Alchemists never succeeded at transmuting base metals into gold. Their
modern counterparts may be more successful in their efforts to produce not
gold but brainpower. They focus on silicon rather than a base metal, but
their efforts have been in vain.

These modern alchemists, the artificial intelligence community, do not
use potions but incantations of predicate logic, Lisp and the lambda calculus.
The Age of Intelligent Machines is a celebration of progress to date, the
machines and their inventors. Its glossy appearance and plentiful illustrations
give it the look of a coffee-table book, but do not detract from the consistently
high quality of its content.

The phrase ‘artificial intelligence’ generates as much hype and hot
air as it does progress in computation. It also seems to demand that those
who work in the field ponder the nature of ‘real’ intelligence. But artificial
intelligence rarely concerns the making of machines that share our intellectual
preoccupation. Instead, a popular definition is ‘building devices to do
things that usually require people’.

There are many such machines, doing good works and making people money.
Raymond Kurzweil, editor of this book, for example, surely deserves the
success of his machines that read books to the blind. The Age of Intelligent
Machines celebrates precisely this kind of engineering, such as robots that
spray paint in places no human should ever have to enter. Grappling with
‘real thoughts’, then, need not be a central concern of workers in artificial
intelligence. It is hard enough making simpler programs work.

The book emphasises a traditional symbolist’s approach to artificial
intelligence of programming, logic and plans, though recent advances in
neural networks receive a mention. The burgeoning field of artificial life,
however, is not included. This is a pity, as it promises significant contributions
to the field of robotics through the paradigm of behavioural robotics. Traditional
robots have been successful, in limited areas, because their precise programs
make them reliable. The problem in widening their scope is that most of
the world is far from reliable.

Until recently, received wisdom was that effective robots must have
a powerful reasoning system, with sufficient memory and processing power
to build detailed models of the world in which they operate. This is being
challenged.

As alternatives to a robot that dutifully and stupidly does what it
is told, other approaches are proving more robust. They are based on models
for artificial life such as ‘swarm intelligence’. Instead of breaking down
complex tasks into a fixed sequence of steps conducted in a carefully modelled
world, the robot possesses a set of simple rules for behaviour. These are
applied singly or in combination in response to environmental stimuli. This
is a new paradigm: complex behaviour may stem from complex environments
rather than complex programs.

Perhaps we can allow Kurzweil his speculative chronology about tomorrow’s
technology, with which he closes the book because he has been so successful.
Forecasting the future of technologies is a fraught business – I know, it
is my job. One prediction in Kurzweil’s list that I doubt will be achieved
fully is that by the early 2000s we shall have ‘translating telephones (that)
allow two people across the globe to speak to each other even if they do
not speak the same language’.

This is an energetic and proselytising book that deserves to be read.
It is nonspecialist, and its readers require no programming skills nor a
degree in logic.

Specialists will nevertheless enjoy the many contributions from major
figures in the history of the discipline, and the pleasing photographs.
It is good to put faces to some of the names that dominate the literature
of artificial intelligence. One of the most appealing photographs appears
in the chapter ‘Electronic Roots’, a description of the struggles of early
programmers to debug software, is followed by a picture from the notebook
of Grace Hopper, developer of the programming language Cobol, who died this
year. ‘First actual case of bug being found,’ reads her note, beside a dead
moth stuck to the page after being found in relay number 70.

Nick Beard is an information technology consultant with Price Waterhouse.

In the tangle of philosophy, psychology and technology that is cognitive
science, many disciplines merge. This is a field in a state of flux.

Cognitive science includes a combination of brain physiology, computing
and mathematics with a smattering of statistical physics. This pick-and-mix
origin has led to the development of some powerful techniques, many of which
are being used to solve real-world problems and make people money. But here
a split emerges: is the purpose of artificial intelligent research to uncover
principles of human cognition? Or is it to advance technology?

A primary task of cognitive science is the understanding of cognition
from a computational perspective, defining what computational problems are
being solved. We also have to decide which computational mechanisms (algorithms)
are being used in cognition and how brain physiology implements them. All
these questions have valuable technological implications beyond their contribution
to the study of human reason.

Cognitive science, then, hopes to conduct experiments, where ethics
and biology conspire to prevent them. The next best thing is simulation
– modelling brains with computers.

The results of this approach have been impressive, and Foundations of
Cognitive Science offers a thorough grounding in the whole field. The two
principal strands of cognitive science research, the symbolic and the connectionist,
are both given a thorough treatment. The exponents of the symbolic approach
argue that intelligence and its manifestations are a consequence of the
logical manipulation of symbols – inner representations of problems, objects,
words and problems that are juggled and fumbled to produce answers. The
alternative is connectionism, based on the idea of emergent properties of
networks of nerve cells, derived from the study of computer simulated neural
networks.

After an introductory survey by Herbert Simon and Craig Kaplan, the
first part covers the basics, setting the scene for more specific studies
of domains, such as language acquisition, discourse, concepts and induction,
vision and memory. Finally, assessment includes cultural cognition, an assessment
of the impact of this research on anthropology and vice versa, and a more
philosophical chapter, tastefully tucked away at the end so that those who
cannot face the mind-body problem again can ignore it.

Although it is a survey of the foundations of the subject, Foundations
of Cognitive Science is not beginner’s guide. The content is advanced, and
the target readership is likely to be undergraduate level and beyond. It
contains some light logic, though never enough to seriously intimidate the
asymbolic reader. The book includes contributions from many of the leading
authorities in the field, including Zenon Pylyshyn, David Rumelhart, Terrence
Sejnowski, Steven Pinker, Phillip Johnson-Laird and Shimon Ullman. The result
is a beautifully lucid account, and is highly recommended.

Nick Beard trained in medicine and psychiatry before joining the health
IT group at Coopers & Lybrand Deloitte.

The Computing Neuron edited by Richard Durbin, Christopher Miall and
Graeme Mitchinson, Addison-Wesley, pp 417, Pounds sterling 25.95

Parallel Distributed Processing edited by R. G. M. Morris, Oxford, pp
339, Pounds sterling 35

SOMEWHERE between neurotransmitters and neuroanatomy, cognition arises.
There remains much debate, however, about precisely where – and how – to
look for it, and indeed what manner of thing it might be. For example, John
Searle, a philosopher, has expressed doubts about the whole enterprise we
call cognitive science, though he concedes that neural network modelling
looked promising.

The renaissance during the past 10 years of simulations of neural networks
has spawned a new area of research based on computer modelling techniques,
for the study of the human nervous system. It is known as connectionism,
or parallel-distributed processing, after McClelland and Rumelhart’s seminal
text (published by MIT Press in 1986) of the same title.

The behaviour of large assemblies of nerve cells or neurons is not easy
to study experimentally, so researchers often need to develop a model to
understand what is happening. The pioneering work of Peter Getting demonstrates
this. Studying the sea slug Tritonia diomedea, Getting attempted to establish
the mechanisms of neural control for its rhythmic swimming. Unlike the huge
numbers of neurons in human heads, Tritonia has only 14. These are larger
than mammalian nerve cells, so he found it easier to insert probes to monitor
the activity of the slug nerve cells, and thus link neural activity directly
to behaviour. Getting was to discover, however, that even with small numbers
of ‘components’ (14 neurons), understanding the sea slug circuitry was very
difficult. Despite his structural knowledge of the network, it was not obvious
how it worked. Not until researchers made detailed models could they understand
the circuit. This has led to attempts to replicate the network in silicon.

Like the brain, simulated webs of computationally simple units can learn,
classify and remember. In such models, each unit roughly corresponds to
a single neuron. The abstract mathematical description of the behaviour
of these models owes much to the physics community. In particular, it was
John Hopfield, a physicist, who demonstrated the analogy between neural
network models and spin glass physics.

Daniel Amit developed Hopfield’s analysis further and solved some of
the key equations in Hopfield’s work. Amit’s book, Modelling Brain Function,
summarises the technical and mathematical results of recent theoretical
research into neural tests. He concentrates on attractor networks, a subset
of the wider scope of connectionist models. These are feedback nets, which
are more amenable to fuller mathematical analysis than the more commonly
used ‘input output’ networks.

Amit sets himself a demanding goal: to present the abstract elegance
of net theory plus the required background. There is much pedagogical material
on general ideas of dynamics, a swift but thorough tour of attractor networks,
and surveys of more speculative application areas, such as psychiatry. For
example, he describes Hoffman’s attempts to characterise differences between
manic and schizophrenic speech using attractor models. Amit’s work targets
mathematical rigour before biological accuracy, but aims to move ever closer
to biological robustness. It is demanding, especially for readers less familiar
with mathematically flavoured texts.

A quick survey of elementary connectionism and elementary neurobiology
reveals a substantial gap between the computational capabilities of the
‘neurons’ of artificial neural systems, and those of real nerve cells. This
is the topic of The Computing Neuron, the proceedings of a 1989 Cambridge
conference on ‘The neuron as a computational unit’. As the editor Richard
Durbin points out, we have only a sparse knowledge of neurophysiological
detail, but what we do know still fuels inappropriate attacks on, and defences
of, neural models. This book, deliberately interdisciplinary, contains 21
papers that attempt to capture some of the available correspondence between
neurons and their models. In spite of the title, the volume is less concerned
with the properties of single neurons than with the behaviour of networks.
Alan Roberts, for example, discusses a mechanism for turning swimming on
and off in tadpoles. The book is likely to be of less interest to computer
scientists than to neurobiologists, the emphasis being on real rather than
simulated neurons.

Similarly biological, but with greater emphasis on psychological models,
is Parallel Distributed Processing, edited by Richard Morris. The book splits
into three parts: formal models, which is an overview of connectionist modelling;
implications for psychology; and implications for neurobiology. The book
emerged from an Oxford meeting of the Experimental Psychology Society, at
which the connectionist paradigm was assessed. Its tone is more critical
than the other two books, being concerned with the evaluation of the entire
paradigm than presentation of specific technical results.

It is the study of psychology, rather than the anatomical and physiological
substrates of thought, that benefits most from connectionist modelling.
Of particular note in this realm is the Rumelhart-McClelland connectionist
model of language acquisition, which is said to account successfully for
first-language acquisition in children. It is refreshing to read a vigorously
sceptical analysis, such as that included here from Steven Pinker and Alan
Prince. They argue quite simply that the mechanisms invoked by the Rumelhart-McClelland
model are inconsistent with available psycholinguistic data.

These three books present a different perspective on what has been described
as ‘neural network revolution’. The paradigm has advanced since McClelland
and Rumelhart’s ‘connectionist bible’ appeared; these volumes will bring
anyone up to date.

Nick Beard is an information consultant with Price Waterhouse.

He proposed a digital computing architecture, a logical arrangement
of the operating system, so successful that it is rare to find computers
based on anything else. Machines with a Von Neumann architecture have a
central processing unit that operates on data and instructions taken from
memory locations that are explicitly labelled.

Some tasks, however, prove difficult for computers with this kind of
operating system. Paradoxically, these often include things that people
do well, easily and with little or no training. The way in which people
can recall names, faces, perfumes and songs, for example, when remembering
a failed romance from long ago, is almost impossible to replicate with a
computer. This skill is called associative memory. Something trivial, such
as a phrase from a song, may trigger a long chain of detailed memories.
It is just one of the impressive capabilities of a different style of computing
based on networks of artificial neuron-like processing units.

Densely tangled webs of these computationally crude units act in concert
to display many of the features traditionally found only in physiological
systems – brains. Associative (sometimes called content-addressable) memory,
pattern recognition, learning by example and tolerance of ‘noisy’ (inexact)
data or partial failure of a processor are all possible with artificial
neural systems.

Software is available that simulates neural networks on standard sequential
computers such as the PC or Mac. You can use these programs either as research
tools or as commercial products to solve business problems. They offer opportunities
for researchers in the study of brain circuitry and psychology, as well
as having technological value in fields as diverse as speech processing,
image recognition, which includes patterns of data in mortgage applications
or the hand writing on cheques; and many other uses.

Parallel Distributed Processing, or PDP, is effectively synonymous with
neurocomputing or connectionism. So the publication of James McClelland
and David Rumelhart’s Parallel Distributed Processing a few years ago contributed
much to the widespread popularity of this field. The authors are leading
authorities on neurocomputing, and Explorations in Parallel Distributed
Processing is a collection of software to accompany their books. It provides
an excellent and affordable introduction to an exciting field.

The authors offer numerous different models of networks. For those already
familiar with neurocomputing, the programs include interactive competition
and learning, Boltzmann machines, a pattern associator, a competitive learning
system, and a back prop net.

The programs also include demonstrations of some classic neurocomputing
problems in neurocomputing, and C source code. The provision of C allows
the programs to be altered or re-compiled for a different type of computer.
The authors give details of all file formats to encourage users to experiment
with, adapt and customise the programs. The package is academic software,
intended primarily for didactic purposes, but you can use it on real problems.
I have used PDP to construct a net that I trained to recognise certain abnormal
electrocardiograms. But you do need an ASCII text editor to write small
files to control the networks, and the mechanics of this task requires some
computing experience. These are real, working nets, at minimum cost, and
I highly recommend them – even if they are not for the complete computing
novice.

More expensive, but far more versatile and sophisticated, is NeuralWorks
Professional II. This is an integrated suite of networks, in a powerful
development environment. The whole package is easy to install and is driven
by both mouse and menu. It also includes a wide selection of configurable
networks. The package includes Adaptive Resonance Theory (ART) and Kohonen
nets, Boltzmann machines, Hopfield, Adaline and probabilistic neural nets
(PNNs) and many implementations of back prop nets, as well as the recently
developed functional link network.

Data input is flexible as the package accepts several different file
formats. NeuralWorks can also be used to generate stand-alone network systems
that do not then require the package. Using the NeuralWorks Designer Pack,
you can also produce networks with NeuralWorks and then turn your work into
C source code to use with a standard C compiler.

The documentation provided with NeuralWorks is excellent, a thorough
introduction to neural networks, as well as to the software. This is the
best neurocomputing software on the market, and will satisfy anyone who
is interested in solving real problems – business or scientific – using
neural networks.

In a different vein comes Brain Simulator. This program allows you to
design networks on screen, and specify the types of connections between
them, but it has none of the powerful multi-layered network training routines
of the workhorse nets of the other two packages. The way that it allows
users to experiment with circuits is akin to the methods used in the early
neural network studies by McCulloch and Pitts, who made the first serious
attempts to simulate brain circuits electrically in the 1940s. Brain Simulator
is easy to use, and the designers offer sample nets that make it easy to
speculate about the possible logical building blocks of cognition. It is
of less value as a research tool, but could be popular in biology or computing
departments in schools and colleges.

Neurocomputing is not about to replace conventional digital systems,
but will doubtless supplement them in certain areas. These packages are
all useful at the different stages along the route to understanding this
fast developing field.

I tested these products on a 20Mhz Elonex 386 IBM compatible PC.

Explorations in Parallel Distributed Processing, MIT Press 01-730 9200;
IBM PC version Pounds sterling 21.95; Mac Pounds sterling 29.95. Hardware
required: IBM compatible, PC; 256K RAM, DOS 1.0 or above.

NeuralWorks Professional II Costs Pounds sterling 16.50 plus VAT from
Recognition Research, 140 Church Lane, Marple, Stockport SK6 7LA. 061

449 8628. Hardware: PC: 512K RAM, DOS 3.0 or above, CGA, EGA or VGA
screen. Hard disc. Also available for Mac, Sun and Neube.

Brain Simulator Abbot Foster & Hausermann Co, 0101 (509) 458 4660
(USA), 44 Montgomery Street, 5th Floor, San Francisco, CA 94104. Hardware:
IBM compat, 240K free RAM, DOS 3.0 or above. $99 plus $10 for shipment outside
US.

Nick Beard works in the knowledge of engineering group at management
consultants Coopers & Lybrand.

Prolog is a language for writing fixed logical laws about problems,
plus a mechanism that deduces new knowledge from those laws autonomously.
It splits an algorithm into ‘logic + control’. The programmer specifies
the logic of the problem, the computer controls the exploration of that
logic. It has a deduction method, built in, called resolution. If you give
Prolog a detailed description of the problem, it solves it without any further
directions from you. Of course, there are limitations – the problem must
be stated in a particular style. However, the style is flexible, and for
a certain class of problems, invests tremendous power in the language.

Prolog allows programmers to state in a few short sentences what would
take many lines of code in a conventional language. It is based on first-order
or classical logic. Said to be the language of formal reasoning, this sets
programming on a rigorous footing: programs written in Prolog resemble statements
of this logic. Prolog was developed in the 1970s by Alain Colmerauer in
Marseilles. Robert Kowalski subsequently enhanced the language and gave
it a thorough underpinning of logic.

There are now several versions of the language. Turbo Prolog comes from
the Borland stable. It provokes combined attacks of angst and snobbery among
the ‘true Prolog’ community – it is not a ‘proper’ Prolog. This response
is not simply academic grandeur; Prolog has capabilities not available in
this version. Users never experience Prolog’s real power with Turbo Prolog.
The principal difference is that Turbo Prolog demands strict types of data,
whereas standard Prolog can use data of any type. This has important effects
on the scope of the programs that can be written.

In its favour, Turbo Prolog has an easy ‘user-interface’, in which programs
are constructed. More professional software can display the computer’s legendary
unfriendliness at its worst. The package comes with a pair of thick and
well-written manuals, and a number of sample programs and tutorials.

Logic Programming Associates produced another version of Prolog. The
designers have their roots firmly placed in Imperial College London, the
home of much research into this area of computing. LPA Prolog is a highly
professional implementation of the language, aimed at serious users. It
includes powerful debugging facilities, and links to allow C programs to
be interfaced directly to Prolog applications. Reflecting perhaps its serious
origins and aspirations, it is less immediately friendly than Turbo Prolog.

Rather than have you learn a new text editor to create and edit programs,
LPA includes a routine which enables you to call your existing editor from
within Prolog. Other facilities offered include: tools to assist in natural
language processing; a common prolog application area; and advanced graphics
routines. It also provides a set of ‘humancomputer interface’ tools.

Prolog is notoriously greedy with memory and LPA Prolog comes with support
for an expanded memory. The designers are developing a version that takes
full advantage of the Intel 386 chip, now in many faster PCs. LPA Prolog
is also available for the Apple Macintosh, and for larger machines.

LPA also developed a ‘frame-based’ system, written in Prolog, call Flex,
which has become the de facto standard for this approach. It is designed
to aid development of expert systems.

Those with a penchant for desk-vermin would welcome mouse support, and
online help would have been nice. Generally, this is a high quality implementation
of Prolog, which it is hard to fault.

SD Prolog is not as fast as LPA Prolog but an easier environment to
work in. It was developed by Quintec, which also produces an advanced version
of Prolog for workstations. It is an obviously professional Prolog that
combines the ease of a menu-driven user-interface with heavy-duty programming
tools.

SD Prolog has an impressive set of help screens, which provide thorough
documentation for the system on line. Apart from the basic Prolog system,
which arrives on three floppies, Quintec also provides a set of ‘man-machine
interface utilities.’ These enable the simple construction of menu-driven
programs, to make Prolog programs more user-friendly. Graphics routines
are included, and interface packages to enable Prolog programs to be linked
to dBASE and other database programs. Quintec also provides a tool for program
analysis, called Xref, which checks programs for syntactic errors.

Quintec includes Analyst, a demonstration program that is a version
of the famous ‘artificial-intelligence’ program Eliza.

Turbo Prolog is from Borland International. Telephone: 0734 320022.
It costs Pounds sterling 99.95 + VAT.

LPA Prolog comes from LPA Ltd. Tel: 01-871 2016. LPA Prolog Professional
Pounds sterling 495 + VAT. Full system, with Flex Pounds sterling 1485.
Academic discounts available, negotiable between 25 per cent and 50 per
cent. System requirements: IBM compatible PC, with at least 512K RAM, and
twin floppies. Full 640K RAM + hard disc drive recommended.

SD (Quintec) Prolog Quintec Systems Ltd. Tel: 0865 791565. System requirements:
IBM compatible PC, with at least 384K RAM and dual floppy drives. 512K RAM
and hard disc recommended. Pounds sterling 799, academic discounts negotiable.

Nick Beard is in the knowledge engineering group at Coopers & Lybrand.