Book Review: Artificial Intelligence: A Modern Approach

The third book that I read about Artificial intelligence is a rather academic one, written by Stuart Russell and Peter Norvig.

With more than a thousand pages, it is too heavy for an introduction to the field. It is more suitable for people looking for an academic, or theoretical research of the AI field.

The book focuses quite much on information theory, searching, representation of knowledge and how to handle uncertainties. All those subjects are interesting, but all the theory in the first 850 pages feel a bit overwhelming.

The interesting chapters are in the end (natural language, image perception, robotics and the road ahead).

If I'd have more time to spend reading the book (I'm on a parental leave now and the book loan is four weeks only), I would have given it more time. Now, I skipped 3/4 of the book. That is more related to my personal situation than the qualities of the book.

I think that a combination of the two other books that I've written about will give a good introduction to the AI field.

One Sidenote
There is one aspect of AI that I've been thinking above that I don't see in the books: We, the humans. In many cases we don't know, or we don't admit what goals we want to pursue. Often the goals, or the focus changes quickly, and we often choose actions that will prevent the goals.

For example, the overall goal with customer support centres is to resolve the end users issues. When KPIs are introduced to monitor the efficiency (often defined as time to resolve tickets), the agents have incentives to close tickets before the issue is solved, forcing the end user to call another agent and describe the problem again. If we request an AI to target several conflicting goals, we will see very... interesting results.

I've also seen cases where policymakers want to dodge responsibility for their decicions by forcing their own agencies to make cruel decisions. When the decisions are critizised, the policymakers can blame the agency.

On the other hand, when defining utility functions and their relative priority and preferences, hopefully a clever AI system, or its designers will be able to confront the customer and make them make up their mind.

Book Review: Python Machine Learning Blueprints

Python Machine Learning Blueprints is written by Alexander T Combs and is illustrating how to do some machine learning, data science and advanced web scraping using Python.

It shows some real-world examples using some very useful packages:
Python Requests  - Makes it easy to send HTTP requests
Jupyter Notebook - Useful for visualizing data inspection and creating documents with live code and visualisation.
Pandas - Useful for data analysis
Matplotlib - Based on Matlab,  useful for plotting data.
Seaborn - Useful for visual analysis of data, offering for example violin diagrams.
Statsmodels - Useful for working with models and statistical tests.
Scikit Learn - Useful for visualizations, regression, model selection, clustering etc.

feedback@packtpub.com

The examples are clearly presented, along with both code and diagrams showing how the data was analyzed. Some of them are: Finding underpriced apartments, Analyzing and predicting viral content, Building an image similarity engine and more.

One flaw with the book was the chapter about deep learning. It seems that the author didn't get the numbers right, which made it very hard to follow the calculations.

The expected output isn't specified in the second section/second sentence. The weight function got W and X wrong and W2 was set to 0.2 in the text but 0.4 in the equation. This makes it very hard to follow the example.

I have some other pet projects that take time now, so I won't be able to do serious data mining/machine learning now, but once I have time to explore these topics, I'll buy this book,

Book Review: Superintelligence - Paths, Dangers, Strategies

Regarding AI, I was naive before reading this book. Now I'm perhaps more alarmistic.

Nick Boström is a Swedish professor in Philosophy at the University of Oxford with focus on superintelligence and the existensial risks we might face when developing agents that are far more intelligent than we humans are together.

The problem is still on a hypothetical level, and a general superintelligence will probably take time to happen, if ever. However, it is important to address the risks and find powerful safeguards (as Boström explains, this is difficult),

What's AI?
Artificial Intelligence is a buzz-word that is hard to define. One definition might be the ability to solve problems humans can solve. The field has been explored formally since a workshop on Darmouth College in 1956. Since then, there has been several periods of extensive research followed by AI winters.

There is a distinction between general and narrow AI, where narrow AI adresses more specific problems, such as chess games, trading algorithms and Apple's Siri. General AI refers to an ability to solve general problems that humans can solve. Most AI systems of today are narrow AI.

General intelligence is much more complicated to develop. This includes common sense, which is hard to formulate.

Superintelligence is an actor/agent that has an intelligence far above the most intelligent human beings.

Scales of Intelligence
Here we are biased by the human scale of intelligence, that ranges from Lloyd Christmas and Harry Dunne up to sir Isaac Newton and Ada Lovelace. This scale apply to human intelligence, but there are more life forms than that. Fungus, for example have much weaker intellectual capability than all human beings. At the other end of the scale, there might be agents with far higher intelligence than the brightest human beings.

Analogy: Human beings are generally up to two meters tall, with a world record of 272 cm. Blue whales twelve times longer has been recorded. Redwood trees can exceed 91 meters. This illustrates that human scales are very narrow. If we can meet creatures that are twelve or fourty times longer than we are, it would be possible to meet creatures twelve or fourty times more intelligent.

Types of Superintelligence
Nick Boström discusses different types of superintellligence:

• Speed (doing what we can do but faster)
• Collective (many smaller intellects cooperating to form a superintelligence)
• Quiality superintelligence (doing what we can do but with better quality)

He foresees that the speed of the AI takeoff (speed for the system to gain superintelligence itself) will happen very quickly. That will happen when the system finds ways to improve its capacity, develop new code and convincing people to assign more resources to that system.

Things Getting Nasty?

An agent that is smart enough will learn how when to conceal its intelligence when it finds it suitable. It may also try to escape from its sandbox environment, for example by convincing a technician to plug in a network cable, or bring a flash disk with a copy of some data.

Boström is also discussing the goals of superintelligence systems. The targets that we give the system can easily become harmful. For example, an AI system that is supposed to make people smile can cheat by applying plastic surgery dooming people to statically grinning faces, Further, if a superintelligence has different goals than humans have, the system will find ways to secure the resources. Perhaps not nice or friendly ways.

There are several movies describing clashes between AI and humans. If we ever ceate a superintelligence powerful enough to harm us, it will probably do its homework enough not to give its opponent any chance.

AI Research

The AI research has several agents: governments, companies, criminals, academics. Many of them are in an arms-race with each other. They cannot afford losing that race, and there is a significant risk that many projects will rush themselves to reach superintelligence. There is a risk that projects will ignore the safeguards that may prevent AI from becoming harmful.

Artificial intelligence has already been used to manipulate elections. AI is changin our world now. In other countries, AI is used to rank citizens after how correct their behaviour is. I strongly recommend everyone affected by AI (everyone) to read this book.

Book Review: Designing User Interfaces for an Aging Population

After improving TrafficControl, I found this book on my local library:

As the authors are emphasizing, the aging population is a very diverse group with a lot of individuals with different abilities, knowledge and attitude to technology. On the other hand, there are some design principles that are useful for developping software for other populations.

The book refers to Nielsen's 10 Usability Heuristics for Application Development. They discuss how vision, hearing, motoric skills and cognition is changing when people get richer in years. Based on those factors, the book presents some recommendations and examples of good and bad design.

TrafficControl isn't explicitly designed for any age group, but many of the recommendations in the book can be applied to a good UI design anyways.

The book also discusses some different tech generations.

As I was born in 1978, I belong to the "Digital computer" generation. Of course, I can use the Internet and social media, but that is something that I've learnt after being a child.

I'll use some of the advices from the book when designing my pet projects.