Natural Cognition and Artificial Intelligence: What can biologists learn from AI?
logo by Jolyon Troscianko
Dr Jackie Chappell
Cognitive Adaptations Research GroupSchool of BiosciencesUniversity of Birmingham
What problem are we interested in?
Photo: Milwaukee County Zoo
Photo: Honda
cognitive mechanisms unclear
cognitive mechanisms well
understood
behaviour rich and fairly well understood
behaviour simple and usually optimised for particular
environment/context
Biomimetic robots as tools for probing animal behaviour
45°
Honey bee waggle dance
Taken from: Tanner & Visscher 2010
‘RoboBee’
- angle of waggle run- length of waggle run- vibrations and
changes in air currents- thoracic temperature
increase- pheromones
Which of these are decoded by watching bees?
Photo: http://robobee.eu
Robots to test hypotheses about mechanisms
Phonotaxis in crickets (Gryllus spp.)
♂
♀Photo: sanmartin@flickr
Photo: elchip@flickr From: physiology.org
The existing models
turn left turn right
left ear right ear
turn left turn right
if left > right
AND
AND
recogniser
if right > left
left ear right ear
recogniser recogniser
if left > right if right > left
Adapted from: Webb & Scutt (2000) Biol Cybern 82, 247-269.
The robot model
• Is the turn determined by the firing rate of auditory neurons on each side or a comparison of the time taken for sound to reach each ear?
• How is the song (characteristic repetition rate of syllables) recognised?
• Robot uses biologically-plausible 4 neuron system
• No mechanism for comparing songs or recognition
• Processes sound fast enough to use real cricket song as a stimulus
• Result: phonotaxis just like a cricket
Taken from: Webb & Scutt (2000) Biol Cybern 82, 247-269.
How do rats use their whiskers to explore?
Grant et al. (2009) J Neurophysiol 101, 862-874.
exploring floor exploring wall
min. spread perpendicular to
surface
min. spread perpendicular to
surface
max. spread parallel to surface
max. spread parallel to
surface
Using AI as a tool for thinking about cognition
Can orangutans plan their actions?
L M MS
gaps
closedend
openend
forward-facingtrap
backward-facingtrap
Tecwyn, Thorpe and Chappell (2012) Animal Cognition 15: 121-133
Amos
Puzzle tube
Tecwyn, Thorpe and Chappell (2012) Animal Cognition 15: 121-133
Puzzle tube
Tecwyn, Thorpe and Chappell (2012) Animal Cognition 15: 121-133
Using AI to approach the problem differently
problem to be posed to
animals in experiment
decompose problem using AI
techniques
compare results
simulate results using AI
techniques
refine problem
pose problem to test animals
finalised problem
problem to be posed to
animals in experiment
decompose problem using AI
techniques
simulate results using AI
techniques
refine problem
MAPL/MAPSIM planner
Collect resultsRun planner (MAPSIM)
Define individual problems
Decompose task domain
facts
facts
facts
states
preconditions
effects
actions
DOMAIN MODEL
goal state
initial state
facts known by agent
PROBLEM DESCRIPTION
(ONE OF MANY)
Generate plans
Simulate execution of
plans
OUTCOME: smallest
number of actions
necessary to attain goal state
effects of actions change the state of the world and preconditions for further action
1
2
3 4
Chappell & Hawes (2012) Philos Trans R Soc Lond B Biol Sci 367, 2723-2732.
How does this approach help?
• Work out:
• What problems the environment poses
• What information environment provides
• How animals access (some of ) this information
• How animals might solve these problems
• Then in a better position to hypothesise about mechanisms which would allow animals to solve problems in a particular way
What have we learned?
• Using robots/AI can be useful, but you have to be clear about what you gain
• Can be difficult to find projects where the problems are equally interesting to biologists and computer scientists
• Modelling behaviour in a robot can lead to important insights, but does not prove that a particular mechanism is used in animals
• Perception and action are relatively easy to model/produce in robots, but cognition is really hard
• Need to use AI to help define what the problem is first