Machine Learning Q and AI 1686653642

Machine Learning Q and AI
Expand Your Machine Learning & AI

Knowledge With 30 In-Depth Questions
and Answers
Sebastian Raschka, PhD

This book is for sale at
http://leanpub.com/machine-learning-q-and-ai
This version was published on 2023-06-11
This is a Leanpub book. Leanpub empowers authors and

publishers with the Lean Publishing process. Lean Publishing is
the act of publishing an in-progress ebook using lightweight tools
and many iterations to get reader feedback, pivot until you have
the right book and build traction once you do.
© 2016 - 2023 Sebastian Raschka, PhD

This book is dedicated to those who tirelessly contribute to advanc-
ing the field of machine learning through research and development.
Your passion for discovery and innovation and your commitment to
sharing knowledge and resources through the open-source commu-
nity is an inspiration to us all.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Who Is This Book For? . . . . . . . . . . . . . . . . . . . . 2
What Will You Get Out of This Book? . . . . . . . . . . . 3
How To Read This Book . . . . . . . . . . . . . . . . . . . 4
Sharing Feedback and Supporting This Book . . . . . . . 6
Acknowledgements . . . . . . . . . . . . . . . . . . . . . 7
About the Author . . . . . . . . . . . . . . . . . . . . . . . 8
Copyright and Disclaimer . . . . . . . . . . . . . . . . . . 9
Credits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Chapter 1. Neural Networks and Deep Learning . . . . . . 12

Q1. Embeddings, Representations, and Latent Space . . . 13
Q2. Self-Supervised Learning . . . . . . . . . . . . . . . . 14
Q3. Few-Shot Learning . . . . . . . . . . . . . . . . . . . 15
Q4. The Lottery Ticket Hypothesis . . . . . . . . . . . . . 16
Q5. Reducing Overfitting with Data . . . . . . . . . . . . 17
Q6. Reducing Overfitting with Model Modifications . . . 18
Q7. Multi-GPU Training Paradigms . . . . . . . . . . . . 19
Q8. The Keys to Success of Transformers . . . . . . . . . 26
Q9. Generative AI Models . . . . . . . . . . . . . . . . . . 27
Q10. Sources of Randomness . . . . . . . . . . . . . . . . 28
Chapter 2. Computer Vision . . . . . . . . . . . . . . . . . . 29

Q11. Calculating the Number of Parameters . . . . . . . 30
CONTENTS
Q12. The Equivalence of Fully Connected and Convolu-

tional Layers . . . . . . . . . . . . . . . . . . . . . 31
Q13. Large Training Sets for Vision Transformers . . . . 32
Chapter 3. Natural Language Processing . . . . . . . . . . 33

Q14. The Distributional Hypothesis . . . . . . . . . . . . 34
Q15. Data Augmentation for Text . . . . . . . . . . . . . 35
Q16. “Self”-Attention . . . . . . . . . . . . . . . . . . . . . 36
Q17. Encoder- And Decoder-Style Transformers . . . . . 37
Q18. Using and Finetuning Pretrained Transformers . . . 47
Q19. Evaluating Generative Language Models . . . . . . 48
Chapter 4. Production, Real-World, And Deployment

Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Q20. Stateless And Stateful Training . . . . . . . . . . . . 50
Q21. Data-Centric AI . . . . . . . . . . . . . . . . . . . . . 51
Q22. Speeding Up Inference . . . . . . . . . . . . . . . . . 52
Q23. Data Distribution Shifts . . . . . . . . . . . . . . . . 53
Chapter 5. Predictive Performance and Model Evaluation 54

Q24. Poisson and Ordinal Regression . . . . . . . . . . . 55
Q25. Confidence Intervals . . . . . . . . . . . . . . . . . . 56
Q26. Confidence Intervals Versus Conformal Predictions 57
Q27. Proper Metrics . . . . . . . . . . . . . . . . . . . . . 58
Q28. The K in K-Fold Cross-Validation . . . . . . . . . . 59
Q29. Training and Test Set Discordance . . . . . . . . . . 60
Q30. Limited Labeled Data . . . . . . . . . . . . . . . . . 61
Afterword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Appendix A: Reader Quiz Solutions . . . . . . . . . . . . . 76
Appendix B: List of Questions . . . . . . . . . . . . . . . . . 77

Preface
Over the years, I shared countless educational nuggets about ma-
chine learning and deep learning with hundreds of thousands
of people. The positive feedback has been overwhelming, and I
continue to receive requests for more. So, in this book, I want to
indulge both your desire to learn and my passion for writing about
machine learning¹.
¹I will use machine learning as an umbrella term for machine learning, deep learning, and
artificial intelligence.
Who Is This Book For? 2
Who Is This Book For?

This book is for people with a beginner or intermediate background
in machine learning who want to learn something new. This book
will expose you to new concepts and ideas if you are already
familiar with machine learning. However, it is not a math or coding
book. You won’t need to solve any proofs or run any code while
reading. In other words, this book is a perfect travel companion or
something you can read on your favorite reading chair with your
morning coffee.
What Will You Get Out of This Book? 3
What Will You Get Out of This Book?

Machine learning and AI are moving at a rapid pace. Researchers
and practitioners are constantly struggling to keep up with the
breadth of concepts and techniques. This book provides bite-sized
nuggets for your journey from machine learning beginner to ex-
pert, covering topics from various machine learning areas. Even
experienced machine learning researchers and practitioners will
encounter something new that they can add to their arsenal of
techniques.
How To Read This Book 4
How To Read This Book

The questions in this book are mainly independent, and you can
read them in any desired order. You can also skip individual
questions for the most part. However, I organized the questions to
bring more structure to the book.
For instance, the first question deals with embeddings, which we
refer to in later questions on self-supervised learning and few-
shot learning. Therefore, I recommend reading the questions in
sequence.
The book is structured into five main chapters to provide additional
structure. However, many questions could appear in different
chapters without affecting the flow.
Chapter 1, Deep Learning and Neural Networks covers questions
about deep neural networks and deep learning that are not specific
to a particular subdomain. For example, we discuss alternatives to
supervised learning and techniques for reducing overfitting.
Chapter 2, Computer Vision focuses on topics mainly related to
deep learning but are specific to computer vision, many of which
cover convolutional neural networks and vision transformers.
Chapter 3, Natural Language Processing covers topics around
working with text, many of which are related to transformer
architectures and self-attention.
Chapter 4, Production, Real-World, And Deployment Scenarios con-
tains questions pertaining to practical scenarios, such as increasing
inference speeds and various types of distribution shifts.
Chapter 5, Predictive Performance and Model Evaluation dives a
bit deeper into various aspects of squeezing out predictive perfor-
mance, for example, changing the loss function, setting up k-fold
cross-validation, and dealing with limited labeled data.
If you are not reading this book for entertainment but for machine
How To Read This Book 5
learning interview preparation, you may prefer a spoiler-free look

at the questions to quiz yourself before reading the answers. In this
case, you can find a list of all questions, without answers, in the
appendix.
Sharing Feedback and Supporting This Book 6
Sharing Feedback and Supporting

This Book
I enjoy writing, and it is my pleasure to share this knowledge with
you.
If you obtained a free copy and like this book, you can
support me by buying a digital copy on Leanpub at
https://leanpub.com/machine-learning-q-and-ai/².
For an author, there is nothing more valuable than your honest
feedback. I would really appreciate hearing from you and appreci-
ate any reviews! And, of course, I would be more than happy if you
recommend this book to your friends and colleagues or share some
nice words on your social channels.
Moreover, I value feedback of every type, both positive and neg-
ative. Positive feedback serves as motivation, while constructive
criticism provides insights to learn and grow. If you have questions,
please don’t hesitate to reach out, preferably on the Discussions
page on GitHub³.
²https://leanpub.com/machine-learning-q-and-ai/
³https://github.com/rasbt/MachineLearning-QandAI-book/discussions
Acknowledgements 7
Acknowledgements
Writing a book is an enormous undertaking. This project would not
have been possible without the help of the open source and machine
learning communities who collectively created the technologies
that this book is about. Moreover, I want to thank everyone who
encouraged me to share my flashcard decks, as this book is an
improved and polished version of these.
I also want to thank the following readers for helpful feedback on
the manuscript:
• Anton Reshetnikov for suggesting a cleaner layout for the

supervised learning flowchart in Q30.
• Juan M. Bello-Rivas for sharing various typographical errors.
About the Author 8
About the Author
Sebastian Raschka is a machine learning

and AI researcher with a strong passion
for education. As Lead AI Educator at
Lightning AI, he is excited about making
AI and deep learning more accessible
and teaching people how to utilize these
technologies at scale.
Before dedicating his time fully to Light-
ning AI, Sebastian held a position as
Assistant Professor of Statistics at the University of Wisconsin-
Madison, where he specialized in researching deep learning and
machine learning. You can find out more about his research on his
website⁴.
Moreover, Sebastian loves open-source software and has been
a passionate contributor for over a decade. Next to coding, he
also loves writing and authored the bestselling Python Machine
Learning book and Machine Learning with PyTorch and Scikit-
Learn.
If you like to find out more about Sebastian and what
he is currently up to, please visit his personal website at
https://sebastianraschka.com. You can also find Sebastian on
Twitter (@rasbt)⁵ and LinkedIn (sebastianraschka)⁶.
⁴https://sebastianraschka.com/publications/
⁵https://twitter.com/rasbt
⁶https://www.linkedin.com/in/sebastianraschka
Copyright and Disclaimer 9
Copyright and Disclaimer

Machine Learning Q and AI by Sebastian Raschka
Copyright © 2023 Sebastian Raschka. All rights reserved.
No part of this book may be reproduced or transmitted in any form
or by any means, electronic or mechanical, including photocopying,
recording or by any information storage and retrieval system,
without written permission from the author.
The information contained within this book is strictly for educa-
tional purposes. If you wish to apply ideas contained in this book,
you are taking full responsibility for your actions. The author has
made every effort to ensure the accuracy of the information within
this book was correct at time of publication. The author does not
assume and hereby disclaims any liability to any party for any loss,
damage, or disruption caused by errors or omissions, whether such
errors or omissions result from accident, negligence, or any other
cause.
Credits 10
Credits
Cover image by ECrafts / stock.adobe.com.
Introduction
Thanks to rapid advancements in deep learning, we have seen a
significant expansion of machine learning and AI in recent years.
On the one hand, this rapid progress is exciting if we expect these
advancements to create new industries, transform existing ones,
and improve the quality of life for people around the world.
On the other hand, the rapid emergence of new techniques can
make it challenging to keep up, and keeping up can be a very time-
consuming process. Nonetheless, staying current with the latest
developments in AI and deep learning is essential for professionals
and organizations that use these technologies.
With this in mind, I began writing this book in the summer of 2022
as a resource for readers and machine learning practitioners who
want to advance their understanding and learn about useful tech-
niques that I consider significant and relevant but often overlooked
in traditional and introductory textbooks and classes.
I hope readers will find this book a valuable resource for obtaining
new insights and discovering new techniques they can implement
in their work.
Happy learning,
Sebastian
Chapter 1. Neural
Networks and Deep
Learning
Q1. Embeddings, Representations, and Latent Space 13
Q1. Embeddings, Representations,

and Latent Space
This content is not available in the sample book. The book can be
purchased on Leanpub at http://leanpub.com/machine-learning-q-
and-ai.
Q2. Self-Supervised Learning 14
Q2. Self-Supervised Learning

and-ai.
Q3. Few-Shot Learning 15
Q3. Few-Shot Learning

and-ai.
Q4. The Lottery Ticket Hypothesis 16
Q4. The Lottery Ticket Hypothesis

and-ai.
Q5. Reducing Overfitting with Data 17
Q5. Reducing Overfitting with Data

and-ai.
Q6. Reducing Overfitting with Model Modifications 18
Q6. Reducing Overfitting with Model

Modifications
and-ai.
Q7. Multi-GPU Training Paradigms 19
Q7. Multi-GPU Training Paradigms

> Q:
What are the different multi-GPU training paradigms, and what
are their respective advantages and disadvantages?
> A:
The multi-GPU training paradigms can be categorized into two
groups: (1) dividing data for parallel processing with multiple GPUs
and (2) dividing the model among multiple GPUs to handle memory
constraints when the model size surpasses that of a single GPU.
Data parallelism falls into the first category, model parallelism and
tensor parallelism fall into the second category, and techniques like
pipeline parallelism borrow ideas from both categories. In addition,
current software implementations such as DeepSpeed⁷, Colossal
AI⁸, and others blend multiple approaches into a hybrid technique.
Model parallelism
Model parallelism (also referred to as inter op parallelism) is
perhaps the most intuitive form of parallelization across devices.
For example, suppose you have a simple neural network that only
consists of 2 layers: a hidden layer and an output layer. Here, we
keep one layer on one GPU and the other layer on another GPU. Of
course, this can scale to an arbitrary number of layers and GPUs.
This is a good strategy for dealing with limited GPU memory where
the complete network does not fit into one GPU. However, there are
more efficient ways of using multiple GPUs because of the chain-
like structure (layer 1 on GPU 1 → layer 2 on GPU 2 → …).
A major disadvantage of model parallelism is that the GPUs have to
wait for each other – they cannot efficiently work parallel as they
depend on each other’s outputs.
⁷https://github.com/microsoft/DeepSpeed
⁸https://github.com/hpcaitech/ColossalAI
Data parallelism
Data parallelism has been the default mode for multi-GPU train-
ing for several years. Here, we divide a minibatch into smaller
microbatches. Then, each GPU processes a microbatch separately
to compute the loss and loss gradients for the model weights. After
the individual devices process the microbatches, the gradients are
combined to compute the weight update for the next round.
An advantage of data parallelism over model parallelism is that the
GPUs can run in parallel – each GPU processes a portion of the
training minibatch, a microbatch. However, a caveat is that each
GPU requires a full copy of the model. This is obviously not feasible
if we have large models that don’t fit into the GPU’s VRAM.
Tensor parallelism
Tensor parallelism (also referred to as intra op parallelism) is a more
efficient form of model parallelism (inter op parallelism). Here,
the weight and activation matrices are spread across the devices
instead of distributing whole layers across devices. Specifically,
the individual matrices are split, so we split an individual matrix
multiplication across GPUs.
Figure 7.1. Comparison of model parallelism (left), data parallelism (center),

and tensor parallelism (right). In model parallelism, we put different layers
onto different GPUs to work around GPU memory limitations. In data par-
allelism, we split batch across GPUs to train copies of the model in parallel,
averaging gradients for the weight update afterwards. In tensor parallelism, we
split matrices (inputs and weights) across different GPU for parallel processing
when models are too large to fit into GPU memory.
There are several ways we can implement tensor parallelism. For

example, using basic principles of linear algebra, we can split a
matrix multiplication across two GPUs in a row- or column-wise
fashion, as illustrated in the figure below. (Note that this concept
can be extended to an arbitrary number of GPUs.)
Figure 7.2. Illustration how we can distribute matrix multiplication across

different devices. For simplicity, the figure depicts two GPUs, but the concepts
extends to an arbitrary number of devices.
An advantage of tensor parallelism is that we can work around

memory limitations similar to model parallelism. And at the same
time, we can also execute operations in parallel, similar to data
parallelism.
A small weakness of tensor parallelism is that it can result in high
communication overhead between the multiple GPUs across which
the matrices are split or sharded. For instance, tensor parallelism
requires frequent synchronization of the model parameters across
the devices, which can slow down the overall training process.
Pipeline parallelism
Pipeline parallelism can be seen as a form of model parallelism

that tries to minimize the sequential-computation bottleneck. In
that sense, we can think of pipeline parallelism as a form of model
parallelism that enhances the parallelism between the individual
layers sitting on different devices. However, note that it also
borrows ideas from data parallelism, such as splitting minibatches
further into microbatches.
How does it work? During the forward pass, the activations are
passed similar to model parallelism, however, the twist is that
the gradients of the input tensor are passed backwards to prevent
the devices from being idle. In a sense, pipeline parallelism is a
sophisticated hybrid between data and model parallelism, which is
described in more detail in the GPipe paper⁹ or DeepSpeed pipeline
parallelism tutorial¹⁰.
Figure 7.3. A conceptual illustration of pipeline parallelism, which aims to

reduce the idle time of GPUs compared to regular model parallelism.
A disadvantage of pipeline parallelism is that it may require sig-

nificant effort to design and implement the pipeline stages and
associated communication patterns. Additionally, the performance
gains from pipeline parallelism may not be as substantial as those
from other parallelization techniques, such as pure data parallelism,
especially for small models or in cases where the communication
⁹Huang, Cheng, Bapna, Firat, Chen, Chen, Lee, Ngiam, Le, Wu, and Chen
(2018). GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism,
https://arxiv.org/abs/1811.06965.
¹⁰Pipeline Parallelism: https://www.deepspeed.ai/tutorials/pipeline/
overhead is high.
Pipeline parallelism is definitely an improvement over model paral-
lelism, even though it is not perfect, and there will be idle bubbles.
However, for modern architectures that are too large to fit into
GPU memory, it is nowadays more common to use a blend of data
parallelism and tensor parallelism (as opposed to model parallelism)
techniques.
Sequence parallelism
Sequence parallelism is a new concept developed for transformer
models¹¹. One shortcoming of transformers is that the self-
attention mechanism (the original scaled-dot product attention)
scales quadratically with the input sequence length. There are,
of course, more efficient alternatives to the original attention
mechanism that scales linearly¹²¹³; however, they are less popular,
and most people prefer the original scaled-dot product attention
mechanism.
Sequence parallelism splits the input sequence into smaller chunks
that can be distributed across GPUs to work around memory
limitations as illustrated in the figure below.
¹¹Li, Xue, Baranwal, Li, and You (2021). Sequence Parallelism: Long Sequence Training from
[a] System[s] Perspective, https://arxiv.org/abs/2105.13120.
¹²Tay, Dehghani, Bahri, and Metzler (2020). Efficient Transformers: A Survey,
¹³Zhuang, Liu, Pan, He, Weng, and Shen (2023). A Survey on Efficient Training of Trans-
formers, https://arxiv.org/abs/2302.01107.
Figure 7.4. A conceptual illustration of sequence parallelism, which aims to

reduce computation memory constraints of self-attention mechanisms.
Sequence parallelism is less well-studied than other parallelization

techniques. However, conceptually, it would have similar advan-
tages and disadvantages as tensor parallelism. It enables us to train
larger models when memory is a constraint due to the sequence
length; however, it also introduces additional communication over-
heads. On the other hand, it also shares shortcoming of data
parallelism where we have to duplicate the model and make sure it
fits into the device memory.
Another disadvantage of sequence parallelism (depending on the
implementation) for multi-GPU training of transformers is that
breaking up the input sequence into smaller subsequences can
decrease the model’s accuracy – mainly when the model is applied
to longer sequences.
Which techniques should we use in practice?
Practical recommendations depend on the context. If we train small
models that fit onto a single GPU, then data parallelism strategies
may be the most efficient. Performance gains from pipeline paral-
lelism may not be as significant as those from other parallelization
techniques such as data parallelism, especially for small models or
in cases where the communication overhead is high.
If models are too large to fit into the memory of a single GPU, we
need to explore model or tensor parallelism. Tensor parallelism is

naturally more efficient since the GPUs can work in parallel – there
is no sequential dependency as in model parallelism.
Modern multi-GPU strategies typically combine data parallelism
and tensor parallelism (popular examples include DeepSpeed stages
2 and 3¹⁴).
> Reader quiz:
7-A
Suppose we are implementing our own version of tensor paral-
lelism, which works great when we train our model with an SGD
(standard stochastic gradient descent) optimizer.
However, when we try the Adam¹⁵ optimizer, we encounter an out-
of-memory device. What could be a potential problem explaining
this issue?
7-B
Suppose we don’t have access to a GPU and are considering using
data parallelism on the CPU. Is this a good idea?
¹⁴https://www.deepspeed.ai/tutorials/zero/
¹⁵Kingma and Ba (2014). Adam: A Method for Stochastic Optimization,
Q8. The Keys to Success of Transformers 26
Q8. The Keys to Success of

Transformers
and-ai.
Q9. Generative AI Models 27
Q9. Generative AI Models

and-ai.
Q10. Sources of Randomness 28
Q10. Sources of Randomness

and-ai.
Chapter 2. Computer
Vision
and-ai.
Q11. Calculating the Number of Parameters 30
Q11. Calculating the Number of

Parameters
and-ai.
Q12. The Equivalence of Fully Connected and Convolutional
Layers 31
Q12. The Equivalence of Fully

Connected and Convolutional Layers
and-ai.
Q13. Large Training Sets for Vision Transformers 32
Q13. Large Training Sets for Vision

Transformers
and-ai.
Chapter 3. Natural
Language Processing
and-ai.
Q14. The Distributional Hypothesis 34
Q14. The Distributional Hypothesis

and-ai.
Q15. Data Augmentation for Text 35
Q15. Data Augmentation for Text

and-ai.
Q16. “Self”-Attention 36
Q16. “Self”-Attention
and-ai.
Q17. Encoder- And Decoder-Style Transformers 37
Q17. Encoder- And Decoder-Style

Transformers
> Q:
What are the differences between encoder- and decoder-based
language transformers?
> A:
Fundamentally, both encoder- and decoder-style architectures use
the same self-attention layers to encode word tokens. However, the
main difference is that encoders are designed to learn embeddings
that can be used for various predictive modeling tasks such as
classification. In contrast, decoders are designed to e new texts, for
example, answering user queries.
The original transformer
The original transformer architecture¹⁶, which was developed for
English-to-French and English-to-German language translation,
utilized both an encoder and a decoder, as illustrated in the figure
below.
¹⁶Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin (2017). Atten-
tion Is All You Need, https://arxiv.org/abs/1706.03762
Figure 17.1. Illustration of the original transformer architecture.
In the figure above, the input text (that is, the sentences of the
text that is to be translated) is first tokenized into individual word
tokens, which are then encoded via an embedding layer¹⁷ before
it enters the decoder part. Then, after adding a positional encoding
vector to each embedded word, the embeddings go through a multi-
head self-attention layer. The multi-head attention layer is followed
by an “Add & normalize” step, which performs a layer normaliza-
tion and adds the original embeddings via a skip connection (also
known as a residual or shortcut connection). Finally, after entering
a “fully connected layer,” which is a small multilayer perceptron
¹⁷See Q1 for more details about embeddings.
consisting of two fully connected layers with a nonlinear activation

function in between, the outputs are again added and normalized
before they are passed to a multi-head self-attention layer of the
decoder part.
The decoder part in the figure above has a similar overall structure
as the encoder part. The key difference is that the inputs and outputs
are different. The encoder receives the input text that is to be
translated, and the decoder generates the translated text.
Encoders
The encoder part in the original transformer, illustrated in the
preceding figure, is responsible for understanding and extracting
the relevant information from the input text. It then outputs a con-
tinuous representation (embedding) of the input text that is passed
to the decoder. Finally, the decoder generates the translated text
(target language) based on the continuous representation received
from the encoder.
Over the years, various encoder-only architectures have been
developed based on the encoder module of the original transformer
model outlined above. Notable examples include BERT¹⁸ and
RoBERTa¹⁹.
BERT (Bidirectional Encoder Representations from Transformers)
is an encoder-only architecture based on the Transformer’s encoder
module. The BERT model is pretrained on a large text corpus using
masked language modeling (illustrated in the figure below) and
next-sentence prediction tasks.
¹⁸BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
(2018). Devlin, Chang, Lee, and Toutanova, https://arxiv.org/abs/1810.04805.
¹⁹RoBERTa: A Robustly Optimized BERT Pretraining Approach (2018). Liu, Ott, Goyal, Du,
Joshi, Chen, Levy, Lewis, Zettlemoyer, and Stoyanov https://arxiv.org/abs/1907.11692.
Figure 17.2. Illustration of the masked language modeling pretraining objective

used in BERT-style transformers.
The main idea behind masked language modeling is to mask (or

replace) random word tokens in the input sequence and then train
the model to predict the original masked tokens based on the
surrounding context.
Next to the masked language modeling pretraining task illustrated
in the figure above, the next-sentence prediction task asks the
model to predict whether the original document’s sentence order
of two randomly shuffled sentences is correct. For example, two
sentences, in random order, are separated by the [SEP] token:
• [CLS] Toast is a simple yet delicious food [SEP] It’s often

served with butter, jam, or honey.
• [CLS] It’s often served with butter, jam, or honey. [SEP] Toast
is a simple yet delicious food.
The [CLS] token is a placeholder token for the model, prompting

the model to return a True or False label indicating whether the
sentences are in the correct order or not.
The masked language and next-sentence pretraining objective²⁰

allow BERT to learn rich contextual representations of the input
texts, which can then be finetuned for various downstream tasks
like sentiment analysis, question-answering, and named entity
recognition.
RoBERTa (Robustly optimized BERT approach) is an optimized
version of BERT. It maintains the same overall architecture as BERT
but employs several training and optimization improvements, such
as larger batch sizes, more training data, and eliminating the
next-sentence prediction task. These changes resulted in RoBERTa
achieving better performance on various natural language under-
standing tasks than BERT.
Decoders
Coming back to the original transformer architecture outlined at
the beginning of this section, the multi-head self-attention mech-
anism in the decoder is similar to the one in the encoder, but it
is masked to prevent the model from attending to future positions,
ensuring that the predictions for position i can depend only on the
known outputs at positions less than i. As illustrated in the figure
below, the decoder generates the output word by word.
²⁰It shall be noted that this pretraining is a form of self-supervised learning (see Q2 for more
details).
Figure 17.3. Illustration of the next-sentence prediction task used in the original
transformer.
This masking (shown explicitly in the figure above, although it

happens internally in the decoder’s multi-head self-attention mech-
anism) is essential to maintain the autoregressive property of the
transformer model during training and inference. The autoregres-
sive property ensures that the model generates output tokens one
at a time and uses previously generated tokens as context for
generating the next word token.
Over the years, researchers have built upon the original encoder-
decoder transformer architecture and developed several decoder-
only models that have proven to be highly effective in various
natural language processing tasks. The most notable models include
the GPT family.
The GPT (Generative Pre-trained Transformer) series are decoder-
only models pretrained on large-scale unsupervised text data and

finetuned for specific tasks such as text classification, sentiment
analysis, question-answering, and summarization. The GPT mod-
els, including GPT-2, GPT-3²¹, and the more recent GPT-4, have
shown remarkable performance in various benchmarks and are
currently the most popular architecture for natural language pro-
cessing.
One of the most notable aspects of GPT models is their emergent
properties. Emergent properties refer to the abilities and skills
that a model develops due to its next-word prediction pretraining.
Even though these models were only taught to predict the next
word, the pretrained models are capable of text summarization,
translation, summarization, question answering, classification, and
more. Furthermore, these models can perform new tasks without
updating the model parameters via in-context learning, which is
discussed in more detail in Q18.
Encoder-decoder hybrids
Next to the traditional encoder and decoder architectures, there
have been advancements in the development of new encoder-
decoder models that leverage the strengths of both components.
These models often incorporate novel techniques, pre-training ob-
jectives, or architectural modifications to enhance their perfor-
mance in various natural language processing tasks. Some notable
examples of these new encoder-decoder models include BART²²
and T5²³.
Encoder-decoder models are typically used for natural language
processing tasks that involve understanding input sequences and
²¹Brown et al. (2020). Language Models are Few-Shot Learners,
https://arxiv.org/abs/2005.14165
²²Lewis, Liu, Goyal, Ghazvininejad, Mohamed, Levy, Stoyanov, and Zettlemoyer (2018).
BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Trans-
lation, and Comprehension, https://arxiv.org/abs/1910.13461.
²³Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li, and Liu (2019).
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer,
generating output sequences, often with different lengths and struc-

tures. They are particularly good at tasks where there is a complex
mapping between the input and output sequences and where it is
crucial to capture the relationships between the elements in both
sequences. Some common use cases for encoder-decoder models
include text translation and summarization.
Terminology and jargon
All of these methods, encoder-only, decoder-only, and encoder-
decoder models, are sequence-to-sequence models (often abbrevi-
ated as seq2seq). Note that while we refer to BERT-style methods
as encoder-only, the description encoder-only may be misleading
since these methods also decode the embeddings into output tokens
or text during pretraining.
In other words, both encoder-only and decoder-only architectures
are “decoding.” However, the encoder-only architectures, in con-
trast to decoder-only and encoder-decoder architectures, are not
decoding in an autoregressive fashion. Autoregressive decoding
refers to generating output sequences one token at a time, condition-
ing each token on the previously generated tokens. Encoder-only
models do not generate coherent output sequences in this manner.
Instead, they focus on understanding the input text and producing
task-specific outputs, such as labels or token predictions.
Conclusion
In brief, encoder-style models are popular for learning embeddings
used in classification tasks, encoder-decoder-style models are used
in generative tasks where the output heavily relies on the input
(for example, translation and summarization), and decoder-only
models are used for other types of generative tasks including
Q&A. Since the first transformer architecture emerged, hundreds
of encoder-only, decoder-only, and encoder-decoder hybrids have
been developed, as summarized in the figure below.
Figure 17.4. An overview of some of the most popular large language trans-
formers organized by architecture type and developers.
While encoder-only models gradually lost in popularity, decoder-

only models like GPT exploded in popularity thanks to break-
through in text generation via GPT-3, ChatGPT, and GPT-4. How-
ever, encoder-only models are still very useful for training predic-
tive models based on text embeddings versus generating texts.
> Reader quiz:

17-A
As discussed earlier, BERT-style encoder models are pretrained
using masked language modeling and next-sentence prediction
pretraining objectives. How could we adopt such a pretrained
model for a classification task, for example, predicting whether a
text has a positive or negative sentiment?
17-B
Can we finetune or a decoder-only model like GPT for classifica-
tion?
Q18. Using and Finetuning Pretrained Transformers 47
Q18. Using and Finetuning

Pretrained Transformers
and-ai.
Q19. Evaluating Generative Language Models 48
Q19. Evaluating Generative

Language Models
and-ai.
Chapter 4. Production,
Real-World, And
Deployment Scenarios
and-ai.
Q20. Stateless And Stateful Training 50
Q20. Stateless And Stateful Training

and-ai.
Q21. Data-Centric AI 51
Q21. Data-Centric AI
and-ai.
Q22. Speeding Up Inference 52
Q22. Speeding Up Inference

and-ai.
Q23. Data Distribution Shifts 53
Q23. Data Distribution Shifts

and-ai.
Chapter 5. Predictive
Performance and Model
Evaluation
and-ai.
Q24. Poisson and Ordinal Regression 55
Q24. Poisson and Ordinal Regression

and-ai.
Q25. Confidence Intervals 56
Q25. Confidence Intervals

and-ai.
Q26. Confidence Intervals Versus Conformal Predictions 57
Q26. Confidence Intervals Versus

Conformal Predictions
and-ai.
Q27. Proper Metrics 58
Q27. Proper Metrics

and-ai.
Q28. The K in K-Fold Cross-Validation 59
Q28. The K in K-Fold Cross-Validation

and-ai.
Q29. Training and Test Set Discordance 60
Q29. Training and Test Set

Discordance
and-ai.
Q30. Limited Labeled Data 61
Q30. Limited Labeled Data

> Q:
Suppose we plotted a learning curve²⁴ and found that the machine
learning model overfits and could benefit from more training data.
Name different approaches for dealing with limited labeled data in
supervised machine learning settings.
> A:
Next to collecting more data, there are several methods more or less
related to regular supervised learning that we can use in limited-
labeled data regimes.
1) Label more data
Collecting additional training examples is often the best way to
improve the performance of a model²⁵. However, this is often
not feasible in practice. Listed below are various alternative ap-
proaches.
2) Bootstrapping the data
Similar to the topics discussed in Q5, Reducing Overfitting with
Data, it can be helpful to “bootstrap” the data by generating
modified (augmented) or artificial (synthetic) training examples to
boost the performance of the predictive model.
Of course, improving the quality of data can also lead to improved
predictive performance of a model, as discussed in Q21, Data-
Centric AI.
3) Transfer learning
Transfer learning describes training a model on a general dataset
(e.g., ImageNet) and then finetuning the pretrained target dataset
(e.g., a specific dataset consisting of different bird species). Transfer
²⁴See Q5, Figure 14 for a refresher of what a learning curve looks like.
²⁵A learning curve is a good diagnostic to find out whether the model can benefit from more
data. See Figure 15 in Q5 for details.
learning is usually done in the context of deep learning, where

model weights can be updated. This is in contrast to tree-based
methods since most decision tree algorithms are nonparametric
models that do not support iterative training or parameter up-
dates²⁶.
Figure 30.1. Illustration of transfer learning.
4) Self-supervised learning
Similar to transfer learning, self-supervised learning, the model
is pretrained on a different task before it is finetuned to a target
task for which only limited data exists. However, in contrast to
transfer learning, self-supervised learning usually relies on label
information that can be directly and automatically extracted from
unlabeled data. Hence, self-supervised learning is also often called
unsupervised pretraining. Common examples include “next word”
(e.g., used in GPT) or “masked word” (e.g., used in BERT) prediction
in language modeling. Or, an intuitive example from computer
vision includes inpainting: predicting the missing part of an im-
age that was randomly removed. (For more details about self-
supervised learning, also see Q2.)
²⁶While decision trees for incremental learning are not commonly
implemented, algorithms for training decision trees in an iterative fashion do exist
(https://en.wikipedia.org/wiki/Incremental_decision_tree).
Figure 30.2. Illustration of inpainting for self-supervised learning.
5) Active learning
In active learning, we typically involve manual labelers or users for
feedback during the learning process. However, instead of labeling
the entire dataset upfront, active learning includes a prioritization
scheme for suggesting unlabeled data points for labeling that
maximize the machine learning model’s performance.
The name active learning refers to the fact that the model is actively
selecting data for labeling in this process. For example, the simplest
form of active learning selects data points with high prediction
uncertainty for labeling by a human annotator (also referred to as
an oracle).
Figure 30.3. Illustration active learning.
6) Few-shot learning
In a few-shot learning scenario, we often deal with extremely small

datasets where we usually only have a handful of examples per
class. In research contexts, 1-shot (1 example per class) and 5-shot
(5 examples per class) are very common.
An extreme case of few-shot learning is zero-shot learning, where
no labels are provided. A recently popular example of zero-shot
learning is GPT-3 and related language models. Here, the user has
to provide all the necessary information via the input prompt, as
illustrated in the figure below.
Figure 30.4. Example of zero-shot classification with ChatGPT.
(For more details about self-supervised learning, also see Q3.)

7) Meta-learning
We can think of meta-learning as “learning to learn” – we develop
methods that learn how machine learning algorithms can best learn
from data.
Over the years, the machine learning community developed several
approaches for meta-learning. To further complicate matters, meta-
learning can refer to different processes.
Meta-learning is one of the main subcategories of few-shot learning
(mentioned above). Here, the focus is on learning a good feature

extraction module. The feature extraction module converts support
and query images into vector representations. These vector repre-
sentations are optimized for determining the predicted class of the
query example via comparisons with the training examples in the
support set. (This form of meta-learning is illustrated in Q3, Figure
13.)
Another branch of meta-learning, unrelated to the few-shot learn-
ing approach above, is focused on extracting meta-data (also called
meta-features) from datasets for supervised learning tasks. The
meta-features are descriptions of the dataset itself. For example,
these can include the number of features and statistics of the
different features (kurtosis, range, mean, etc.).
The extracted meta-features provide information for selecting a ma-
chine learning algorithm for the given dataset at hand. Using this
approach, we can narrow down the algorithm and hyperparameter
search spaces, which helps reduce overfitting when the dataset is
small.
Figure 30.5. Illustration of the meta-learning process involving the extraction

of meta-data.
8) Weakly supervised learning

Weakly supervised learning is a procedure where we use an exter-

nal label source to generate labels for an unlabeled dataset. Often,
the labels created by a weakly supervised labeling function are
more noisy or inaccurate than those produced by a human or
domain expert; hence, the term weakly supervised.
Often, we can develop or adopt a rule-based classifier to create
the labels in weakly supervised learning – these rules usually only
cover a subset of the unlabeled dataset.
Figure 30.6. Illustration of weakly supervised learning.
Imagine the context of email spam classification as an example of

a rule-based approach for data labeling. In weak supervision, we
could design a rule-based classifier based on the keyword “SALE”
in the email subject header line to identify a subset of spam emails.
Note that while we may use this rule to label certain emails as spam-
positive, we should not apply this rule to label emails without SALE
as non-spam but leave those either unlabeled or apply a different
rule to these²⁷.
In short, weakly supervised learning is an approach for increasing
the number of labeled instances in the training set. Hence, other
techniques, such as semi-supervised, transfer, active, and zero-shot
learning, are fully compatible with weakly supervised learning.
9) Semi-supervised learning
Semi-supervised learning is closely related to weakly supervised
learning described above: we create labels for unlabeled instances
in the dataset. The main difference between weakly supervised and
semi-supervised learning is how we create the labels²⁸.
In weak supervision, we create labels using an external labeling
function that is often noisy, inaccurate or only covers a subset of the
data. In semi-supervision, we do not use an external label function
but leverage the structure of the data itself.
Figure 30.7. Illustration of semi-supervised learning.
In semi-supervised learning, we can, for example, label additional

data points based on the density of neighboring labeled data points,
as illustrated in the figure below.
While we can apply weak supervision to an entirely unlabeled
dataset, semi-supervised learning requires at least a portion of
²⁷There is a subcategory of weakly supervised learning referred to as PU-learning. In PU-
learnining, which is short for positive-unlabeled learning, we only label and learn from positive
examples.
²⁸Semi-supervised learning is sometimes referred to as a subcategory of weakly supervised
learning and vice versa.
the data to be labeled. In practice, it is possible first to apply

weak supervision to label a subset of the data and then use semi-
supervised learning to label instances that were not captured by
the labeling functions.
10) Self-training
Self-training is a category that falls somewhere between semi-
supervised learning and weakly supervised learning. In self-
training, we train a model or adopt an existing model to label the
dataset. This model is also referred to as a pseudo-labeler.
Since the model used in self-training does not guarantee accu-
rate labels, self-training is related to weakly supervised learning.
Moreover, while we use or adopt a machine learning model for
this pseudo-labeling, self-training is also related to semi-supervised
learning.
An example of self-training is knowledge distillation, discussed in
Q6.
11) Multi-task learning
Multi-task learning trains neural networks on multiple, ideally
related tasks. For example, suppose we are training a classifier to
detect spam emails; here, spam classification is the main task. In
multi-task learning, we can add one or more related tasks the model
has to solve. These additional tasks are also referred to as auxiliary
tasks. If the main task is email spam classification, an auxiliary task
could be classifying the email’s topic or language.
Typically, multi-task learning is implemented via multiple loss
functions that have to be optimized simultaneously – one loss
function for each task. The auxiliary tasks serve as an inductive
bias, guiding the model to prioritize hypotheses that can explain
multiple tasks. This approach often results in models that perform
better on unseen data²⁹.
²⁹Caruana (1997). Multi-task learning. Machine Learning. 28: 41–75.
https://doi.org/10.1023%2FA%3A1007379606734
There are two subcategories of multi-task learning: (1) multi-task

learning with hard parameter sharing and (2) multi-task learning
with soft parameter sharing³⁰.
Figure 30.8. Illustration of the two main types of multi-task learning. For
simplicity, the figure depicts only two tasks, but multitask learning can be used
for any number of tasks.
The figure above illustrates the difference between hard and soft
parameter sharing. In hard parameter sharing, only the output
layers are task-specific, while all tasks share the same hidden
layers and neural network backbone architecture. In contrast, soft
parameter sharing uses separate neural networks for each task, but
regularization techniques such as distance minimization between
parameter layers are applied to encourage similarity among the
networks.
12) Multi-modal learning
While multi-task learning involves training a model with multiple
tasks and loss functions, multi-modal learning focuses on incorpo-
rating multiple types of input data.
Common examples of multi-modal learning are architectures that
take both image and text data as input³¹. Depending on the task,
³⁰Ruder (2017). An Overview of Multi-Task Learning in Deep Neural Networks.
https://www.ruder.io/multi-task/.
³¹Multi-modal learning is not restricted two only two modalities and can be used for any
number of inputs modalities.
we may employ a matching loss that forces the embedding vectors

(Q1) between related images and text to be similar, as shown in the
figure below.
Figure 30.9. Illustration of multi-modal learning with a matching loss that

forces embeddings from different types of inputs to be similar.
The figure above shows image and text encoders as separate com-
ponents. The image encoder can be a convolutional backbone or
a vision transformer, and the language encoder can be a recurrent
neural network or language transformer. However, it’s common
nowadays to use a single transformer-based module that can simul-
taneously process image and text data³².
Optimizing a matching loss, as shown in the previous figure, can
be useful for learning embeddings that can be applied to various
tasks, such as image classification or summarization. However, it is
also possible to directly optimize the target loss, like classification
or regression, as the figure below illustrates.
³²For example, VideoBERT is a model that with a joint module that processes both video and
text for action classification and video captioning. Reference: Sun, Myers, Vondrick, Murphy,
Schmid (2019). VideoBERT: A Joint Model for Video and Language Representation Learning,
Figure 30.10. Illustration of multi-modal learning for optimizing a supervised

learning objective.
Intuitively, models that combine data from different modalities

generally perform better than uni-modal models because they can
leverage more information. Moreover, recent research suggests that
the key to the success of multi-modal learning is the improved
quality of the latent space representation³³.
13) Inductive biases
Choosing models with stronger inductive biases can help to lower
data requirements by making assumptions about the structure of
the data. For example, due to their inductive biases, convolutional
networks require less data than vision transformers as discussed in
Q13.
Which techniques should we use?
Now that we covered several techniques for lowering the data
requirements, which ones should we use?
Collecting more data and techniques such as data augmentation
and feature engineering are compatible with all the methods dis-
cussed above. Also, multi-task learning and multi-modal inputs can
be used with the other learning strategies outlined above. If the
³³Huang, Du, Xue, Chen, Zhao, Huang, (2021). What Makes Multi-Modal Learning Better
Than Single (Provably), https://arxiv.org/abs/2106.04538.
model suffers from overfitting, techniques from Q5 and Q6 should

also be included.
How about active learning, few-shot learning, transfer learning,
self-supervised learning, semi-supervised learning, and weakly su-
pervised learning? Which technique(s) to try highly depends on the
context, and the figure below provides an overview that can be used
for guidance.
Figure 30.11. Recommendations for choosing a supervised learning techniques.

The black boxes are not terminal nodes but arch back to Evaluate model
performance (the arrows were omitted to avoid visual clutter).
> Reader quiz:

30-A
Given the task of constructing a machine learning model that

utilizes images to detect manufacturing defects on the outer shells
of tablet devices similar to iPads, we have access to the following
data:
• Millions of images of various computing devices, including

smartphones, tablets, and computers, which are not labeled.
• Thousands of labeled pictures of smartphones depicting vari-
ous types of damage.
• Hundreds of labeled images specifically related to the target
task of detecting manufacturing defects on tablet devices.
How could we approach this problem using self-supervised learn-

ing or transfer learning?
30-B
In active learning, selecting difficult examples for human inspec-
tion and labeling is often based on confidence scores. Neural
networks can provide such scores by using the logistic sigmoid or
softmax function in the output layer to calculate class-membership
probabilities. However, it is widely recognized that deep neu-
ral networks exhibit overconfidence on out-of-distribution data³⁴,
rendering their use in active learning ineffective. So, what are
some other methods to obtain confidence scores using deep neural
networks for active learning?
³⁴Nguyen, Yosinski, Clune, (2015). Deep Neural Networks Are Easily Fooled: High Confidence
Predictions for Unrecognizable Image, https://arxiv.org/abs/1412.1897.
Afterword
and-ai.
Appendix A: Reader Quiz
Solutions
and-ai.
Appendix B: List of
Questions
and-ai.

Machine Learning Q and AI 1686653642

Uploaded by

Copyright:

Available Formats

Machine Learning Q and AI 1686653642

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning Q and AI 1686653642

Uploaded by

Copyright:

Available Formats

Machine Learning Q and AI

Expand Your Machine Learning & AI

Sebastian Raschka, PhD

This version was published on 2023-06-11

This is a Leanpub book. Leanpub empowers authors and

© 2016 - 2023 Sebastian Raschka, PhD

Chapter 1. Neural Networks and Deep Learning . . . . . . 12

Chapter 2. Computer Vision . . . . . . . . . . . . . . . . . . 29

Q12. The Equivalence of Fully Connected and Convolu-

Chapter 3. Natural Language Processing . . . . . . . . . . 33

Chapter 4. Production, Real-World, And Deployment

Chapter 5. Predictive Performance and Model Evaluation 54

Appendix A: Reader Quiz Solutions . . . . . . . . . . . . . 76

Appendix B: List of Questions . . . . . . . . . . . . . . . . . 77

Who Is This Book For?

What Will You Get Out of This Book?

How To Read This Book

learning interview preparation, you may prefer a spoiler-free look

Sharing Feedback and Supporting

• Anton Reshetnikov for suggesting a cleaner layout for the

About the Author

Sebastian Raschka is a machine learning

Copyright and Disclaimer

Q1. Embeddings, Representations,

Q2. Self-Supervised Learning

Q3. Few-Shot Learning

Q4. The Lottery Ticket Hypothesis

Q5. Reducing Overfitting with Data

Q6. Reducing Overfitting with Model

Q7. Multi-GPU Training Paradigms

Figure 7.1. Comparison of model parallelism (left), data parallelism (center),

There are several ways we can implement tensor parallelism. For

Figure 7.2. Illustration how we can distribute matrix multiplication across

An advantage of tensor parallelism is that we can work around

Pipeline parallelism can be seen as a form of model parallelism

Figure 7.3. A conceptual illustration of pipeline parallelism, which aims to

A disadvantage of pipeline parallelism is that it may require sig-

Figure 7.4. A conceptual illustration of sequence parallelism, which aims to

Sequence parallelism is less well-studied than other parallelization

need to explore model or tensor parallelism. Tensor parallelism is

Q8. The Keys to Success of

Q9. Generative AI Models

Q10. Sources of Randomness

Q11. Calculating the Number of

Q12. The Equivalence of Fully

Q13. Large Training Sets for Vision

Q14. The Distributional Hypothesis

Q15. Data Augmentation for Text

Q17. Encoder- And Decoder-Style

Figure 17.1. Illustration of the original transformer architecture.

consisting of two fully connected layers with a nonlinear activation

Figure 17.2. Illustration of the masked language modeling pretraining objective

The main idea behind masked language modeling is to mask (or

• [CLS] Toast is a simple yet delicious food [SEP] It’s often

The [CLS] token is a placeholder token for the model, prompting

The masked language and next-sentence pretraining objective²⁰

This masking (shown explicitly in the figure above, although it

only models pretrained on large-scale unsupervised text data and

generating output sequences, often with different lengths and struc-

While encoder-only models gradually lost in popularity, decoder-