Start free trial
Trending New Genres Authors
EnglishEnglish
EspañolSpanish
简体中文Chinese
繁體中文Chinese (Traditional)
FrançaisFrench
DeutschGerman
日本語Japanese
PortuguêsPortuguese
ItalianoItalian
한국어Korean
РусскийRussian
NederlandsDutch
العربيةArabic
PolskiPolish
हिन्दीHindi
Tiếng ViệtVietnamese
SvenskaSwedish
ΕλληνικάGreek
TürkçeTurkish
ไทยThai
ČeštinaCzech
RomânăRomanian
MagyarHungarian
УкраїнськаUkrainian
IndonesiaIndonesian
DanskDanish
SuomiFinnish
БългарскиBulgarian
עבריתHebrew
NorskNorwegian
HrvatskiCroatian
CatalàCatalan
SlovenčinaSlovak
LietuviųLithuanian
SlovenščinaSlovenian
СрпскиSerbian
EestiEstonian
LatviešuLatvian
فارسیPersian
മലയാളംMalayalam
தமிழ்Tamil
اردوUrdu
Searching...
SoBrief
Deep Learning

Deep Learning

Understand how deep learning works, where it excels, and when its confident predictions may mislead.
by John D. Kelleher 2019 296 pages
3.89
473 ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Use deep learning when raw data is vast and hand-built features fail.

Learn from raw inputs. Deep learning trains large neural networks to find useful patterns in complex data. Unlike traditional machine learning, which often depends on people designing informative features, deep learning can learn features from raw inputs such as image pixels, speech signals, or text. That ability is especially valuable when the data has many dimensions and examples are plentiful.

AlphaGo made the case. Go has an enormous number of possible board positions, making it far harder to search than chess. AlphaGo used neural networks to evaluate positions and choose moves, beating Lee Sedol in 2016. The book notes that computer Go progressed from advanced amateur strength in 2009 to world-champion level in just seven years. The leap illustrates what data-driven pattern recognition can do when paired with substantial computation.

Analysis

The important distinction is not that a neural network eliminates human judgment, but that it shifts judgment toward choosing the task, data, and evaluation criteria. That can make deep learning powerful in domains where useful patterns are difficult to specify in advance. It also makes the technology easy to overstate: success at a bounded task such as Go does not, by itself, demonstrate general intelligence. The achievement is remarkable precisely because the system is specialized, trained, and measured within a well-defined environment.

Think of every model as a learned input-to-output mapping, not a mind.

A model maps values. A function is a deterministic mapping from inputs to outputs. A spam filter maps an email to a spam decision; a speech recognizer maps sound to text. In machine learning, a model is a computer program that encodes a function learned from examples. A neural network is one way to represent that function, with many small calculations joined together.

Learning has two stages. During training, the algorithm adjusts the model to match examples in a dataset. Once training ends, the model is fixed. During inference, it applies the learned mapping to new examples and produces estimates. Calling the process “learning” can suggest human-like understanding, but the book’s technical account is narrower: parameters are adjusted to improve the match between input patterns and target outputs.

Analysis

This distinction helps separate performance from understanding. A system can produce useful outputs without possessing a human account of why those outputs make sense. The function perspective also gives a practical test: identify what goes in, what comes out, and what the output is supposed to support. If those boundaries are vague, claims about a model’s abilities are difficult to evaluate. The limitation is that a simple mapping description can conceal the social meaning of its outputs, especially when a prediction is treated as a decision.

Audit the data, model choices, and scoring rule before trusting an answer.

Three ingredients shape learning. A machine learning project needs examples, a set of candidate functions, and a measure of how well each function fits. In practice, the dataset’s features matter greatly: too few can omit useful information, while irrelevant or redundant features can encourage spurious patterns. The book notes that data preparation can consume as much as 80% of a project’s budget.

Algorithms bring assumptions. An inductive bias is the set of assumptions that guides an algorithm toward some functions rather than others. Without such assumptions, many different functions may fit the same limited examples. A bias that is too strong can cause underfitting, where the model misses real complexity. A bias that is too weak can cause overfitting, where it learns noise in the training data and performs poorly on new examples.

Analysis

This is a useful corrective to the idea that more data automatically produces better decisions. The data must represent the task, and the scoring rule must reward the behavior that actually matters. A model can optimize its stated measure while missing the purpose behind it, so technical validation and domain judgment need to work together. The book’s three ingredients also clarify where scrutiny belongs: errors can originate in the examples, in the model family, or in the definition of success.

A network gets expressive by stacking simple calculations with nonlinear gates.

Each neuron is simple. An artificial neuron multiplies its inputs by weights, adds the results, then passes the sum through an activation function. The activation function introduces nonlinearity, allowing the network to represent relationships more complex than a straight-line mapping. Without it, stacking layers would still produce only a linear function.

Layers build representations. Each neuron transforms information for the next layer, so a network can combine simple patterns into more abstract ones. In image recognition, early layers may respond to edges or curves, while deeper layers combine them into parts and then objects. The XOR logic problem shows why this matters: a single threshold neuron cannot separate its true and false cases with one straight boundary, but a multilayer network can combine intermediate results to represent it.

Analysis

The compositional view makes neural networks less mysterious without making them trivial. Their power comes from arranging many simple transformations so that the whole system can represent a complicated mapping. This resembles a general strategy in mathematics and engineering: build a difficult operation from reusable components. Yet the resulting components do not necessarily correspond to human-readable concepts. A layer may be useful because it helps the next layer predict correctly, not because its internal representation has an obvious interpretation.

Depth is useful only when training signals reach the network's early layers.

Credit must travel backward. Training a network requires determining how each weight contributed to the final error, including weights far from the output. Backpropagation addresses this credit assignment problem by sending error information backward through the network. In deep networks, however, repeated multiplication can shrink that signal. This is the vanishing gradient problem: early layers receive tiny updates and may learn very slowly.

Practical advances helped. The history of neural networks includes repeated waves of promise and disappointment. In 1969, Minsky and Papert highlighted limits of single-layer perceptrons, including their inability to learn XOR. Multilayer networks could represent richer functions, but training them was difficult. ReLU activations helped gradients flow through active units, while improved weight initialization and skip-connections supported deeper models. These techniques made depth more trainable, not automatically more useful.

Analysis

Depth is a resource, not a guarantee. The network must receive a usable learning signal, and the architecture and training choices determine whether that signal survives. This is a broader lesson about complex systems: adding components increases capacity, but also creates more ways for coordination to fail. The historical account also resists a simple tale of steady progress. A promising representation can remain impractical until the training method, data supply, and computational resources make it workable.

Match the architecture to the data: images need locality, sequences need memory.

CNNs reuse visual detectors. A convolutional neural network applies the same set of weights, called a kernel, to small regions across an image. Each region is a receptive field. Reusing weights lets the network detect a feature wherever it appears; the resulting feature map records where that detector responded. Pooling can reduce the map’s size and discard some exact location information. Deeper layers combine local detections into more complex visual patterns.

RNNs carry context forward. A recurrent neural network processes a sequence one element at a time, passing information from one step to the next. Long short-term memory units use gates to control which information is retained, updated, or output, helping with long-range dependencies. In sequence-to-sequence translation, one network encodes the source sentence and another generates the translation. These architectures build assumptions about the structure of their data into the network.

Analysis

Architecture is a form of prior knowledge. A CNN assumes that nearby pixels and repeated visual patterns matter; an RNN assumes that sequence order and past context matter. Those assumptions reduce the space of possible functions the model must search, which can make learning more efficient. The tradeoff is that a useful assumption in one domain can be restrictive in another. The design question is therefore not simply which architecture is newest, but which structure most closely reflects the task’s meaningful regularities.

Train by tracing each error backward, then taking measured steps downhill.

Gradient descent adjusts weights. A training algorithm measures model error, then changes parameters in a direction expected to reduce it. The learning rate controls the size of each change: steps that are too large can overshoot, while steps that are too small can make training slow. For a simple linear model, the error surface can resemble a bowl, and the search can move toward its lowest point.

Backpropagation assigns blame. First, an example passes forward through the network and produces an output. The algorithm compares that output with the expected answer, then propagates error information backward to estimate each neuron’s contribution. Gradient descent uses those estimates to update weights. Neural networks have complex error surfaces with multiple valleys, so training is not guaranteed to find the global best weights. Repeating training with different starting weights can help explore alternatives.

Analysis

The two algorithms answer different questions: backpropagation estimates how changes affect error, while gradient descent uses those estimates to change the parameters. Keeping that distinction clear makes the training process easier to reason about. The “downhill” metaphor is helpful, but it can imply a simple landscape and reliable destination. In a complex model, the terrain may have many local low points, and practical success means finding a useful solution, not proving that the best possible one has been found.

Reuse learned features when labels are scarce, and treat synthetic data cautiously.

Labels cost time and money. Supervised learning requires examples paired with the correct answers. ImageNet, which supported major image-recognition advances, contains more than 14 million manually annotated images, with more than one million also marked with object locations. Such labeling is expensive to repeat for every new task.

Reuse and generation offer alternatives. Transfer learning repurposes information learned on a related task. For example, early layers of a pretrained image model may already detect edges and other general visual features, so a new project can adapt later layers using a smaller labeled dataset. Unsupervised methods such as autoencoders learn useful representations without target labels. Generative adversarial networks train one model to create synthetic examples and another to distinguish them from real data. Such models can create useful images, but can also produce deceptive deepfake videos.

Analysis

These methods address different parts of the label shortage. Transfer learning carries forward patterns that may generalize, while generative approaches create additional examples. Neither removes the need to ask whether the resulting data fits the intended task. Synthetic examples can expand a dataset, but their usefulness depends on how faithfully they reflect relevant variation. Likewise, transfer works best when the source and target tasks share structure. Reuse saves effort when the inherited patterns are genuinely relevant, not merely because a model is already available.

Deep learning advances when data, algorithms, and specialized hardware reinforce each other.

The gains compound. Neural networks perform many matrix operations, which graphics processing units can accelerate. Faster hardware makes it feasible to train larger networks on larger datasets. More data, in turn, can make flexible neural networks more effective, while algorithmic improvements help them learn. The book describes this as a reinforcing cycle that helped move deep learning from research into widely used applications.

Hardware also sets limits. Training and deployment require computation and energy. The book discusses neuromorphic chips, which use networks of spiking neurons and aim to reduce energy use through sparse, time-dependent activity. It also describes research into quantum computing, while noting that useful commercial-scale systems remained a future prospect at the time of writing. A model’s progress therefore depends not only on its mathematics, but also on what machines can practically run.

Analysis

This account makes clear that technical breakthroughs rarely arrive in isolation. A new algorithm matters more when hardware can run it, and abundant data matters more when a model can use it. The cycle also invites a caution: increasing computational scale can improve a defined task without resolving whether that task is socially valuable or fairly measured. Efficiency is not just about producing a more capable system; it also concerns the resources consumed and the consequences of putting the system to work.

Require explanations when opaque predictions govern health, credit, or privacy.

Accuracy is not the whole result. Deep learning models can process medical images, assess credit-related data, and analyze traces of online behavior. Yet their many connected neurons and distributed internal representations can make it difficult to explain how a particular input produced a decision. When decisions affect people, knowing the output alone may not be enough. The book connects this challenge to privacy and the right to explanations for automated decisions under European data protection rules.

Interpretation has tools. Feature visualization looks for inputs that activate particular parts of a network. Attribution estimates how components contribute to a prediction, often with saliency maps. Dimensionality reduction projects high-dimensional representations into a view that people can inspect. These methods can reveal patterns, possible bias, or failure cases, but interpreting the inner workings of a network remains a developing challenge.

Analysis

Interpretability is not decorative documentation. It can help expose when a model relies on unexpected signals or fails on particular cases, and it matters most when people bear the cost of an error. Still, a visualization of influential features is not automatically a complete explanation of a decision. The standard of explanation should depend on the stakes: a low-impact recommendation and a consequential medical or credit decision do not require the same level of scrutiny.

Analysis

The book’s central achievement is to replace mystique with a working mental model. Deep learning is presented not as a digital brain, but as a flexible method for fitting input-output functions: neurons calculate weighted sums, activation functions introduce nonlinearity, layers compose transformations, and training adjusts weights against examples. This account helps a general reader understand both the reach and the limits of the technology.

Its narrative also makes technical progress contingent. Neural networks did not simply improve because researchers kept adding layers. The field encountered representational limits, trouble training hidden layers, vanishing gradients, and periods when competing approaches were easier to use. The modern rise of deep learning depended on better training methods, large datasets, and hardware suited to matrix calculations. That history is a useful antidote to technological inevitability: capability emerges from interacting conditions, not a single clever idea.

The most important tension is between flexibility and accountability. A model’s capacity to learn features from raw data reduces the burden of hand-designing them, but it also makes the learned representation harder to interpret. A high score on a benchmark may establish performance on that benchmark, but it does not settle whether the data are representative, whether the evaluation captures the real goal, or whether the prediction should govern a human decision. The book recognizes these questions through its discussion of privacy and interpretability.

A reader should therefore leave with two complementary instincts. First, ask how the model works at the level of its task: what inputs it learns from, what output it produces, and how error changes its parameters. Second, ask what its success does not prove. Deep learning can be extraordinarily effective in complex, data-rich domains without becoming general intelligence, supplying a causal explanation, or guaranteeing fair outcomes. Its power is specific, conditional, and increasingly consequential.

Summary in 30 Seconds
Deep learning fits input-to-output mappings by adjusting network weights against examples. Layers combine simple calculations into complex patterns, while backpropagation and gradient descent update the model to reduce error. Images and sequences benefit from different architectures, and pretrained features can help when labeled examples are scarce. Data preparation can consume up to 80% of a project, while ImageNet contains more than 14 million annotated images. Strong results still depend on representative data, sound evaluation, and scrutiny of consequential predictions.
Contains spoilers

Last updated:

Report Issue

Review Summary

3.89 out of 5
Average of 473 ratings from Goodreads and Amazon.

Deep Learning receives mixed reviews, with an average rating of 3.91/5. Many praise it as an informative introduction to deep learning concepts, especially for those with some technical background. However, some criticize it for being too technical for general readers, despite being marketed as accessible. Readers appreciate the historical context and explanations of neural networks, but some find the math challenging. The book is commended for its comprehensive overview but criticized for lacking sufficient coverage of real-world applications and ethical considerations.

Your rating:
4.42
154 ratings
Want to read the full book?

Glossary

Deep Learning

Layered neural-network learning

A subfield of machine learning focused on training neural networks with multiple hidden layers. These networks learn functions that map inputs to outputs, often by deriving useful features from large datasets. Their layered structure allows them to represent complex, nonlinear relationships and supports applications such as image recognition, speech processing, and machine translation.

Inductive bias

Assumptions guiding model choice

The assumptions an algorithm uses to prefer some candidate functions over others when the available examples do not uniquely determine a solution. Different algorithms encode different biases. A bias that is too restrictive can cause underfitting, while an overly permissive model can fit noise and overfit the training data.

Overfitting

Learning noise as pattern

A model overfits when it becomes so closely adapted to its training examples that it captures noise or sample-specific quirks rather than patterns that generalize. Such a model may perform well on the examples used to train it but poorly on new data. The risk is affected by model flexibility, data quality, and dataset size.

Activation function

Nonlinear neuron output mapping

A function that receives a neuron's weighted sum of inputs and converts it into the neuron's output, or activation. Nonlinear activation functions allow layers of neurons to represent complex nonlinear mappings. Common examples discussed in the book include logistic, tanh, and rectified linear functions.

Backpropagation

Backward error attribution

An algorithm for calculating how the network’s error changes in response to weights throughout a neural network, including weights in hidden layers. It first passes an example forward to produce an output, then propagates error information backward through the network. The resulting gradients can be used by an optimization method, such as gradient descent, to update weights.

Gradient descent

Error-reducing parameter updates

An optimization algorithm that adjusts model parameters in a direction expected to reduce error. It uses the gradient of an error measure to determine how each weight should change. The learning rate controls the size of the updates. In neural networks, gradient descent is commonly paired with backpropagation, which calculates the needed error gradients.

Convolutional Neural Network

Shared filters for spatial data

A neural network architecture commonly used for images. A convolutional layer applies shared weights, called a kernel, to local regions of an input. The resulting feature map records where a learned pattern is detected. Nonlinear activations and pooling are often added, and deeper layers combine local features into more complex representations.

Recurrent Neural Network

Network with sequence memory

A neural network architecture designed to process sequential data. It passes information from one input step to the next, allowing each step to be processed in the context of earlier inputs. This recurrent flow provides a form of memory. Long sequences can make training difficult because error signals may vanish as they are propagated backward.

Long Short-Term Memory (LSTM)

Gated sequence memory

A recurrent-network unit designed to help retain information across longer sequences and reduce the effects of vanishing gradients. It stores activations in a cell and uses gates to control which information is forgotten, added, or used in the output. LSTM networks have been used for language tasks, including machine translation.

Transfer learning

Reuse across related tasks

The use of information or representations learned on one task to assist learning on a related task. In image processing, early layers of a pretrained network may already detect broadly useful visual features. Reusing them can reduce the amount of new labeled data and training effort required for a specialized application.

Generative adversarial network (GAN)

Competing data-generation models

A generative approach built from two neural networks trained in opposition. The generator creates synthetic examples, while the discriminator tries to distinguish generated examples from real data. The generator learns to produce examples that can fool the discriminator. GANs can synthesize images, but the same capabilities can also support deceptive deepfake media.

Vanishing gradient

Diminishing learning signal

A training problem in which error gradients shrink as they are propagated backward through many network layers or across long sequences. When gradients become very small, weights in early layers receive tiny updates and may learn slowly. Recurrent networks are especially vulnerable over long sequences. Activation functions and network designs can affect the severity of the problem.

Distributed representation

Concept encoded across units

A representation in which a concept is encoded by a pattern of activations across multiple neurons, rather than by a single neuron assigned to that concept. Each neuron can contribute to multiple concepts. Such representations can be compact and useful for learning, though the meaning of an individual hidden unit may not be readily interpretable.

About the Author

John D. Kelleher is a Professor of Computer Science and Academic Leader at the Dublin Institute of Technology's Information, Communication, and Entertainment Research Institute. He co-authored "Fundamentals of Machine Learning for Predictive Data Analytics" published by MIT Press. Kelleher's expertise in computer science, particularly in machine learning and artificial intelligence, is evident in his work. His book "Deep Learning" is part of the MIT Press Essential Knowledge series, aiming to provide concise introductions to complex topics. Kelleher's approach combines technical depth with efforts to make the subject accessible, though some readers find the balance challenging. His work contributes to the growing body of literature explaining advanced AI concepts to a broader audience.

Download PDF

To save this Deep Learning summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.41 MB     Pages: 17

Download EPUB

To read this Deep Learning summary on your e-reader device or app, download the free EPUB. The .epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.
Download EPUB
File size: 1.53 MB     Pages: 22
Want to read the full book?
Follow
Listen1 mins
Home
Swipe
Library
Get App
Try Full Access for 3 Days
Listen, bookmark, and more
Compare Features Free Pro
📖 Read Summaries
Read unlimited summaries. Free users get 3 per month
🎧 Listen to Summaries
Listen to unlimited summaries in 40 languages
—
❤️ Unlimited Bookmarks
Free users are limited to 4
—
📜 Unlimited History
Free users are limited to 4
—
📥 Unlimited Downloads
Free users are limited to 1
—
Risk-Free Timeline
Today: Get Instant Access
Listen to full summaries of 26,000+ books. That's 12,000+ hours of audio!
Day 2: Trial Reminder
We'll send you a notification that your trial is ending soon.
Day 3: Your subscription begins
You'll be charged on Nov 24,
cancel anytime before.
Consume 2.8× More Books
2.8× more books Listening Reading
Our users love us
1M+ readers
Trustpilot Rating
TrustPilot
4.6 Excellent
This site is a total game-changer. I've been flying through book summaries like never before. Highly, highly recommend.
— Dave G
Worth my money and time, and really well made. I've never seen this quality of summaries on other websites. Very helpful!
— Em
Highly recommended!! Fantastic service. Perfect for those that want a little more than a teaser but not all the intricate details of a full audio book.
— Greg M
Save 62%
Yearly
$119.88 $44.99/year/yr
$3.75/mo
Monthly
$9.99/mo
Start a 3-Day Free Trial
3 days free, then $44.99/year. Cancel anytime.
Unlock a world of fiction & nonfiction books
26,000+ books for the price of 2 books
Read any book in 10 minutes
Discover new books like Tinder
Request any book if it's not summarized
Read more books than anyone you know
#1 app for book lovers
Lifelike & immersive summaries
30-day money-back guarantee
Download summaries in EPUBs or PDFs
Cancel anytime in a few clicks

Why one summary?

Scanner
Find a barcode to scan

⭐
⭐
⭐
We have a special gift for you
Open
38% off
DISCOUNT FOR YOU
⭐
⭐
⭐
$79.99
$49.99/year
Only $4.16 per month
Continue
2 taps to start, super easy to cancel
Settings
General
Widget
Loading...