1.An ML approximation is useful in situations where
the lookup table will be too large to effectively use.
2.If your neural network is able to learn a set of parameters
where the loss function is zero, then that parameter set
determines the actual solution of the PDE in question.
3.When we use checkpoints, we usually also use virtual
epochs, wherein we decide to carry out the inner loop of
the fit() function, not on the full training dataset but on a
fixed number of training examples.
4.Transfer learning works because it lets us stand on the
shoulders of giants, utilizing models that have already been
trained on extremely large, labeled datasets.
5.Instead of using early stopping or checkpoint selection, it
can be helpful to try to add L2 regularization to your model
so that the validation error does not increase and the model
never gets into phase 3.
6.The keras-tuner library implements Bayesian optimization
to do hyperparameter search directly in Keras.

Scan to Download
7.As the number of workers increases, the time to minimum
training loss dramatically decreases.
Chapter 5 | Quotes From Pages -264
1.The purpose of a machine learning model is to use
it to make inferences on data it hasn’t seen during
training.
2.Software that is deployed into production environments is
expected to be resilient and require little in the way of
human intervention to keep it running.
3.The Stateless Serving Function design pattern makes it
possible for a production ML system to synchronously
handle thousands to millions of prediction requests per
second.
4.Because stateless components don’t have any state, they
can be shared by multiple clients.
5.The Continued Model Evaluation design pattern handles
the common problem of detecting when a deployed model
is no longer fit-for-purpose.
6.The Two-Phase Predictions design pattern provides a way

Scan to Download
 to address the problem of keeping models sophisticated and
performant when they have to be deployed onto distributed
devices.
7.The Keyed Predictions design pattern is a necessity to
scalably implement several of the design patterns discussed
in this chapter.
Chapter 6 | Quotes From Pages 265-334
1.... this sort of reproducibility is difficult in machine
learning.
2.The Transform design pattern makes moving an ML model
to production much easier by keeping inputs, features, and
transforms carefully separate.
3.We have to know that we need to send in '3' for dayofweek.
That '3' ...is that Tuesday or Wednesday? Depends on
which library was used by the model, or what we consider
the start of a week!
4.In order to avoid training–serving skew, it is preferable to
use the same SQL in a tumbling window function in a
streaming pipeline.

Scan to Download
5.To handle this, we’ll need a way to track how the model is
performing in production and compare this with previous
iterations.
6.By keeping multiple versions of a model deployed at a
given time, those users relying on older versions will still
be able to use the service.
7.With each pipeline component in its own container,
different team members can build and test separate pieces
of a pipeline in parallel.
8.A feature store acts as a version-controlled repository for
feature datasets, allowing the same CI/CD practices of code
and model development to be applied to the feature
engineering process.
9.The added advantage of feature reusability that tf.transform
does not have.
10.In the absence of any extra information, we will assume
that the frequency distribution is the same for all the
historical examples.

Scan to Download
Chapter 7 | Quotes From Pages 335-374
1.With great power comes great responsibility.
2.Responsible AI patterns are not limited to the three that we
discuss in this chapter.
3.A good heuristic benchmark should be intuitively easy to
understand and relatively trivial to compute.
4.The Explainable Predictions design pattern increases user
trust in ML systems by providing users with an
understanding of how and why models make certain
predictions.
5.Bias becomes harmful when it affects different groups of
people differently.
6.The goal of Model Cards is to improve model transparency
by providing details on scenarios where a model should
and should not be used.
Chapter 8 | Quotes From Pages -398
1.Patterns don’t exist in isolation. Many of them are
closely related to one another either directly or
indirectly and often complement one another.

Scan to Download
2.A successful discovery stage requires collaboration
between the business domain experts as well as machine
learning experts to assess the viability of an ML approach.
3.Machine learning is not the answer to all problems, and
sometimes a rule-based heuristic is hard to beat.
4.The model staleness can occur for many reasons.
Retraining models periodically can help to improve their
performance over time.
5.It is important to anticipate these challenges [in
productionizing models], and developing a comprehensive
solution requires significant investment from the business
operations side to make the transition as easy as possible
and increase the speed to market.

Scan to Download
Machine Learning Design Patterns
Questions
View on Bookey Website

Chapter 1 | The Need for Machine Learning Design

Patterns| Q&A
1.Question
What is the main purpose of machine learning design
patterns?
Answer:Machine learning design patterns serve to
capture best practices and solutions to common
problems encountered in machine learning
engineering. They codify the experiences and
knowledge of experts, providing a framework that
practitioners can follow to improve the efficiency
and effectiveness of their work.

2.Question
How do design patterns relate to traditional
programming and architecture?
Answer:Design patterns in machine learning are inspired by
concepts from architecture and traditional programming.

Scan to Download
Like architectural design patterns that provide solutions for
building layouts, ML design patterns offer reusable solutions
for common issues faced during ML model development,
much like object-oriented design patterns do in software
engineering.

3.Question
Can you explain the ‘garbage in, garbage out’ concept in
the context of data quality?
Answer:The phrase 'garbage in, garbage out' indicates that
the quality of a machine learning model's output is directly
proportional to the quality of the input data. If the training
data is of poor quality—meaning it's incomplete, inaccurate,
or biased—then the model's predictions will also be flawed,
reflecting those same limitations.

4.Question
What are the different types of machine learning
problems?
Answer:Machine learning problems can be broadly classified
into supervised and unsupervised learning. Supervised

Scan to Download
learning involves labeled training data, used for tasks like
classification and regression, while unsupervised learning
deals with data without pre-existing labels, aimed at
discovering patterns or groupings.

5.Question
What role does data validation play in the machine
learning process?
Answer:Data validation is crucial as it involves analyzing the
dataset used for training and testing, ensuring it has balanced
statistical properties, verifying data types for each feature,
and identifying potential biases or inaccuracies that could
affect model performance.

6.Question
How does one achieve reproducibility in machine
learning?
Answer:Reproducibility in machine learning can be achieved
by controlling various elements like the random seed value,
ensuring consistent environmental conditions, and
standardizing framework dependencies. Setting fixed

Scan to Download
parameters minimizes randomness and allows for consistent
results across different training runs.

7.Question
What is data drift and why is it a challenge in machine
learning?
Answer:Data drift occurs when the statistical properties of
the input data change over time, causing models trained on
historical data to become less effective. Keeping models
relevant requires regularly updating training datasets,
retraining models, and adjusting their parameters as needed.

8.Question
In what ways can machine learning design patterns assist
practitioners?
Answer:Machine learning design patterns assist practitioners
by providing a common vocabulary, framing solutions to
frequently encountered problems, and offering best practices
that enhance collaboration and communication within teams,
ultimately leading to more efficient and effective machine
learning workflows.

Scan to Download
9.Question
What are some common roles in a machine learning
team, and how do they collaborate?
Answer:Common roles in a machine learning team include
data scientists, data engineers, and machine learning
engineers. Data scientists focus on data analysis and model
building, data engineers manage the infrastructure and data
pipelines, while ML engineers deploy models and handle
production aspects. They collaborate closely throughout the
machine learning workflow—from data ingestion to model
serving.

10.Question
How does understanding multiple objectives help in
model optimization?
Answer:Recognizing and balancing multiple objectives
across teams within an organization can inform model
optimization strategies. By identifying differing
priorities—such as minimizing loss for data scientists and
reducing customer costs for product teams—practitioners can

Scan to Download
align model goals to satisfy various stakeholders effectively.
Chapter 2 | Data Representation Design Patterns|
Q&A
1.Question
Why is it important to understand the types of input data
for machine learning models?
Answer:Understanding input data types is crucial
because machine learning models rely on specific
data formats. The mathematical core of many
models uses boolean variables or numerical values,
and if the incoming data doesn't match these
formats, the model won't function correctly. For
example, decision trees operate on boolean
variables, so categorical data must be transformed
into a compatible format.

2.Question
What is feature engineering, and why is it significant in
the context of machine learning?
Answer:Feature engineering involves selecting, modifying,
or creating features from raw data to improve model

Scan to Download
performance. It is significant because well-designed features
can enhance a model's ability to learn patterns from the data,
leading to more accurate predictions.

3.Question
Can you explain how to make categorical variables
compatible with machine learning models?
Answer:Categorical variables can be transformed into
boolean values or numerical representations compatible with
machine learning models. For instance, we can convert a
categorical variable such as 'hospital location' into a boolean
format (e.g., IS 'in France') or through techniques like
one-hot encoding where each category is represented as a
binary vector.

4.Question
What are embeddings, and how do they benefit machine
learning models?
Answer:Embeddings are learned data representations that
map high-cardinality data into a lower-dimensional space
while preserving information that is relevant for the task.

Scan to Download
They benefit machine learning models by capturing
relationships between categories or items, thus improving
performance on tasks like recommendation or classification.

5.Question
How do feature crosses enhance model predictiveness?
Answer:Feature crosses create new features by combining
existing variables, allowing models to learn interactions
between features more effectively. This can uncover
non-linear relationships in the data that individual features
may not capture, thus improving predictive performance.

6.Question
What is the advantage of using multimodal inputs in
machine learning models?
Answer:Multimodal inputs allow models to leverage
different types of data (like images, text, or tabular data)
simultaneously. This can provide a richer representation of
the problem domain, potentially leading to better model
performance since it captures a wider scope of information.

7.Question
Why is scaling numerical inputs beneficial in machine

Scan to Download
learning?
Answer:Scaling numerical inputs to a standard range (like
[-1, 1]) can speed up model convergence during training. It
ensures that gradient descent works effectively by preventing
features with larger ranges from disproportionately
influencing the optimization process.

8.Question
What are the trade-offs of using hashed features in a
machine learning model?
Answer:The trade-offs of using hashed features include a loss
of precision due to bucket collisions, which can reduce the
model's accuracy. However, they help manage high
cardinality and cold-start scenarios by providing a consistent,
compact representation of categorical variables without the
need for complete vocabulary knowledge.

9.Question
Why would a model benefit from both bag-of-words and
embedding representations for text data?
Answer:Using both representations allows the model to

Scan to Download
capture different aspects of the text data: bag-of-words
focuses on the presence and absence of keywords, while
embeddings capture semantic relationships and contextual
information. Combining them can enhance the model's
understanding and improve prediction accuracy.

10.Question
How does feature engineering aid in addressing skewed
data distributions?
Answer:Feature engineering can involve using
transformations or creating new features that help make data
distributions more uniform or normally distributed. This may
include applying nonlinear transformations or selecting
representative samples from the data to support better model
training and performance.
Chapter 3 | Problem Representation Design
Patterns| Q&A
1.Question
What is the central theme of Chapter 3 in 'Machine
Learning Design Patterns'?
Answer:Chapter 3 focuses on various problem

Scan to Download
representation design patterns in machine learning,
examining how different types of machine learning
problems influence model architectures based on
input and output types, especially through
reframing, ensemble methods, and handling
imbalanced datasets.

2.Question
What are the advantages of using the Reframing design
pattern?
Answer:The Reframing design pattern allows changing the
output representation of a machine learning problem (e.g.,
from regression to classification), which can improve model
performance by capturing probabilistic distributions of
outcomes instead of just predicting single values,
accommodating complex data behaviors.

3.Question
Can you explain the difference between Reframing and
Multilabel design patterns?
Answer:Reframing changes how a single problem is

Scan to Download
represented (like switching between regression and
classification), while the Multilabel design pattern allows an
instance to belong to multiple classes simultaneously, often
requiring sigmoid outputs for probability estimates on each
class.

4.Question
How does the Ensemble design pattern contribute to
model performance?
Answer:The Ensemble design pattern enhances performance
by combining multiple models (e.g., through bagging,
boosting, or stacking), utilizing the diversity of model
predictions to reduce error, improve accuracy, and achieve a
better overall predictive capability compared to individual
models.

5.Question
Why is handling imbalanced datasets crucial in machine
learning?
Answer:Imbalanced datasets can lead to models that perform
well on majority classes while neglecting minority classes,

Scan to Download
leading to misleading accuracy. Techniques like rebalancing,
upsampling, downsampling, or adjusting class weights are
crucial to ensure models learn to recognize all classes
adequately.

6.Question
What is the purpose of the Neutral Class design pattern?
Answer:The Neutral Class design pattern introduces a third
class in classification tasks to capture uncertain or arbitrary
instances that do not fit neatly into the primary categories,
improving the model's accuracy and providing a more
comprehensive understanding of classification uncertainties.

7.Question
How can the Cascade design pattern be advantageous in
certain machine learning applications?
Answer:The Cascade design pattern breaks complex
problems into simpler subproblems, allowing distinct models
to be trained on different segments of data (e.g., typical vs.
atypical behaviors), which can enhance performance and
accuracy for nuanced tasks that have distinct behaviors.

Scan to Download
8.Question
What metrics should be prioritized when evaluating
models on imbalanced datasets?
Answer:For imbalanced datasets, accuracy can be
misleading. Metrics like precision, recall, F-measure, or
average precision-recall should be prioritized, as they provide
a clearer picture of the model's performance on minority
classes.

9.Question
What are the potential trade-offs when implementing
ensemble methods in a model?
Answer:While ensemble methods can improve predictive
power, they can also increase complexity, require more
computational resources, and complicate model
interpretability, which should be weighed against the
performance benefits.

10.Question
In what situations should the Cascade design pattern be
avoided?
Answer:The Cascade design pattern should be avoided in

Scan to Download
scenarios where a single model suffices, where deterministic
inputs can be added without cascading, or where internal
consistency among predictions is not critical, as it adds
unnecessary complexity.

Scan to Download
Chapter 4 | Model Training Patterns| Q&A
1.Question
What is the purpose of the training loop in machine
learning?
Answer:The training loop is essential for iteratively
refining the parameters of a machine learning model
to minimize error and achieve convergence on a
model that can generalize well to unseen data.

2.Question
Why is stochastic gradient descent (SGD) preferred for
training large datasets?
Answer:SGD operates on mini-batches, which allows it to
efficiently handle large datasets by scaling computations and
improving convergence speed compared to traditional
gradient descent methods.

3.Question
What is overfitting, and why might it be useful in specific
scenarios?
Answer:Overfitting occurs when a model learns the training
data too well, capturing noise instead of the underlying data

Scan to Download
distribution. In scenarios where the complete input space can
be covered, such as simulating physical systems with no
unseen data, overfitting can lead to perfect predictions and is
intentionally beneficial.

4.Question
What role does a validation dataset play in the training
process?
Answer:A validation dataset is used to monitor the model's
performance during training, helping to identify overfitting
by comparing training error with validation error to ensure
the model generalizes well.

5.Question
How does the Keras training loop typically function?
Answer:The Keras training loop initializes the model,
compiles it with an optimizer and loss function, fits the
model to the training data in epochs, evaluates on a test set,
and saves the model state.

6.Question
What is the purpose of checkpoints in model training?
Answer:Checkpoints save the model's state at certain

Scan to Download
intervals during training, allowing for recovery from
interruptions, early stopping, and potentially reusing partially
trained models.

7.Question
Explain the concept of transfer learning in machine
learning.
Answer:Transfer learning utilizes a pre-trained model as a
starting point for a similar task. It involves freezing initial
layers that capture general features, allowing the model to
adapt to a new, often smaller dataset by training only the
final layers.

8.Question
What are distribution strategies and why are they
important?
Answer:Distribution strategies enable the training loop of
large models to scale across multiple workers, improving
training efficiency and reducing training time significantly by
parallelizing computations.

9.Question
How does hyperparameter tuning improve machine

Scan to Download
learning models?
Answer:Hyperparameter tuning optimizes the settings that
dictate model training, such as learning rates and layer
configurations, to enhance model performance, ensuring that
the algorithm finds the best configuration efficiently.

10.Question
Why is it crucial to monitor both training and validation
metrics during the training process?
Answer:Monitoring both metrics is critical to ensure the
model is learning effectively, allowing for adjustments to
avoid overfitting and ensuring that the model accurately
generalizes to new, unseen data.

11.Question
What are the trade-offs associated with overfitting?
Answer:While overfitting can be useful in scenarios where
complete data coverage exists, it can lead to poor
generalization in other cases, necessitating careful
consideration of model complexity relative to dataset size.

12.Question
What is gradient descent and why is it fundamental to

Scan to Download
training machine learning models?
Answer:Gradient descent is an optimization algorithm used
to minimize the loss function of a model by iteratively
updating parameters based on the gradient of the loss with
respect to those parameters, essential for effective training.

13.Question
What is Bayesian optimization, and how does it enhance
hyperparameter tuning?
Answer:Bayesian optimization is a method for optimizing
black-box functions that can efficiently find the best set of
hyperparameters for a model by building a surrogate
function, reducing the number of costly direct model
evaluations needed.

14.Question
Describe the difference between synchronous and
asynchronous training in distributed learning
environments.
Answer:In synchronous training, all workers operate in
unison, waiting for each other to complete before updating

Scan to Download
the model parameters, while in asynchronous training,
workers independently compute updates without waiting,
increasing throughput but potentially introducing stale
updates.

15.Question
How can using checkpoints facilitate model retraining
and fine-tuning?
Answer:Checkpoints allow models to resume training from a
previous state after interruptions or to fine-tune an existing
model by continuing training from a saved point rather than
starting over, saving time and computational resources.

16.Question
What is the significance of using a testing dataset separate
from training and validation sets?
Answer:A testing dataset provides an unbiased evaluation of
the model's performance on previously unseen data, ensuring
that metrics reported are reflective of how the model will
perform in real-world applications.

17.Question
Explain the concept of 'virtual epochs'.

Scan to Download
Answer:Virtual epochs refer to the method of treating subsets
of data as if they were multiple epochs of training, allowing
for intermediate checkpoints without fully traversing the
entire dataset repeatedly, optimizing efficiency.

18.Question
Why might a practitioner choose to overfit a small batch
during model debugging?
Answer:Overfitting a small batch is a valuable sanity check
that confirms the model code and input pipeline are
functioning correctly; success in overfitting indicates that the
model architecture is capable of learning the training data
effectively.

19.Question
What are common parameters that can be adjusted in
hyperparameter tuning?
Answer:Common hyperparameters include learning rate,
number of epochs, batch size, number of layers in the
network, types of activation functions, and regularization
methods.

Scan to Download
Chapter 5 | Design Patterns for Resilient Serving|
Q&A
1.Question
What is the main purpose of deploying a machine
learning model in a production environment?
Answer:The primary purpose is to enable the model
to make predictions on unseen data that it did not
encounter during training, thereby providing
real-time inferences in response to incoming
requests.

2.Question
Why is resilience important for machine learning models
in production?
Answer:Resilience is crucial because production models need
to operate continuously with minimal human intervention,
ensuring they can handle large volumes of requests, maintain
performance, and adapt to changing conditions.

3.Question
What are Stateless Serving Functions and how do they
enhance scalability?

Scan to Download
Answer:Stateless Serving Functions are designed to handle
thousands or millions of prediction requests per second
synchronously without retaining state, which allows servers
to efficiently manage resources and scale as needed.

4.Question
How does the Batch Serving design pattern differ from
stateless serving?
Answer:Batch Serving allows for asynchronous processing of
large volumes of predictions at once, rather than handling
individual requests immediately, which is particularly useful
for non-real-time predictions.

5.Question
What challenges can arise from deploying a machine
learning model that has concept drift?
Answer:Concept drift can lead to a situation where the
relationship between the model inputs and outputs changes,
causing the model to produce inaccurate or less useful
predictions over time.

6.Question
What is the role of the Continued Model Evaluation

Scan to Download
design pattern?
Answer:This pattern continuously monitors the model's
performance against new data and determines when
retraining is necessary to ensure accuracy and relevance of
predictions.

7.Question
In the context of edge devices, what is the Two-Phase
Predictions design pattern?
Answer:The Two-Phase Predictions design pattern splits a
prediction task into two phases: one simpler model runs on
the edge device for quick inferences, while a more complex
model running in the cloud is used for detailed analyses
when necessary.

8.Question
What benefits do Keyed Predictions provide when
making predictions?
Answer:Keyed Predictions allow clients to attach unique
identifiers to their inputs, ensuring that even when
predictions are processed out of order, they can still be

Scan to Download
matched back to the relevant inputs, facilitating accurate
output management.

9.Question
How does frequent retraining contribute to maintaining
model performance?
Answer:Frequent retraining addresses and mitigates issues
caused by data drift and concept drift, allowing the model to
adapt to new patterns in data and maintain its predictive
accuracy over time.

10.Question
What practical strategies can be employed for continuous
evaluation of models in production?
Answer:Strategies include sampling predictions in real-time
to evaluate against ground truth, using human labeling
services for accuracy assessment, and implementing
automated systems to trigger retraining based on
performance drops.
Chapter 6 | Reproducibility Design Patterns| Q&A
1.Question
Why is reproducibility important in machine learning

Scan to Download
model training and deployment?
Answer:Reproducibility ensures that model results
can be replicated consistently across different runs,
which is crucial for validating the performance and
reliability of machine learning models. Without
reproducibility, models may yield different outputs
due to randomness in initialized parameters, data
splitting, or feature transformations, making it
difficult to assess their effectiveness in real-world
applications.

2.Question
What is the purpose of the Transform design pattern in
machine learning?
Answer:The Transform design pattern captures the data
preparation transformations from the training pipeline to the
serving pipeline. By explicitly detailing how raw inputs are
converted into model features, it ensures that the same
transformations are consistently applied during both training
and inference, thereby reducing training-serving skew.

Scan to Download
3.Question
How does the Repeatable Splitting design pattern
mitigate information leakage in model evaluation?
Answer:The Repeatable Splitting design pattern uses a
deterministic hash function to split data into training,
validation, and test sets based on a well-distributed column.
This approach keeps related rows together, preventing
leakage where training examples could inadvertently
influence evaluation results, thus maintaining the integrity of
the model evaluation.

4.Question
What is the role of versioning in the context of machine
learning models?
Answer:Versioning allows multiple iterations of a model to
be deployed concurrently, enabling backward compatibility.
This is important for users relying on specific model outputs,
as it allows them to continue using a version that meets their
needs while also facilitating testing and transitioning to
improved models.

Scan to Download
5.Question
How do feature stores improve the efficiency of machine
learning workflows?
Answer:Feature stores centralize and manage features across
different projects, allowing for shared access and reuse. This
reduces redundant feature engineering efforts, ensures
consistent feature transformations, and enables better
governance of sensitive data, ultimately speeding up the
development of new machine learning models.

6.Question
What challenges does the Windowed Inference design
pattern address?
Answer:The Windowed Inference design pattern manages the
computation of features that depend on historical data in a
time-sensitive manner. It ensures that necessary features,
calculated over specific time windows, are consistently and
correctly processed both in training and during real-time
inference.

7.Question
Can you describe how the Bridged Schema design pattern

Scan to Download
works in integrating older and newer datasets?
Answer:The Bridged Schema design pattern allows for
combining datasets that adhere to different schemas by
inferring or transforming older data to match newer data
representations. This enables practitioners to utilize both
historic data and newly collected features, maintaining model
accuracy while leveraging time-sensitive improvements in
data collection.

8.Question
What are the trade-offs of using a Feature Store
compared to traditional feature engineering methods?
Answer:While Feature Stores promote reusability and
consistency across features, they introduce additional
complexity due to the need for infrastructure and governance
around feature data. They can also require more initial setup
and overhead when compared to ad-hoc feature engineering
methods, which may be faster for quick experiments but are
unsustainable at scale.

Scan to Download
Chapter 7 | Responsible AI| Q&A
1.Question
What are the different stakeholders involved in machine
learning models according to the chapter?
Answer:The stakeholders include model builders
(data scientists and ML researchers), ML engineers
(involved in deploying models), business decision
makers (evaluate model fit for business use), end
users (customers and employees who utilize the
model's predictions), and regulatory and compliance
agencies (require summaries of model
decision-making for compliance).

2.Question
What is the purpose of the Heuristic Benchmark design
pattern?
Answer:The Heuristic Benchmark pattern allows for
comparing a model's performance against a simple, intuitive
heuristic to provide context to business decision makers,
helping them understand whether a model's performance is

Scan to Download
acceptable or not.

3.Question
How does the Explainable Predictions design pattern
enhance trust in ML systems?
Answer:By providing insights into how and why models
make predictions, this pattern utilizes feature attributions to
clarify the contributions of various features to a model's
output, fostering user trust and compliance with regulatory
standards.

4.Question
What is problematic bias and why is it significant in
machine learning?
Answer:Problematic bias is when a model's predictions
adversely affect different groups of people, often inherent
due to biased training datasets. It is crucial to address this
bias as it can lead to unfair outcomes in real-world
applications.

5.Question
Can you give an example of a heuristic benchmark?
Answer:An example would be predicting the duration of

Scan to Download
bicycle rentals by using the mean or median rental duration
over historical data as a benchmark to compare against a
newly developed ML model's predictions.

6.Question
Why is it important to define a baseline in explainability
methods?
Answer:Defining a baseline allows explanation values to be
understood in context, highlighting how much a specific
feature contributed to a model's output compared to a
standard or average prediction.

7.Question
What does the chapter suggest regarding monitoring ML
systems after deployment?
Answer:It emphasizes the importance of continued
evaluation to identify and rectify any biases or issues that
might arise post-deployment, ensuring models remain fair,
accurate, and appropriate for use.

8.Question
How can human expert evaluations be applied in model
assessment?

Scan to Download
Answer:In scenarios requiring nuanced understanding, such
as medical image diagnoses, evaluating model predictions
against a consensus from multiple human experts can provide
a standard to measure the model's performance.

9.Question
What is meant by 'counterfactual analysis' in the context
of model explanations?
Answer:Counterfactual analysis involves finding instances
within a dataset that are similar but resulted in different
predictions, helping to analyze how small changes in features
affect model outcomes.

10.Question
Can you summarize the role of the Fairness Lens design
pattern?
Answer:The Fairness Lens provides techniques and
frameworks for assessing model predictions and ensuring
that these predictions are equitable across various
demographics, aiming to mitigate bias throughout the
modeling process.

Scan to Download
Chapter 8 | Connected Patterns| Q&A
1.Question
What are the main stages of the machine learning (ML)
life cycle and their importance?
Answer:The ML life cycle consists of three main
stages: discovery, development, and deployment.
Each stage is crucial; discovery focuses on aligning
business goals with ML applications, development
involves creating and refining models with proper
data handling and feature engineering, and
deployment is about integrating models into
production to solve real business problems
effectively.

2.Question
How do design patterns in machine learning relate to
each other?
Answer:Design patterns in machine learning are
interconnected; they often complement one another. For
example, the Hashed Feature design pattern can work

Scan to Download
alongside the Embeddings pattern for handling
high-cardinality categorical features, while the Bridge
Schema pattern can cooperate with Transfer Learning to
manage changes in data schema.

3.Question
What is the role of feature engineering in the
development phase of the ML lifecycle?
Answer:Feature engineering is crucial as it involves
transforming raw input data into effective features that align
closely with the model’s objectives. It includes techniques
for handling missing values, creating categorical features,
and applying hashing or embedding methods to enhance
model performance.

4.Question
How can organizations assess their AI readiness and what
are the three phases involved?
Answer:Organizations can assess their AI readiness by
identifying which of the three phases they are in: tactical
(manual development on narrow use cases), strategic

Scan to Download
(utilizing automated pipelines with alignment to business
goals), and transformational (fully automated processes
leading to innovation and broader AI application across all
departments).

5.Question
Why is monitoring and maintaining a deployed model
important?
Answer:Monitoring a deployed machine learning model is
essential because it ensures the model's performance stays
relevant over time. Data and environmental changes can lead
to model staleness, necessitating routine checks to maintain
accuracy and effectiveness in real-world applications.

6.Question
What considerations should be made when designing a
machine learning workflow?
Answer:When designing a machine learning workflow,
organizations should consider aspects such as data quality
and accessibility, the iteration nature of the workflow stages,
the impact of decisions made during the discovery stage, and

Scan to Download
the need for robust pipelines for data handling and model
training.

7.Question
What challenges do machine learning teams face in the
deployment stage?
Answer:In the deployment stage, teams face challenges such
as integrating ML models into existing systems, managing
legacy infrastructures, ensuring model retraining is efficient,
and mitigating issues related to network connectivity or data
streaming.

8.Question
How important is collaboration between business experts
and ML practitioners during the discovery phase?
Answer:Collaboration is crucial; without input from both
business experts and ML practitioners, there is a risk that the
project does not align with business goals or fails to address
important technical challenges, which could compromise the
efficacy of the ML solution.

9.Question
What role does hyperparameter tuning play in machine

Scan to Download
learning?
Answer:Hyperparameter tuning is essential for optimizing a
model's performance. It involves systematically finding the
best configuration of parameters that control the learning
process, ultimately improving model accuracy and
efficiency.

10.Question
How can explainable AI techniques enhance trust in
machine learning systems?
Answer:Explainable AI techniques help demystify how
models arrive at predictions, aiding users in understanding
and trusting the model's decisions, particularly in regulated
industries where compliance and accountability are
paramount.

Scan to Download
Machine Learning Design Patterns Quiz
and Test
Check the Correct Answer on Bookey Website

Chapter 1 | The Need for Machine Learning Design

Patterns| Quiz and Test
1.Design patterns in machine learning are derived
solely from software engineering principles and
have no relation to architecture.
2.The book 'Machine Learning Design Patterns' serves as a
reference for common machine learning problems and their
effective solutions.
3.In machine learning, the primary focus is on unsupervised
learning rather than supervised learning.
Chapter 2 | Data Representation Design Patterns|
Quiz and Test
1.Machine learning models require effective data
representation for mathematical functions to work
with real-world data.
2.Embeddings are simple representations that do not capture

Scan to Download
 the relationships between data points.
3.Feature engineering involves creating features from inputs
to improve model performance.
Chapter 3 | Problem Representation Design
Patterns| Quiz and Test
1.The Reframing Design Pattern helps in improving
predictions by capturing probability distributions
instead of single metrics.
2.The Multilabel Design Pattern applies softmax activation
for the final output to create multi-hot encoded labels.
3.The Cascade Design Pattern is used to handle complex
problems by breaking them into smaller tasks, but errors
can propagate through the cascade if not designed
carefully.

Scan to Download
Chapter 4 | Model Training Patterns| Quiz and Test
1.Training machine learning models involves a
process called the training loop.
2.Using checkpoints during model training is unnecessary as
it does not provide any benefits during interruptions.
3.Transfer learning can help improve efficiency when
working with small datasets by reusing parts of pre-trained
models.
Chapter 5 | Design Patterns for Resilient Serving|
Quiz and Test
1.The Stateless Serving Function design pattern
enables production ML systems to handle
thousands or millions of prediction requests per
second.
2.Batch Serving is designed for real-time predictions and
should not be used for non-time-sensitive use cases.
3.Keyed Predictions helps manage outputs from batch
requests and facilitates scalability by tracking predictions
more effectively.

Scan to Download
Chapter 6 | Reproducibility Design Patterns| Quiz
and Test
1.The Transform Design Pattern aims to maintain a
clear separation between raw inputs and
transformed features in order to ensure consistent
feature creation during training and serving.
2.The Workflow Pipeline Design Pattern does not consider
the containerization of workflow steps, making it less
reproducible.
3.Model versioning only allows one model version to be used
at any given time in order to maintain consistency across
different applications.

Scan to Download
Chapter 7 | Responsible AI| Quiz and Test
1.Machine learning models have no impact on
stakeholders beyond the development teams.
2.The Explainable Predictions pattern helps enhance user
trust in machine learning models.
3.The Fairness Lens pattern addresses biases arising only
during the model deployment phase.
Chapter 8 | Connected Patterns| Quiz and Test
1.The chapter provides a reference catalog of
machine learning design patterns that help solve
recurring problems in the design, training, and
deployment of machine learning models.
2.The patterns discussed in the chapter do not have any
interrelations or interactions with each other.
3.The ML life cycle phases outlined in the chapter include
discovery, development, and deployment, each building on
the previous one.

Scan to Download