Econstudentlog

Artificial intelligence (I?)

This book was okay, but nothing all that special. In my opinion there’s too much philosophy and similar stuff in there (‘what does intelligence really mean anyway?’), and the coverage isn’t nearly as focused on technological aspects as e.g. Winfield’s (…in my opinion better…) book from the same series on robotics (which I covered here) was; I am certain I’d have liked this book better if it’d provided a similar type of coverage as did Winfield, but it didn’t. However it’s far from terrible and I liked the authors skeptical approach to e.g. singularitarianism. Below I have added some quotes and links, as usual.

“Artificial intelligence (AI) seeks to make computers do the sorts of things that minds can do. Some of these (e.g. reasoning) are normally described as ‘intelligent’. Others (e.g. vision) aren’t. But all involve psychological skills — such as perception, association, prediction, planning, motor control — that enable humans and animals to attain their goals. Intelligence isn’t a single dimension, but a richly structured space of diverse information-processing capacities. Accordingly, AI uses many different techniques, addressing many different tasks. […] although AI needs physical machines (i.e. computers), it’s best thought of as using what computer scientists call virtual machines. A virtual machine isn’t a machine depicted in virtual reality, nor something like a simulated car engine used to train mechanics. Rather, it’s the information-processing system that the programmer has in mind when writing a program, and that people have in mind when using it. […] Virtual machines in general are comprised of patterns of activity (information processing) that exist at various levels. […] the human mind can be understood as the virtual machine – or rather, the set of mutually interacting virtual machines, running in parallel […] – that is implemented in the brain. Progress in AI requires progress in defining interesting/useful virtual machines. […] How the information is processed depends on the virtual machine involved. [There are many different approaches.] […] In brief, all the main types of AI were being thought about, and even implemented, by the late 1960s – and in some cases, much earlier than that. […] Neural networks are helpful for modelling aspects of the brain, and for doing pattern recognition and learning. Classical AI (especially when combined with statistics) can model learning too, and also planning and reasoning. Evolutionary programming throws light on biological evolution and brain development. Cellular automata and dynamical systems can be used to model development in living organisms. Some methodologies are closer to biology than to psychology, and some are closer to non-reflective behaviour than to deliberative thought. To understand the full range of mentality, all of them will be needed […]. Many AI researchers [however] don’t care about how minds work: they seek technological efficiency, not scientific understanding. […] In the 21st century, […] it has become clear that different questions require different types of answers”.

“State-of-the-art AI is a many-splendoured thing. It offers a profusion of virtual machines, doing many different kinds of information processing. There’s no key secret here, no core technique unifying the field: AI practitioners work in highly diverse areas, sharing little in terms of goals and methods. […] A host of AI applications exist, designed for countless specific tasks and used in almost every area of life, by laymen and professionals alike. Many outperform even the most expert humans. In that sense, progress has been spectacular. But the AI pioneers weren’t aiming only for specialist systems. They were also hoping for systems with general intelligence. Each human-like capacity they modelled — vision, reasoning, language, learning, and so on — would cover its entire range of challenges. Moreover, these capacities would be integrated when appropriate. Judged by those criteria, progress has been far less impressive. […] General intelligence is still a major challenge, still highly elusive. […] problems can’t always be solved merely by increasing computer power. New problem-solving methods are often needed. Moreover, even if a particular method must succeed in principle, it may need too much time and/or memory to succeed in practice. […] Efficiency is important, too: the fewer the number of computations, the better. In short, problems must be made tractable. There are several basic strategies for doing that. All were pioneered by classical symbolic AI, or GOFAI, and all are still essential today. One is to direct attention to only a part of the search space (the computer’s representation of the problem, within which the solution is assumed to be located). Another is to construct a smaller search space by making simplifying assumptions. A third is to order the search efficiently. Yet another is to construct a different search space, by representing the problem in a new way. These approaches involve heuristics, planning, mathematical simplification, and knowledge representation, respectively. […] Often, the hardest part of AI problem solving is presenting the problem to the system in the first place. […] the information (‘knowledge’) concerned must be presented to the system in a fashion that the machine can understand – in other words, that it can deal with. […] AI’s way of doing this are highly diverse.”

“The rule-baed form of knowledge representation enables programs to be built gradually, as the programmer – or perhaps an AGI system itself – learns more about the domain. A new rule can be added at any time. There’s no need to rewrite the program from scratch. However, there’s a catch. If the new rule isn’t logically consistent with the existing ones, the system won’t always do what it’s supposed to do. It may not even approximate what it’s supposed to do. When dealing with a small set of rules, such logical conflicts are easily avoided, but larger systems are less transparent. […] An alternative form of knowledge representation for concepts is semantic networks […] A semantic network links concepts by semantic relations […] semantic networks aren’t the same thing as neural networks. […] distributed neural networks represent knowledge in a very different way. There, individual concepts are represented not by a single node in a carefully defined associative net, but by the changing patterns of activity across an entire network. Such systems can tolerate conflicting evidence, so aren’t bedevilled by the problems of maintaining logical consistency […] Even a single mind involves distributed cognition, for it integrates many cognitive, motivational, and emotional subsystems […] Clearly, human-level AGI would involve distributed cognition.”

“In short, most human visual achievements surpass today’s AI. Often, AI researchers aren’t clear about what questions to ask. For instance, think about folding a slippery satin dress neatly. No robot can do this (although some can be instructed, step by step, how to fold an oblong terry towel). Or consider putting on a T-shirt: the head must go in first, and not via a sleeve — but why? Such topological problems hardly feature in AI. None of this implies that human-level computer vision is impossible. But achieving it is much more difficult than most people believe. So this is a special case of the fact noted in Chapter 1: that AI has taught us that human minds are hugely richer, and more subtle, than psychologists previously imagined. Indeed, that is the main lesson to be learned from AI. […] Difficult though it is to build a high-performing AI specialist, building an AI generalist is orders of magnitude harder. (Deep learning isn’t the answer: its aficionados admit that ‘new paradigms are needed’ to combine it with complex reasoning — scholarly code for ‘we haven’t got a clue’.) That’s why most AI researchers abandoned that early hope, turning instead to multifarious narrowly defined tasks—often with spectacular success.”

“Some machine learning uses neural networks. But much relies on symbolic AI, supplemented by powerful statistical algorithms. In fact, the statistics really do the work, the GOFAI merely guiding the worker to the workplace. Accordingly, some professionals regard machine learning as computer science and/or statistics —not AI. However, there’s no clear boundary here. Machine learning has three broad types: supervised, unsupervised, and reinforcement learning. […] In supervised learning, the programmer ‘trains’ the system by defining a set of desired outcomes for a range of inputs […], and providing continual feedback about whether it has achieved them. The learning system generates hypotheses about the relevant features. Whenever it classifies incorrectly, it amends its hypothesis accordingly. […] In unsupervised learning, the user provides no desired outcomes or error messages. Learning is driven by the principle that co-occurring features engender expectations that they will co-occur in future. Unsupervised learning can be used to discover knowledge. The programmers needn’t know what patterns/clusters exist in the data: the system finds them for itself […but even though Boden does not mention this fact, caution is most definitely warranted when applying such systems/methods to data (..it remains true that “Truth and true models are not statistically identifiable from data” – as usual, the go-to reference here is Burnham & Anderson)]. Finally, reinforcement learning is driven by analogues of reward and punishment: feedback messages telling the system that what it just did was good or bad. Often, reinforcement isn’t simply binary […] Given various theories of probability, there are many different algorithms suitable for distinct types of learning and different data sets.”

“Countless AI applications use natural language processing (NLP). Most focus on the computer’s ‘understanding’ of language that is presented to it, not on its own linguistic production. That’s because NLP generation is even more difficult than NLP acceptance [I had a suspicion this might be the case before reading the book, but I didn’t know – US]. […] It’s now clear that handling fancy syntax isn’t necessary for summarizing, questioning, or translating a natural-language text. Today’s NLP relies more on brawn (computational power) than on brain (grammatical analysis). Mathematics — specifically, statistics — has overtaken logic, and machine learning (including, but not restricted to, deep learning) has displaced syntactic analysis. […] In modern-day NLP, powerful computers do statistical searches of huge collections (‘corpora’) of texts […] to find word patterns both commonplace and unexpected. […] In general […], the focus is on words and phrases, not syntax. […] Machine-matching of languages from different language groups is usually difficult. […] Human judgements of relevance are often […] much too subtle for today’s NLP. Indeed, relevance is a linguistic/conceptual version of the unforgiving ‘frame problem‘ in robotics […]. Many people would argue that it will never be wholly mastered by a non-human system.”

“[M]any AI research groups are now addressing emotion. Most (not quite all) of this research is theoretically shallow. And most is potentially lucrative, being aimed at developing ‘computer companions’. These are AI systems — some screen-based, some ambulatory robots — designed to interact with people in ways that (besides being practically helpful) are affectively comfortable, even satisfying, for the user. Most are aimed at the elderly and/or disabled, including people with incipient dementia. Some are targeted on babies or infants. Others are interactive ‘adult toys’. […] AI systems can already recognize human emotions in various ways. Some are physiological: monitoring the person’s breathing rate and galvanic skin response. Some are verbal: noting the speaker’s speed and intonation, as well as their vocabulary. And some are visual: analysing their facial expressions. At present, all these methods are relatively crude. The user’s emotions are both easily missed and easily misinterpreted. […] [An] point [point], here [in the development and evaluation of AI], is that emotions aren’t merely feelings. They involve functional, as well as phenomenal, consciousness […]. Specifically, they are computational mechanisms that enable us to schedule competing motives – and without which we couldn’t function. […] If we are ever to achieve AGI, emotions such as anxiety will have to be included – and used.”

[The point made in the book is better made in Aureli et al.‘s book, especially the last chapters to which the coverage in the linked post refer. The point is that emotions enable us to make better decisions, or perhaps even to make a decision in the first place; the emotions we feel in specific contexts will tend not to be even remotely random, rather they will tend to a significant extent to be Nature’s (…and Mr. Darwin’s) attempt to tell us how to handle a specific conflict of interest in the ‘best’ manner. You don’t need to do the math, your forebears did it for you, which is why you’re now …angry, worried, anxious, etc. If you had to do the math every time before you made a decision, you’d be in trouble, and emotions provide a great shortcut in many contexts. The potential for such short-cuts seems really important if you want an agent to act intelligently, regardless of whether said agent is ‘artificial’ or not. The book very briefly mentions a few of Minsky’s thoughts on these topics, and people who are curious could probably do worse than read some of his stuff. This book seems like a place to start.]

Links:

GOFAI (“Good Old-Fashioned Artificial Intelligence”).
Ada Lovelace. Charles Babbage. Alan Turing. Turing machine. Turing test. Norbert WienerJohn von Neumann. W. Ross Ashby. William Grey Walter. Oliver SelfridgeKenneth Craik. Gregory Bateson. Frank Rosenblatt. Marvin Minsky. Seymour Papert.
A logical calculus of the ideas immanent in nervous activity (McCulloch & Pitts, 1943).
Propositional logic. Logic gate.
Arthur Samuel’s checkers player. Logic Theorist. General Problem Solver. The Homeostat. Pandemonium architecture. Perceptron. Cyc.
Fault-tolerant computer system.
Cybernetics.
Programmed Data Processor (PDP).
Artificial life.
Forward chaining. Backward chaining.
Rule-based programming. MYCIN. Dendral.
Semantic network.
Non-monotonic logic. Fuzzy logic.
Facial recognition system. Computer vision.
Bayesian statistics.
Helmholtz machine.
DQN algorithm.
AlphaGo. AlphaZero.
Human Problem Solving (Newell & Simon, 1970).
ACT-R.
NELL (Never-Ending Language Learning).
SHRDLU.
ALPAC.
Google translate.
Data mining. Sentiment analysis. Siri. Watson (computer).
Paro (robot).
Uncanny valley.
CogAff architecture.
Connectionism.
Constraint satisfaction.
Content-addressable memory.
Graceful degradation.
Physical symbol system hypothesis.

January 10, 2019 - Posted by | Biology, Books, Computer science, Engineering, Language, Mathematics, Papers, Psychology, Statistics

No comments yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: