Book Review: Deep Learning with

作者：Francois Chollet
出版社：Manning
发行时间：December 4th 2017
来源：下载的 pdf 版本
Goodreads：4.7 (148 Ratings)
豆瓣：9.4（50人评价）

Artificial intelligence:
Artificial intelligence was born in the 1950s, when a handful of pioneers from the nascent field of computer science started asking whether computers could be made to “think”—a question whose ramifications we’re still exploring today. A concise definition of the field would be as follows: the effort to automate intellectual tasks normally performed by humans. As such, AI is a general field that encompasses machine learning and deep learning, but that also includes many more approaches that don’t involve any learning. Early chess programs, for instance, only involved hardcoded rules crafted by programmers, and didn’t qualify as machine learning. For a fairly long time, many experts believed that human-level artificial intelligence could be achieved by having programmers handcraft a sufficiently large set of explicit rules for manipulating knowledge. This approach is known as symbolic AI, and it was the dominant paradigm in AI from the 1950s to the late 1980s.
It reached its peak popularity during the expert systems boom of the 1980s. Although symbolic AI proved suitable to solve well-defined, logical problems, such as playing chess, it turned out to be intractable to figure out explicit rules for solving more complex, fuzzy problems, such as image classification, speech recognition, and language translation. A new approach arose to take symbolic AI’s place: machine learning.

Machine learning:
Machine learning arises from this question: could a computer go beyond “what we know how to order it to perform” and learn on its own how to perform a specified task? Could a computer surprise us? Rather than programmers crafting data-processing rules by hand, could a computer automatically learn these rules by looking at data?
This question opens the door to a new programming paradigm. In classical programming, the paradigm of symbolic AI, humans input rules (a program) and data to be processed according to these rules, and out come answers (see figure 1.2). With machine learning, humans input data as well as the answers expected from the data, and out come the rules. These rules can then be applied to new data to produce original answers.
A machine-learning system is trained rather than explicitly programmed. It’s presented with many examples relevant to a task, and it finds statistical structure in these examples that eventually allows the system to come up with rules for automating the task. For instance, if you wished to automate the task of tagging your vacation pictures, you could present a machine-learning system with many examples of pictures already tagged by humans, and the system would learn statistical rules for associating specific pictures to specific tags.
Although machine learning only started to flourish in the 1990s, it has quickly become the most popular and most successful subfield of AI, a trend driven by the availability of faster hardware and larger datasets. Machine learning is tightly related to mathematical statistics, but it differs from statistics in several important ways. Unlike statistics, machine learning tends to deal with large, complex datasets (such as a dataset of millions of images, each consisting of tens of thousands of pixels) for which classical statistical analysis such as Bayesian analysis would be impractical. As a result, machine learning, and especially deep learning, exhibits comparatively little mathematical theory—maybe too little—and is engineering oriented. It’s a hands-on discipline in which ideas are proven empirically more often than theoretically.

Deep learning:
To define deep learning and understand the difference between deep learning and other machine-learning approaches, first we need some idea of what machine learning algorithms do. I just stated that machine learning discovers rules to execute a data-processing task, given examples of what’s expected. So, to do machine learning, we need three things:

• Input data points—For instance, if the task is speech recognition, these data points could be sound files of people speaking. If the task is image tagging, they could be pictures.
• Examples of the expected output—In a speech-recognition task, these could be human-generated transcripts of sound files. In an image task, expected outputs could be tags such as “dog,” “cat,” and so on.
• A way to measure whether the algorithm is doing a good job—This is necessary in order to determine the distance between the algorithm’s current output and its expected output. The measurement is used as a feedback signal to adjust the way the algorithm works. This adjustment step is what we call learning.

A machine-learning model transforms its input data into meaningful outputs, a process that is “learned” from exposure to known examples of inputs and outputs. Therefore, the central problem in machine learning and deep learning is to meaningfully transform data: in other words, to learn useful representations of the input data at hand—representations that get us closer to the expected output. Before we go any further: what’s a representation? At its core, it’s a different way to look at data—to represent or encode data. For instance, a color image can be encoded in the RGB format (red-green-blue) or in the HSV format (hue-saturation-value): these are two different representations of the same data. Some tasks that may be difficult with one representation can become easy with another. For example, the task “select all red pixels in the image” is simpler in the RG format, whereas “make the image less saturated” is simpler in the HSV format. Machine-learning models are all about finding appropriate representations for their input data—transformations of the data that make it more amenable to the task at hand, such as a classification task.

Deep learning is a specific subfield of machine learning: a new take on learning representations from data that puts an emphasis on learning successive layers of increasingly meaningful representations. The deep in deep learning isn’t a reference to any kind of deeper understanding achieved by the approach; rather, it stands for this idea of successive layers of representations. How many layers contribute to a model of the data is called the depth of the model. Other appropriate names for the field could have been layered representations learning and hierarchical representations learning. Modern deep learning often involves tens or even hundreds of successive layers of representations— and they’re all learned automatically from exposure to training data. Meanwhile, other approaches to machine learning tend to focus on learning only one or two layers of representations of the data; hence, they’re sometimes called shallow learning.
In deep learning, these layered representations are (almost always) learned via models called neural networks, structured in literal layers stacked on top of each other. The term neural network is a reference to neurobiology, but although some of the central concepts in deep learning were developed in part by drawing inspiration from our understanding of the brain, deep-learning models are not models of the brain. There’s no evidence that the brain implements anything like the learning mechanisms used in modern deep-learning models. You may come across pop-science articles proclaiming that deep learning works like the brain or was modeled after the brain, but that isn’t the case. It would be confusing and counterproductive for newcomers to the field to think of deep learning as being in any way related to neurobiology; you don’t need that shroud of “just like our minds” mystique and mystery, and you may as well forget anything you may have read about hypothetical links between deep learning and biology. For our purposes, deep learning is a mathematical framework for learning representations from data.

At this point, you know that machine learning is about mapping inputs (such as images) to targets (such as the label “cat”), which is done by observing many examples of input and targets. You also know that deep neural networks do this input-to-target mapping via a deep sequence of simple data transformations (layers) and that these data transformations are learned by exposure to examples. Now let’s look at how this learning happens, concretely.
The specification of what a layer does to its input data is stored in the layer’s weights, which in essence are a bunch of numbers. In technical terms, we’d say that the transformation implemented by a layer is parameterized by its weights (see figure 1.7). (Weights are also sometimes called the parameters of a layer.) In this context, learning means finding a set of values for the weights of all layers in a network, such that the network will correctly map example inputs to their associated targets. But here’s the thing: a deep neural network can contain tens of millions of parameters. Finding the correct value for all of them may seem like a daunting task, especially given that modifying the value of one parameter will affect the behavior of all the others!

Although deep learning is a fairly old subfield of machine learning, it only rose to prominence in the early 2010s. In the few years since, it has achieved nothing short of a revolution in the field, with remarkable results on perceptual problems such as seeing and hearing—problems involving skills that seem natural and intuitive to humans but have long been elusive for machines.
In particular, deep learning has achieved the following breakthroughs, all in historically difficult areas of machine learning:

• Near-human-level image classification
• Near-human-level speech recognition
• Near-human-level handwriting transcription
• Improved machine translation
• Improved text-to-speech conversion
• Digital assistants such as Google Now and Amazon Alexa
• Near-human-level autonomous driving
• Improved ad targeting, as used by Google, Baidu, and Bing
• Improved search results on the web
• Ability to answer natural-language questions
• Superhuman Go playing

Although deep learning has led to remarkable achievements in recent years, expectations for what the field will be able to achieve in the next decade tend to run much higher than what will likely be possible. Although some world-changing applications like autonomous cars are already within reach, many more are likely to remain elusive for a long time, such as believable dialogue systems, human-level machine translation across arbitrary languages, and human-level natural-language understanding. In particular, talk of human-level general intelligence shouldn’t be taken too seriously. The risk with high expectations for the short term is that, as technology fails to deliver, research investment will dry up, slowing progress for a long time.
This has happened before. Twice in the past, AI went through a cycle of intense optimism followed by disappointment and skepticism, with a dearth of funding as a result. It started with symbolic AI in the 1960s. In those early days, projections about AI were flying high. One of the best-known pioneers and proponents of the symbolic AI approach was Marvin Minsky, who claimed in 1967, “Within a generation … the problem of creating ‘artificial intelligence’ will substantially be solved.” Three years later, in 1970, he made a more precisely quantified prediction: “In from three to eight years we will have a machine with the general intelligence of an average human being.” In 2016, such an achievement still appears to be far in the future—so far that we have no way to predict how long it will take—but in the 1960s and early 1970s, several experts believed it to be right around the corner (as do many people today). A few years later, as these high expectations failed to materialize, researchers and government funds turned away from the field, marking the start of the first AI winter (a reference to a nuclear winter, because this was shortly after the height of the Cold War).
It wouldn’t be the last one. In the 1980s, a new take on symbolic AI, expert systems, started gathering steam among large companies. A few initial success stories triggered a wave of investment, with corporations around the world starting their own in-house AI departments to develop expert systems. Around 1985, companies were spending over $1 billion each year on the technology; but by the early 1990s, these systems had proven expensive to maintain, difficult to scale, and limited in scope, and interest died down. Thus began the second AI winter.
We may be currently witnessing the third cycle of AI hype and disappointment— and we’re still in the phase of intense optimism. It’s best to moderate our expectations for the short term and make sure people less familiar with the technical side of the field have a clear idea of what deep learning can and can’t deliver.

In particular, the Random Forest algorithm introduced a robust, practical take on decision-tree learning that involves building a large number of specialized decision trees and then ensembling their outputs. Random forests are applicable to a wide range of problems—you could say that they’re almost always the second-best algorithm for any shallow machine-learning task. When the popular machine-learning competition website Kaggle (http://kaggle.com) got started in 2010, random forests quickly became a favorite on the platform—until 2014, when gradient boosting machines took over. A gradient boosting machine, much like a random forest, is a machine-learning technique based on ensembling weak prediction models, generally decision trees. It uses gradient boosting, a way to improve any machine-learning model by iteratively training new models that specialize in addressing the weak points of the previous models. Applied to decision trees, the use of the gradient boosting technique results in models that strictly outperform random forests most of the time, while having similar properties. It may be one of the best, if not the best, algorithm for dealing with nonperceptual data today. Alongside deep learning, it’s one of the most commonly used techniques in Kaggle competitions.

Around 2010, although neural networks were almost completely shunned by the scientific community at large, a number of people still working on neural networks started to make important breakthroughs: the groups of Geoffrey Hinton at the University of Toronto, Yoshua Bengio at the University of Montreal, Yann LeCun at New York University, and IDSIA in Switzerland.
In 2011, Dan Ciresan from IDSIA began to win academic image-classification competitions with GPU-trained deep neural networks—the first practical success of modern deep learning. But the watershed moment came in 2012, with the entry of Hinton’s group in the yearly large-scale image-classification challenge ImageNet. The ImageNet challenge was notoriously difficult at the time, consisting of classifying highresolution color images into 1,000 different categories after training on 1.4 million images. In 2011, the top-five accuracy of the winning model, based on classical approaches to computer vision, was only 74.3%. Then, in 2012, a team led by Alex Krizhevsky and advised by Geoffrey Hinton was able to achieve a top-five accuracy of 83.6%—a significant breakthrough. The competition has been dominated by deep convolutional neural networks every year since. By 2015, the winner reached an accuracy of 96.4%, and the classification task on ImageNet was considered to be a completely solved problem.
Since 2012, deep convolutional neural networks (convnets) have become the go-to algorithm for all computer vision tasks; more generally, they work on all perceptual tasks. At major computer vision conferences in 2015 and 2016, it was nearly impossible to find presentations that didn’t involve convnets in some form. At the same time, deep learning has also found applications in many other types of problems, such as natural-language processing. It has completely replaced SVMs and decision trees in a wide range of applications. For instance, for several years, the European Organization for Nuclear Research, CERN, used decision tree–based methods for analysis of particle data from the ATLAS detector at the Large Hadron Collider (LHC); but CERN eventually switched to Keras-based deep neural networks due to their higher performance and ease of training on large datasets.

The two key ideas of deep learning for computer vision—convolutional neural networks and backpropagation—were already well understood in 1989. The Long ShortTerm Memory (LSTM) algorithm, which is fundamental to deep learning for timeseries, was developed in 1997 and has barely changed since. So why did deep learning only take off after 2012? What changed in these two decades?
In general, three technical forces are driving advances in machine learning:

• Hardware
• Datasets and benchmarks
• Algorithmic advances

Because the field is guided by experimental findings rather than by theory, algorithmic advances only become possible when appropriate data and hardware are available to try new ideas (or scale up old ideas, as is often the case). Machine learning isn’t mathematics or physics, where major advances can be done with a pen and a piece of paper. It’s an engineering science.

In addition to hardware and data, until the late 2000s, we were missing a reliable way to train very deep neural networks. As a result, neural networks were still fairly shallow, using only one or two layers of representations; thus, they weren’t able to shine against more-refined shallow methods such as SVMs and random forests. The key issue was that of gradient propagation through deep stacks of layers. The feedback signal used to train neural networks would fade away as the number of layers increased.
This changed around 2009–2010 with the advent of several simple but important algorithmic improvements that allowed for better gradient propagation:

• Better activation functions for neural layers
• Better weight-initialization schemes, starting with layer-wise pretraining, which was quickly abandoned
• Better optimization schemes, such as RMSProp and Adam

Only when these improvements began to allow for training models with 10 or more layers did deep learning start to shine.
Finally, in 2014, 2015, and 2016, even more advanced ways to help gradient propagation were discovered, such as batch normalization, residual connections, and depthwise separable convolutions. Today we can train from scratch models that are thousands of layers deep.

The problem we’re trying to solve here is to classify grayscale images of handwritten digits (28 × 28 pixels) into their 10 categories (0 through 9). We’ll use the MNIST dataset, a classic in the machine-learning community, which has been around almost as long as the field itself and has been intensively studied. It’s a set of 60,000 training images, plus 10,000 test images, assembled by the National Institute of Standards and Technology (the NIST in MNIST) in the 1980s. You can think of “solving” MNIST as the “Hello World” of deep learning—it’s what you do to verify that your algorithms are working as expected. As you become a machine-learning practitioner, you’ll see MNIST come up over and over again, in scientific papers, blog posts, and so on. You can see some MNIST samples in figure 2.1.

This branch of machine learning consists of finding interesting transformations of the input data without the help of any targets, for the purposes of data visualization, data compression, or data denoising, or to better understand the correlations present in the data at hand. Unsupervised learning is the bread and butter of data analytics, and it’s often a necessary step in better understanding a dataset before attempting to solve a supervised-learning problem. Dimensionality reduction and clustering are well-known categories of unsupervised learning.

1. Staying up to date in a fast-moving field
2. Practice on real-world problems using Kaggle
3. Read about the latest developments on arXiv
4. Explore the Keras ecosystem