ANIMAL AND MACHINE INTELLIGENCE
Autumn 2001
Office: 3B18 Biols
Tel: 01273-678507
E-mail: T.S.Collett@sussex.ac.uk
AIMS AND OBJECTIVES
The overall aim of this course is to develop your understanding of what
it means for an animal or a machine to behave intelligently, and how brain and
behavioural systems are adapted to enable an animal to cope effectively within
its environment. You explore this topic in lectures and seminars through a
number of case studies that are designed to acquaint you with recent
behavioural and AI literature.
COURSE OUTLINE
Animals engage in a variety of complex behaviours that are intelligent
in the sense of being well-adapted to particular situations. Thus, intelligence
is not a unitary phenomenon, rather animals have multiple, specialised
intelligences that are designed for particular tasks. Animals go in for
computational economy. This means that they usually do not build up detailed
internal models of the world, but rather use sensory information on-line to
guide actions.
Developments in behavioural
neuroscience and in artificial intelligence over the last decade have led to
some convergence of the aims and methods of the two disciplines. In studies of
behaviour, there is increasing acceptance of robot and computer modelling as
experimental tools. In artificial intelligence and cognitive science, the focus
has shifted from high-level human intellectual capabilities toward detailed
studies of the basic behaviours that are common to most animals, and that are
required of autonomous robots. Our interdisciplinary course reflects this new
movement and aims to show how the two fields can inform each other, to give an
understanding of the subtleties of cognition in simple animals and the
challenges faced by scientists trying to create artificial systems with the
same behavioural capabilities as these animals.
We begin by discussing the evolution of views on animal intelligence and
robot‑building. The modern approach emphasises complete systems acting in
the real world. This aims to avoid the pitfalls of earlier methods, even if it
means confining ourselves to relatively simple systems Recent experiments are
used to illustrate the developing methodology by which computer simulations and
robotic experiments can stimulate biology (and vice‑versa).
We will consider in detail several examples of ‘specialised or adaptive
intelligences’ in invertebrates and vertebrates. The study of relatively simple
animals (such as insects) turns out to be challenging rather than restrictive.
Their capabilities often exceed what can be achieved in artificial systems or
fully understood by biologists. We concentrate on navigation and foraging, and
we will examine the behavioural and memory strategies that insects and other
animals have evolved to accomplish these tasks.
Difficulties in controlling behaviour can be lessened by appropriate
design of sense organs and effectors, and by exploiting properties of the
environment in which the animal/machine operates. A telling example is the way
in which the insect visual system has adapted to interpret the visual
consequences of an animal’s own movements. We look at optic flow in insects and
the specialised neural mechanisms that analyse it.
The synaptic computations performed by real neurones in the insect
visual system serve as an introduction to artificial neural networks (ANNs).
ANNs are computer simulations or electronic circuits, at least partially based
on observations of the structure and operation of animals' nervous systems. As
we will see, they turn out to be very powerful computational devices for
prediction, categorisation and learning and are becoming widely used in
engineering and industry. They have also revolutionised our understanding of
the brain in that appropriately configured and trained neural nets can in many
cases substitute for explanations of behaviour that are cast in terms of rules
and symbol manipulation.
We next consider how animals, without the aid of
language, recognise and categorise objects and form concepts. To what extent
can neural net models mimic these abilities?
Intelligence does not only exist at the individual level.
Some tasks demand group co-ordination and intelligent algorithms can be
implemented at the level of a group. A different kind of social intelligence is
when the competence of individuals is enhanced through knowledge obtained from
others. The course ends with a look at three aspects of social intelligence: i)
the social organisation of foraging in bees; ii) culture and observational
learning; iii) social intelligence, mind reading and the evolution of big
brains.
AMI is open both to final-year undergraduates and to COGS MSc students.
Teaching is through lectures, seminars and an essay. All students should attend
all lectures. However, there are separate seminars for undergraduate and
postgraduate students. Lectures can
only introduce a topic and you are expected to develop knowledge and analytical
skills through reading critically the starred paper and a selection of the
other recommended items. It is especially important that you integrate the
different topics of the course. Again, lectures can only give pointers,
developing a conceptual framework of your own is something that requires both
individual thought and discussion with friends and colleagues. Seminars encourage group discussion of
critical issues and are for you to raise and discuss problems that you have
with course material. The essay has
both a learning and an assessment role. It should teach you to organise a large
body of material, or to analyze a problem, in a logical, interesting and
concise manner. We urge you to write a detailed plan or outline of your essay
before turning it into continuous prose and we are more than happy to provide
feedback at that stage.
LECTURERS
Adrian Thompson, adrianth@cogs.sussex.ac.uk
Tom Collett, BIOLS 3B18,
t.s.collett@sussex.ac.uk
Daniel Osorio, BIOLS 3B32, d.osorio@sussex.ac.uk
Ezequiel Di Paolo, ezequiel@cogs.sussex.ac.uk (advisor to MSc students)
Do not hesitate to come with any queries. Use e-mail, if you prefer.
LECTURES
Two per week, weeks 1 to 10: Wednesday at 10:15 in the Biology lecture
theatre; Friday at 15:00 in the Biology lecture theatre. First lecture is on Wednesday 10th October.
Undergraduates:
There are 7 seminars in the term
(i.e. not every week). The weeks in which they occur
are shown in the course time table below.
Group 1: Wednesday 12:30 in PEV1 2C1
Group 2: Thursday 17:00 in PEV1 2A1
Group 3: Friday 9:15 in PEV1 2A1
Seminar topics with suggested
readings and questions for discussion are listed towards the end of the
handout. To make the most of seminars, you need to come prepared having read
and pondered the material!
Postgraduates:
Seminars are weekly. They start in week 1.
Group 1: Thursday 16:00 PEV1 1B6
Group 2: Friday, 10:15 ARUN 404B
Group 3: Friday 14:00 ARUN 1C
Group 4: Friday 16:00 ARUN 1D
There will be an unseen exam that contributes 85%, and
one 3000 (+/- 20%) word assessed essay, contributing 15% of the total course
marks.
Essays must be handed in to the
Biology Subject Office (BIOLS Room 4B 13) between 9.00 am and 5.00 pm on
Thursday 10th January 2002. Please be sure to put your name and exam
number on the essay and to sign the list. Unless there is good reason for late
submission, 10% of the possible marks will be deducted for essays that are
handed in on Friday 12th January. No marks will be given if essays
are any later. This regulation is imposed by the Neuroscience exam board.
Excerpt
from exam board rules: “Any assessed work, which is late due to ill health or
other mitigating circumstances, MUST be submitted directly to the course
organizer with a written explanation of the circumstances and a photocopy of
the Medical Certificate (not self-certificated) or other evidence. This will be passed on to the Neuroscience
Sub-Board for consideration in due course. “
Sample essay titles are provided at the end of this handout. If you prefer, you may choose your own
title, but it should be approved by one of the lecturers. We recommend strongly
that you give a detailed plan of this essay to one of the lecturers well before
the end of this term so that we have time to provide feedback.
Formal assessment is by submission of a short term paper
(max. 3500 words) due by 12 noon on the Monday, week 2 of the Spring Term (Jan 14).
There isn't a suitable textbook. Useful
perspectives are to be found in D. McFarland and T. Bösser. Intelligent behaviour in animals and robots
MIT Press; R. Pfeifer and C. Scheier, Understanding
intelligence, MIT Press; S. Shettleworth, Cognition evolution and behaviour. Oxford University Press.
(But you are not recommended to
buy them). Essential articles, usually one per lecture, and one or two for each
seminar are starred and in bold. These should definitely be read. Some of these
will be available for purchase as a study pack in the COGs library from Celia
Mcinnes (4C2). Others can be down-loaded from the web. All are available for
borrowing from the reserve collection in the Main University Library. The other
papers on the reading list shouldn't be ignored, but you aren't expected to
read them all! Many of the items are to help with essay writing. Single copies
of most of these items (apart from books) are available in the BIOLS school
library. If you have problems getting hold of material or need extra sources,
please contact the lecturer concerned.
TIMETABLE
|
Week |
Lecturer |
Title |
Seminar |
|
1 Oct 8 |
AT AT |
Robots &
Biology Robots &
Biology |
|
|
2 Oct 15 |
TC TC |
Insect navigation
Insect
navigation |
Robot
architecture (AT) |
|
3 Oct 22 |
TC TC |
Path planning Memory and
flexible behaviour |
|
|
4 Oct 29 |
DO DO |
Motion
detection and flow fields in insects 1 and 2 |
Bee navigation (TC) |
|
5 Nov 5 |
DO AT |
Optic
flow Artificial
neural networks 1 |
Memory (TC) |
|
6 Nov 12 |
AT AT |
Artificial
neural networks 2 and 3 |
Neural coding
(DO) |
|
7 Nov 19 |
DO DO |
Object
recognition and categorisation 1 and 2 |
Artificial
neural networks (AT) |
|
8 Nov 26 |
TC TC |
Social
organisation of foraging Social
learning |
Insight and
tool use (TC) |
|
9 Dec 3 |
TC |
Why are primate
brains big? |
Social
intelligence (TC) |
|
10 Dec 10 |
|
Have essay
plans ready!!! |
|
COURSE SYNOPSIS
1.
Thinking about Intelligent Behaviour.
Intelligence, adaptive behaviour,
adaptation. Some key historical frameworks(with special attention to vision):
Cartesianism, Evolution, Cybernetics and dynamical systems theory, `classical'
AI, Connectionism, Behaviour-based,`Artificial Life.' Relationship between
anatomy and function; functionallocalisation. Interface between biology and
engineering. How much do we currently understand all this?
Reading:
(Available on the
web: http://www.ai.mit.edu/people/brooks/papers.html)
Extracts from
*Brooks, R.A. (1995)."Intelligence Without Reason". In Steels, L. and
Brooks, R. (eds.) The Artificial Life Route to Artificial Intelligence:
Building embodied, situated agents, 25‑81, Lawrence Erlbaum Associates.
Arkin R.C.
(1998) Behavior-based robotics. MIT Press.
Boden, M. (ed)
(1996) The philosophy of artificial
life. Oxford University Press
Braitenberg, V.
(1984). "Vehicles: Experiments in Synthetic Psychology", MIT Press.
Library: QU 4588 Bra (1 copy in MAIN).
Ashby, W.R.
(1960). "Design for a brain: the origin of adaptive behaviour",
Chapman. Library: QE 230 Ash (1 copy in RESERVE).
Gibson, J.J.
(1979). "The ecological approach to visual perception", Houghton
Mifflin. Library: QZ 314 Gib (1 copy in RESERVE, 3 in SHORT).
Gray Walter: see
http://gate.uwe.ac.uk:8002/IAS/ on the web.
Proceedings of
the four "Simulation of Adaptive Behaviour (SAB): From animals to
animats" conferences. Library: QZ 1250 Fro (Several copies in MAIN and
RESERVE).
2.
Designing Intelligent Behaviour
Special focus on Brooks' approach.
Reactive and non-reactive control.
Situatedness, embodiment and emergence;
behaviour-based robotics; the
subsumption architecture. Top-down vs.
bottom-up design; hierarchical controlstructures. General purpose vs.
niche-specific; Horswill's `habitat
constrained' vision. The battles: Pros
and cons of information processing andinternal representation perspectives.
Robots and simulations as models ofnature.
Reading :
*Webb, B. (1996). "A Cricket Robot".
Scientific American, December 1996, 62‑67.
Franceschini, N.,
Pichon, J.M., and Blanes, C. (1997). "Bionics of Visuo‑motor
Control". In: Evolutionary Robotics: From intelligent robots to artificial
life (ER'97), Gomi, T. (ed.), 49‑67,
AAI Books.
Deneubourg, J.L.,
et al. (1991). "The dynamics of collective sorting: Robot‑like ants
and ant‑like robots". In Meyer, J‑A., and Wilson, S.W. (eds.),
Proc 1st Int. Conf. on Simulation of Adaptive Behaviour: From Animals to
Animats, 356‑363, MIT Press.
3. Navigation in Insects
Ants and bees are
impressive navigators. They leave their nest to collect food from sites that
may be located hundreds (ants) or thousands (bees) of metres away. They then return accurately to their nest.
To do this, they have at their disposal a repertoire of navigational strategies
that must be properly co-ordinated. A primary one is path integration or dead
reckoning. Unavoidable inaccuracies arising from path integration are reduced
by the insects' use of visual landmarks to specify stereotyped routes. The
study of navigation can tell us much about sensori-motor control and 'situated
cognition' in these animals.
Reading:
*Wehner, R. Michel, B., Antonsen, P.(1996) Visual
navigation in insects: coupling of egocentric and geocentric information. J exp
Biol 199, 129-140.
Wehner, R. (1992)
The arthropods. In Animal Homing (ea. F. Papi). pp. 45-144. Chapman and Hall
Muller, M.,
Wehner R. (1988) Path integration in desert ants, Cataglyphis fortis Proc. Nat.
Acad. Sci., USA. 85:5287-5290.
Wehner, R. and
Srinivasan M. V. (1981) Searching behaviour of desert ants, genus Cataglyphis
(Formicidae, Hymenoptera). J. comp. Physiol A, 142:315-338.
Journal of
Experimental Biology, Symposium volume on Navigation. Jan 1996 vol 199.
(Downloadable from the web: www.biologists.com)
(articles by Srinivasan, M.V. et al., Esch, H
& Burns, J.E)
Srinivasan MV, Zhang SW, Bidwell NJ:
Visually mediated odometry in honeybees navigation en route to the goal: visual
flight control and odometry. J exp Biol
1997 200:2513-2522.
Srinivasan
MV, Zhang SW, Altwein M Tautz J(2000), Honeybee navigation: nature and
calibration of the 'odometer'. Science 287, 851 - 853.
Wohlgemuth S,
Ronacher B, Wehner R (2001) Ant odometry in the third dimension. Nature 141,
795-798.
4. Defining Places by Landmarks
Bees and ants use
landmarks to specify a place. What kinds of representations of landmarks do
they have, how do they acquire these representations and how do they use them
for navigation? Answers to these
questions show how apparently complex tasks can be accomplished in relatively
simple ways and mimicked by model navigational systems simulated through
artificial evolution.
Reading:
*Collett, T.S. (1992) Landmark learning and guidance
in insects. Phil. Trans R. Soc. B 337:295-303.
Dill, M, Wolf R.
and Heisenberg M. (1993) Visual pattern recognition in Drosophila involves
retinotopic matching. Nature 365:751-753.
Journal of
Experimental Biology, Symposium volume on Navigation. Jan 1996 vol 199
(Downloadable
from the web: www.biologists.com)
(especially
articles by Menzel, R. et al, Zeil J. et al, Lehrer, M., and Dyer, F.C.)
Dale, K., Collett, T.S. (2001)
Using artificial evolution and selection to model insect navigation. Current
Biology 11, 1305-1316.
5. Path Planning by Spiders and Frogs
Many animals must
plan efficient routes through cluttered environments and the methods that they
use give insights into what they 'know' about their 3-D environment and how
this knowledge is used in intelligent planning. We will see the very different
strategies and mechanisms adopted by spiders planning routes through complex
3-D mazes and frogs and toads planning detours round barriers, and how planning
strategies can be implemented neurally in simple structures.
Reading:
*Arbib, M. A. and Liaw, J.-S. (1995). Sensorimotor
transformations in the worlds of frogs and robots. Artificial Intelligence
72:53-79.
Jackson, R. R.
(1985) A web-building jumping spider. Scientific American 253 (Sept):106-113.
Hill, D.E. (1979)
Orientation by jumping spiders of the genus Phiddipus during the pursuit of
prey. Behav. Ecol. Sociobiol. 5:301-322.
Tarsitano, M.S.,
Jackson, R.R. (1997) Araneophagic jumping spiders discriminate between detour
routes that do and do not lead to prey. Anim. Behav. 53: 257-266.
Tarsitano, M.S.,
Andrew, R. (1999) Scanning and route selection in the jumping spider Portia
labiata. Anim. Behav. 58, 255-265.
Collett, T.S.
(1982) Do toads plan routes? J. comp Physiol. 146:261-271.
Menzel E.W.
(1973) Chimpanzee spatial memory organisation. Science 182:943-945.
6.
Memory
organisation: procedural, contextual and episodic
Most
intelligent behaviour relies on remembering and utilising previous experiences,
both in the short term (working memory) and in the longer term. The importance
of memory in allowing flexible behaviour can already be seen in insects.
Psychologists divide human long-term memories into two very different
functional classes: procedural (skills and habits) and episodic (memory of
individual events). Detailed study of the memory requirements of caching
behaviour in birds reveals that birds have the ability to remember and utilise
information about specific events.
Reading:
*Griffiths,
H, Dickinson, A. Clayton, N. (1999) Episodic memory: what animals can remember
about their past. Trends in Cognitive Science 3, 74-80.
Eichenbaum, H.
(1997) Declarative memory: insights from cognitive neurobiology. Ann. Rev.
Psychol 48,547-572.
Clayton, N.S.,
Dickinson A. (1998) Episodic-like memory during cache recovery by scrub jays.
Nature 395, 272-274.
Clayton, N.S.,
Dickinson A. (1999) Memory for the content of caches by scrub jays (Aphelocoma coeruslescens). . J. exp.
Psychol. Anim. Behav. Processes 25, 82-91.
Roitblat,
H.L. (1987) Working memory Chapter 5 of: Introduction to comparative
cognition. W.H.Freeman.
Squire, L.R.
& Kandel E.R. (1999) Memory: from mind to molecule. Scientific American
Library.
7. - 9. Visual Coding and Motion Detection in Flies. Neural Pathways, Behaviour, and Algorithms. (3 lectures)
A fundamental
question motivating comparisons between animal and machine intelligence is:
Could we in principle we make a machine the exactly mimics a human or animal
brain? The middle of this century saw developments in universal computing
machines implementing simple logical operations and neurophysiology of synapses
– the low-level end of the machine vs. organism comparison. Workers such as the
mathematician A. Turing and the biologist/philosopher W. McCulloch asked
whether the logical operations required for a universal computer could be
implemented by a brain, and whether there is more to brains than formal logic.
This approach to neural computation is illustrated by work on visual motion,
which asks how single synapses in ‘special purpose’ neural circuitry solve a
specific computational problem. Later AMI deals with neural networks which make
more general comparisons between machine and brain computational architectures.
Specifically these lectures concern
neural mechanisms beneath the insect’s eye. For example, Visual motion
flowfields are derived from retinal stimuli by integrating from local
directional motion signals and are used to stabilise flight. The way local
motion signals are abstracted and how they are integrated into behaviour have
been a test-bed for ideas at the interface of neurobiology, behaviour, formal
modelling and machine vision. We go on to look at how bees and other insects use
visual motion signals to control direct level flight, and also manoeuvres such
as obstacle avoidance landing, and (in some cases) how these controls are
implemented by the nervous system, and have been implemented by designers of
autonomous robots.
Egelhaaf, M. and Borst, A. (1993) Motion
computation and visual orientation in flies. Comp. Biochem. Physiol.,
104A:659-473
Franceschini, N., Pichon, J. M. and Blanes, C.
(1992) From insect vision to robot vision. Phil. Trans. R. Soc. Lond. B.
337:283-293.
Srinivasan MV et al. (2001) Landing strategies
in honeybees, and possible applications to autonomous airborne vehicles. Biol Bull. 200, 216-221
McCulloch,
W. S. (1952) Finality and form in the nervous system. Reprinted in Embodiments
of mind (1965/1988). pp.256-275. MIT Press.
Poggio, T. and Koch, C. (1987) Synapses that compute
motion. Scientific American, May 1987, pp.42-48.
Rind FC, Simmons PJ (1999) Seeing what is
coming: building collision-sensitive neurones Trends Neurosci 22 215-220
10 to 12. Artificial Neural Nets (ANNs)
10.
Basics and History
What ANNs are. Feedforward and recurrent
nets. Learning vs. hardwired. The
Perceptron; training, testing and
generalisation. Weight vectors and error surfaces; gradient-descent learning.
The need for a hidden layer. NETtalk asan example. What ANNs are good for.
11.
Some details of learning mechanisms.
Backpropagation. Kohonen's
self-organising maps. Reinforcement learning (Barto's pole balancer).
Reading for 10 and 11:
*Elman, J.L. et al. (1996) Rethinking Innateness. MIT Press. chap 2: Why connectionism? Pp
47-106.
Sejnowski, T.J.
and Rosenberg, C.R. (1986). "NETtalk: a parallel network that learns to
read aloud". Reprinted in the book below, Chapter 40
Anderson, J.A.
and Rosenfeld, E. (eds) (1988). "Neurocomputing: foundations of
research", MIT Press. Library: QU 4550 Neu (1 copy in MAIN, 1 in SHORT).
McCord Nelson, M.
and Illingworth, W.T. (1991). "A Practical Guide to Neural Nets",
Addison‑Wesley. Library: QZ1335Nel, (1 copy MAIN, 1 SHORT).
Browse the QZ
1335 section in the library.
12.
ANNs and nature.
What's the relation between ANNs and
brain function, anatomy and psychology?
"Biological Plausibility" of
architectures and learning regimes. Hebbian
learning. Local and distributed
representations in ANNs. Graceful degradation.
Symbolic vs. non-symbolic, semantic
grounding. Computational Neuroethology,
ANNs as "artificial nervous
systems"; time and dynamics. Artificial evolution
of ANN designs.
Reading:
(Available on the
web: http://citeseer.nj.nec.com/cliff91computational.html)
*Cliff, D.
(1991). "Computational Neuroethology: A Provisional Manifesto". In
Meyer, J‑A., and Wilson, S.W. (eds.), Proc 1st Int. Conf. on Simulation
of Adaptive Behaviour: From Animals to Animats, 29‑39, MIT Press.
Beale, R. and
Jackson, T. (1990). "Neural Computing ‑ an introduction",
chapter "Kohonen Self‑Organising Networks", IOP Publishing.
Roitblat, H.L. et
al. (1991). "Biomimetic Sonar Processing: From dolphin echolocation to
artificial neural networks." In Meyer, J‑A., and Wilson, S.W.
(eds.), Proc 1st Int. Conf. on Simulation of Adaptive Behaviour: From Animals
to Animats, 66‑76, MIT Press.
13 to 14 Visual
object recognition (2 lectures)
These lectures
are concerned with how animals recognise and categorise objects, and whether
these tasks might be accomplished using simple rules. We ask whether simple
computational models such as neural networks may suggest how animals respond to
visual signals and help us understand how striking visual displays such as the
peacock’s tail have evolved.
Most objects such as rocks and
vegetation, are ‘background’ which animals move past or use as landmarks in
navigation. Objects of special importance include predators, food sources or
members of the same species. Here appearance may give information; for example
about the desirability of a food item or the suitability of a mate. The ability
to place objects in categories and to communicate categorical information has
been linked to human language. Evidence for such abilities in other animals is
therefore of particular interest.
Reading:
*Cerella J (1979) Visual classes and natural
categories in the pigeon. J. Exp. Psychol.: Human Perception and Psychology. 5,
68-77
*Enquist M, Arak A (1993) Selection of
exaggerated male traits by female aesthetic senses. Nature, 361, 446-448
Enquist
M, Arak A (1994) Symmetry, beauty and evolution. Nature, 372,169-172
Herrnstein
RJ (1985) Riddles of natural categorization. Phil. Trans. R Soc Lond B, 308,
129-144
Dawkins
MS, Guilford T (1995) An exaggerated preference for simple neural network
models of evolution. Proc. R Soc. Lond. B. 261, 357-360
Bullock
S, Cliff D (1997) The role of 'hidden preferences' in the artificial
co-evolution of symmetrical signals. Proc. R Soc. Lond. B. 264, 505-511
Enquist
M, Johnstone RA (1997) Generalization and the evolution of symmetry
preferences. Proc. R Soc. Lond. B. 264, 1345-1348
15. The Social Organisation of Honeybee Foraging
The socially
organised behaviour of a hive of honey bees during foraging is a wonderful example
of how simple rules followed by individual bees leads to exquisitely organised
and effective global behaviour without centralised control. Those working in
this field like to consider individual bees as individual neurones and a hive
of bees as a brain. We will mostly emphasise how information concerning the
availability and need for nectar is transferred within the hive, and how good
decision making can arise despite the limited knowledge available to individual
bees.
Reading:
*Seeley, T.D., Towne, W.F. (1991) Collective decision
making in honey bees: how colonies choose among nectar sources. Behav. Ecol.
Sociobiol. 28:277-290.
Seeley, T.D.
(1992) Tactics of dance choice in honey bees: do foragers compare dances.
Behav. Ecol. Sociobiol. 30:59-69.
Michelsen, A., et
al. (1992) How honeybees perceive communication dances, studied by means of a
mechanical model. Behav. Ecol. Sociobiol. 130:143-150.
Seeley, T.D.
(1994) Honey bee foragers as sensory units of their colonies Behav. Ecol.
Sociobiol. 34:51-62.
Seeley, T.D.
(1995) The Wisdom of the hive: the social physiology of honey bee colonies.
Harvard University Press.
Seeley, T.D.
(1999) Group decision making in swarms of honeybees. Behav. Ecol Sociobiol. 45,
19-31.
H.
Esch, S.W. Zhang, M.V. Srinivasan & J. Tautz (2001): Honeybee dances
communicate distances measured by optic flow. Nature (Lond) 411, 581-583.
Seeley, T.D., Buhrman S.C. (2001) Nest-site selection in honey bees: how well
do swarms implement the best-of-N decision rule? Behav. Ecol. Sociobiol. ,
416-427.
16. Culture and imitation
Quite apart
from its role in social organisation, socially guided learning is of immense
benefit to individuals, increasing their behavioural effectiveness in the
world. It speeds up the learning of environmental affordances, motor skills,
and makes it possible to accumulate information and skills across individuals
and generations. There are some tricks for making social learning easy, but
often it needs the underpinning of complex neural mechanisms.
Reading:
* Boesch, C. The emergence of culture
among wild chimpanzees. Proc. Brit.Acad 88, 251-268.
Multi-authored
feature article in Science 1999 vol 284, 2070-2076. Chimps in the wild show
stirrings of culture.
McGrew, W.C.
(1998) Culture in non-human primates. Ann. Rev. Anthropol 27, 301-328.
Tomasello, M.
(1996) Do apes ape? In C.M.Heyes and B.G.Galef (eds) Social learning in
animals: the roots of culture (pp. 319-346)
Whiten, A (1998)
Imitation of the sequential structure of actions by chimpanzees (Pan
troglodytes). J. comp. Psychol. 112, 270-271.
Miklosi, A.
(1999) The ethological analysis of imitation. Biol. Rev. 74, 347-374.
Rizzolatti, G
and M.A.Arbib (1998) Language within our grasp. Trends in Neurosci. 21,
188-194.
17. Big Brains, Social Intelligence and the Theory of
Mind
In order to
discover why primate brains have become big, attempts have been made to
correlate the average brain size of a species with its lifestyle. The most
promising result so far are that brain size varies with: (a) the complexity of
foraging skills, especially hunting, and (b) the size of a species' social
group. The latter finding perhaps giving some support to the controversial
notion that increased primate 'intelligence' was driven by the problems of
coping with a complex social life, in which problems are likely to become
harder as social skills evolve. To be socially skilled, one needs to anticipate
the behaviour of others. Prediction might be easier if an animal can guess what
another animal is thinking or intending and there have been explicit
suggestions that apes are 'mind-readers'. How can this be investigated
experimentally?
Reading:
*Dunbar, R.I.M. (1998) The social brain hypothesis.
Evol. Anthropol. 6,178-190
Adolphs, R.
(2001). The neurobiology of social cognition. Current Opinion in Neurobiology
11, 231-239.
Allman, J. (1999)
Evolving brains. Scientific American Books
Povinelli, D.J.,
Preuss, T.M. (1995). Theory of mind: evolutionary history of a cognitive
specialization. Trends in Neurosci. 18:418-424.
Byrne, R.B.
(1995) The thinking ape. Oxford University Press.
Whiten, A. (1996)
When does smart behaviour-reading become mind-reading? In: Theories of theories
of mind (ed P.Carruthers, P.K.Smith). pp. 300-324. Cambridge , Cambridge
University Press.
Gigerenzer,
G.(1997). The modularity of social intelligence. In: Machiavellian Intelligence
II (ed Whiten, A, Byrne R.W.). pp. 264-288. Cambridge , Cambridge University
Press.
Machiavellian
Intelligence l and ll (eds Whiten A., Byrne R.W.) Cambridge , Cambridge
University Press.
Kaplan, H. Hill,
K., Lancaster, J., Hurtado, A.M. (2000) A theory of human life history
evolution: diet, intelligence, and longevity. Evolutionary Anthropology 9,
156-185.
Call, J. (2001)
Chimpanzee social cognition. Trends in Cog. Sci. 5, 388-393.
1.
Does behaviour-based robotics scale up?
Reading:
(Available on the web: http://icl-server.ucsd.edu/~kirsh/Articles/Earwig/earwig-cleaned.html)
*Kirsh, D., 1991. "Today the earwig,
tomorrow man?" Artificial
Intelligence
47, pp161-184 (Reprinted in Boden's "Philosophy of Artificial
Life"). Additional reading on this:
**Brooks "From Earwigs to Humans"
On the web: http://www.ai.mit.edu/people/brooks/papers.html
We'll divide into two groups, one arguing
Kirsh's case and the other that of
Brooks. Reading as much as possible
before the seminar of Kirsh's paper,
and
the Brooks reading for Lecture 1, will
help a lot. Here's just some of the
things to think about while reading
Kirsh:
*
Are there bits you don't understand? (Lecturer will explain)
*
Does Kirsh's statement of Brooks' case square with your reading of
him?
*
Is it right to drive a wedge between the sense-driven and the
representation-driven, and expect to be
able to make a hybrid having elements
of each?
*
How much does Kirsh's argument appeal to intuition and introspection?
*
Some of Kirsh seems to assume a strong division between the mechanisms
of perception, reasoning, action, etc,
but does he mean this and does it
matter?
*
What can Kirsh mean by "computational cost"?
*
One of Kirsh's strongest points is to do with learning. What would
Brooks say?
*
Who's right?
Don't be frightened by this difficult
argument - we'll do it in a fun way.
2. Flexible
navigation in insects
Reading **Menzel
R et al. (1998) Bees travel homeward routes by integrating separately acquired
vector memories. Anim Behav 55 (part
1):139-152.
Traditionally, flexible navigation in animals has been attributed to
‘cognitive maps’ but users of the term are rarely clear about what it means.
Work on insects has attempted to explore in more detail how it is that they
manage to navigate flexibly, and what the limits on their flexibility might be.
1.
What do you understand by the term cognitive map?
2.
Gould’s experiment (described in the lecture) was one attempt to test
whether bees have cognitive maps. What was the reasoning behind his
experimental method? Unfortunately, his results did not prove to be replicable.
But there are other means by which insects can navigate flexibly to a goal. How
might path integration or landmark navigation give flexibility to an animal’s
path
3.
Path integration and landmark navigation provide very different kinds of
positional information. What benefits are there if the two can be integrated?
Discuss the method that Menzel et al. used to show flexible landmark navigation
in a familiar landscape. What differences were there in the behaviour of hive
arriving and hive departing bees when released at the feeder site? How does
motivational state influence memory retrieval?
4.
What were the purposes of the control manipulations that Menzel et al.
performed.
5.
How do Menzel et al. account for their results? Can you suggest other
possible explanations of the ability to home directly from the unfamiliar
interemediate release site?
6.
Would activitating two vector memories be possible in all landscapes?
7.
One generalisation from these experiments is that bees have vector maps.
I.e. they link scenes within a familiar environment to nest directed
vectors. What limits would such a ‘map’
place on an insect’s ability to plan routes?
Readings
in study pack:
*Griffiths,
H, Dickinson, A. Clayton, N. (1999) Episodic memory: what animals can remember
about their past. Trends in Cognitive Science 3, 74-80.
*Hampton, RR. Rhesus monkeys know when they remember. Proceedings
National Academy of Sciences 2001, 98 5359-5362
The aim of this tutorial is for you to discuss and clarify differences
between these types of memory, to consider how the different types might be
stored and recalled, to decide whether animals other than humans can be shown
to have episodic memories.
Specific questions to think about when reading the papers:
1. What difficulties arise when trying to show that animals have
episodic memories?
2. How do Clayton and Dickinson deal with them?
3. How do they demonstrate that birds learn ‘what, where and when’?
4. In what different ways might ‘when’ be encoded?
5. Does it make a difference that the jay’s ability may well 6. be
limited to memories needed for food caching?
7. What is Hampton trying to test and does he succeed?
8. What is the reason for including probe trials in which images were
not presented?
9. What does the Hampton study add to the conclusions reached by
Griffiths et al?
4.
Neural maps and neural coding
A
fundamental question is neuroscience is the relationship between neural
activity and perception/action. The two articles by Newsome et al and Sparks et al. from the Cold Spring Harbor Symposium. (vol. 55, 1990) in the study pack review excellent
experimental studies which give two different views of this relationship.
Questions.
1.
Outline
the experimental procedures used by Sparks et al. and Newsome et al. How do
they: a) record neural activity; b) record behavioural responses. c) manipulate neural activity?
2.
How
do the responses of neurons vary across the surfaces of the brain structures
studied by Newsome and by Sparks.
3.
What
is a neural map, and why, from a computational perspective, might it be useful?
(see Sparks)
4. Outline Newsome’s and
Spark’s main conclusions about the
relationship between neural activity and perception. How do they differ and
what do they imply about computational strategies used by the brain?
Readings (in
study pack):
*Heinrich B
(1995) An experimental investigation of insight in common ravens. The Auk
112:994-1003.
*Visalberghi, E., Limongelli, L. (1994) Lack of
comprehension in tool-using capuchin monkeys. J. comp. Psychol 108: 15-22.
(For more on
cognition in birds see two recent books: B. Heinrich. The Mind of the Raven; I
Pepperberg. The Alex Studies)
How do animals solve problems using tools? Do they reach solutions by
trial and error, or do they have insight into what they are doing? How can one
get animals without language skills provide answers to such questions? In this
tutorial, we will look at one study on Capuchin monkeys. The authors conclude
that although these monkeys are very adept at using tools in artificial tasks,
they do not understand what they are doing.
A second study on ravens comes to exactly the opposite conclusion.
1.
What do you understand by insight? Why is it so difficult to demonstrate
either its presence or its absence? How does one exclude the possibility that
apparently insightful behaviour is a consequence of learning or of innate
predispositions?
2.
What evidence have Visalberghi et al. given to show that Capuchin
monkeys don’t have insight into the tube task? Does it convince?
3.
What does the string task require ravens to do?
4.
How do crossed strings increase its difficulty?
5.
What does the sheep’s head test show?
6.
Summarise Heinrich’s evidence for insight in ravens.
7.
Does the presence of large individual differences in behaviour affect
his argument?
8.
Could this apparently insightful behaviour tap into some innate part of
their normal feeding behaviour?
6. Artificial Evolution, neural nets, and robotics
Reading:
"Artificial Evolution: A new path for Artificial Intelligence?"
P. Husbands, I. Harvey, D. Cliff, G. Miller Brain and Cognition Vol. 34, No. 1,
pp130-159
Just extracts: pages 1-17 and 24-27. Files are available to download at:
http://www.informatics.sussex.ac.uk/users/adrianth/AMI/SEMINAR/
To think about:
How is this "artificial evolution"
similar to natural evolution, and to the human breeding of animals and plants,
and how is it different?
How does the design of an ANN though artificial
evolution compare to the other ANN `learning' methods?
Why were the authors able to depart from a
standard regular ANN structure, such as a feedforward layered network, and why
did they think this was worthwhile?
The paper is inspired by nature, but could such
work give anything back to biologists?
How does this compare to Brooks' approaches?
What to you think might be achievable, and
unachievable, using such methods? Are these conclusions affected by exactly
what sorts of evolutionary algorithms and neural networks are chosen?
7. Social organisation, social learning, co-operation
and culture
No extra readings for this seminar, but please be sure
to have read the papers for the relevant lectures.
It has been
suggested that insects need something like an ‘external memory’ for them to
adjust their behaviour to the previous achievements of others. A chemical trail to food, a partly built
nest, a store of nectar could all be considered as external memories. How in
detail might such ‘memories’ be used in organising co-operative behaviour? We’ll look at overheads of different
examples for you to discuss.
In wood ants
foraging trunk trails persist from year to year even though chemical trails are
almost certainly erased over the winter. How is this ‘cultural tradition’
maintained?
How do honeybees’
individual memories fit into the social organisation of foraging?
Honeybees don’t
know the identity of the other bees with which they communicate. How might
interacting with known neighbours enhance social co-operation?
How do you define
culture? What is the evidence for ‘culture’ in non-human primates?
In non human
primates, this culture can only be carried in individual memories. In the
absence of language, how is cultural information transmitted between
individuals? Is generational transmission horizontal or vertical or both ?
What are the
benefits to be obtained by using observational/social learning?
Sample Essay Titles
1. We can build
machines (computer programs) to play chess more successfully than we can devise
a two‑legged robot to walk without falling over. What implications does
this have for those trying to build intelligent machines? The chess‑playing
program works by having perfect knowledge of the chess‑board and applying
a chain of abstract reasoning to work out the best thing to do next, searching
methodically through millions of possibilities. How does this compare with the
way animals behave adaptively in the real world?
2. Attempts to
understand nervous systems, and the way in which animals (and hopefully robots)
behave adaptively or `intelligently', have a long history of being wrong (many
now think). How confident can we be that our current
thinking
represents progress? Why? Part of the current approach is to study
"simple" animals like insects. How might this be able to shed light
on more complex animals like humans, or are we now simply evading the difficult
questions?
3. What do
insects store about their routes through familiar environments (don’t forget
path integration)? How do they use this information for navigation and in what
ways does it allow their navigation to be flexible? What useful general lessons
are there to be learnt from studying insect navigation?
4. Outline
what is meant by the term elementary motion detector (emd). How might emd’s be
implemented in nervous system or flies or other animals, and how might the
principles they embody be relevant to the design of autonomous vehicles, and
artificial intelligence in general?
5. A wealthy agency (the US airforce perhaps)
requests your advice on whether they should invest in implementing principles
derived from the study of insect flight control for the design of autonomous
agents. Make a case outlining general principles, giving examples of work done
so far, and reaching a clear conclusion.
6. To what extent
does the organisation of real nervous systems exemplify Brooks' subsumption
architecture?
7. Discuss the
advantages and disadvantages of the approach of intelligence as adaptive
behaviour, contrasting that approach with knowledge-based AI.
8. Are recent
developments in visual guidance of autonomous agents an improvement on
traditional approaches to problems of path planning, or a seductive dead-end?
9. Is it useful to compare the function of neural
synapses with the logical operators used in computing? Illustrate your answer
with examples of neural mechanisms.
10.
Are neural networks useful as metaphors or models for understanding
brains? Illustrate your answer with
examples from work on signal evolution and/or pole-balancing.
11. Discuss F.
Crick's assertion: 'The brain isn't even a little bit like a computer'
12. Why are
primate brains big?
13. Why might an
artificial neural network model of behaviour‑generation be of more
interest to biologists than, say, a computer program having a set of rules that
generates roughly the same behaviour? Hint: discuss applying constraints of
biological plausibility, investigating failure modes, etc.
14. How might the
potential for learning within an artificial neural network model be useful when
investigating biological phenomena (eg. by using situated robotics)? What might
be the uses of the various training regimes, learning rules, and network
architectures, in this context?
15. I have a feed‑forward
neural network with one hidden layer. I train it using the back‑propagation
rule so that it can control a robot to avoid obstacles. The network has four
inputs, which represent when an obstacle is detected in front, behind, or to
the left or right of the robot. The outputs tell the robot whether to turn 90‑degrees
left or right, or whether to go exactly 10.0 cm forwards in this particular
time‑step. Why is analysis of the learnt network unlikely to shed much
light on how animals achieve similar behaviours? How else would you criticise
this experiment? How might it be improved?
16. ‘The
difference in mind between man and higher animals, great as it is, certainly is
one of degree and not kind.’ Discuss this claim of Darwin.
17. Discuss
differences between procedural and episodic memory in human and non-humans, and
the associated problems of storage (where and how) and recall.
18. Observational
learning/imitation - why is it important and what mechanisms are needed to make
it work?
19
What
is a perceptual category, why might we doubt that (non human) animals can form
them? Discuss experiments that test the hypothesis that animals can form
categories,
20
What
can neural networks and other models tell us about how animal communication has
evolved and operates.
21. How does
global order in the foraging activity of a colony of bees emerge from local
information processing by individuals?
22. There are various kinds of artificial neural
network `learning', such as supervised (eg. the backpropagation technique),
self-organising (eg. Kohonen maps) and reinforcement learning. How do these
relate to the learning seen in animals? (You could discuss at the level of
neural mechnisms, or of behaviour/psychology, or both. Perhaps you could
mention learning in animal-like robots.)