Transcript part_2

Matching brain and body dynamics
• Daniel Wolpert:
– "Why don't plants have brains?"
– "Plants don't have to move!"
• Early phases of embodied artificial intelligence:
– Navigation and orientation
– Wheeled robots: Mismatch in complexity levels of sensory,
neural and motor systems (unsuccessful, and inconsistent
with the “ecological balance principle”)
Matching brain and body dynamics
• Research inspired by rats
– Place cells (O’Keefe and Dostrovsky, 1971)
– Rat-mimicking robots
• Problem 1: complex sensory-motor skills not accounted for
• Problem 2: rat cognition not limited to navigation
Other aspects of development
An important goal of developmental robotics is to construct a
robot which develops into an intelligent adult human-like robot
• Development of sensory and motor systems
– Biological organisms changes shape
• Growth cannot yet be modeled well on robots
– Robot experiment by Max Lungarella
• From low-precision to high-precision motor system
• From low-resolution to high-resolution sensory system
• Using previously aquired skills to learn new ones
• Time
– Many time scales involved (millisecs -> years)
– Today, robot learn a skill very quickly
Other aspects of development
Other aspects of development
• Motivation
– Why would a robot want to develop/improve over time?
– Autotelic principle
• Good for more complicated tasks
• A single motivating principle
• Inspired by the concept of “flow”: Experience/feeling when
performing an expert activity
• Maintain constant flow of information through system
• Example: learn to reach -> learn to grasp -> learn to drink
Learning in embodied systems
• Exploration guided by embodiment facilitates
learning
• Example: Baby learning to categorize objects
– Hypothetically, random signal are sent to arm muscles
– Movements constrained by anatomy and muscle-tendon
properties
– Objects grasped due to (hypothetical) grasp reflex
– At some point, object is brought to mouth, leading to strong
sensory input (sensory-motor correlation)
– Eventually, categorization possible with less interaction
Social interaction
• Important aspect of development
• Imitation
–
–
–
–
–
Powerful strategy for humans
Requires sophisticated perceptual abilities
Mirror neurons (Gallese et al, 1996)
Shared/joint attention
Scaffolding
• Natural language
– Essential for human development
– Can emerge from interaction among robot agents
Summary
• Human-like robots
– Has potential, but are not very human-like
– Haptic system, mouth region, tongue, speech system,
advanced muscle-tendon system
• Design principles for developmental systems
–
–
–
–
–
Time-scale integration principle
Development as an incremental process principle
Discovery principle
Social interaction principle
Motivated complexity principle
Place cells, grid cells, head-direction
cells, border cells, and more..
A manifold of spatial maps in the brain
Derdikman & Moser, TICS, 2010
Place cells
• Fire specifically when animal is
at certain position
• Allow read-out of position
• Located in the hippocampus
• Strongly influenced by distal
sensory cues
• Active both in light and dark
– i.e. not dependent on vision only
Place cells
• Can also encode
– Olfactory information
– Time intervals
• Remapping can occur when environment is changed
Grid cells
• Multiple firing locations, forming a
hexagonal grid of the entire space
• Located in the parahippocampal
cortex (input/output to hippocampus)
• Grid space, orientation and phase
• Modulated by velocity
KTH research contributing
to grid cell research
Head-direction cells
• Cells that only fire when head is
pointing in certain direction
• Located in the parahippocampal
cortex
Border cells
• Cells that only fire when animal is
close to a geometric boundary
• Located in the parahippocampal
cortex
• Discovered in 2008
Multi-compartment environment
Rapid transition between
spatial maps
Development of spatial maps
Critical spatial representation components are already present when
rat explores open environment outside the nest for the first time
(Langston et al, & Wills et al, Science, 2010)
• Head-direction cells:
– present 2 days after eye-opening
• Place cells:
– present 2 days after eye-opening
• Grid cells:
– present 4-8 days after eye-opening
• In line with Immanuel Kant’s theory that the concept of
space does not require experience
Questions remaining
• How does place cells relate to grid cells?
• Do humans have neurons with similar functions?
– Place cells (Ekstrom et al, 2003)
– Grid cells (Doeller et al, 2010)
– Path cells (Jacobs et al, 2010)
• Clockwise/counterclockwise paths
• Neurosurgical patients in VR-game
• Entorhinal cortex
Mouse place cells in virtual reality experiment
Virtual reality environment created
with Quake2 video game engine
youtube
Mouse place cells in virtual reality experiment