Brain-Based Robots A Means to Creating More Intelligent Machines - - PowerPoint PPT Presentation

Brain-Based Robots A Means to Creating More Intelligent Machines

Jeff Krichmar

Cognitive Anteater Robotics Laboratory (CARL) Department of Cognitive Sciences Department of Computer Science

1

Who is Better at Recognizing Objects?

3

Who is Better at Cooperating?

4

Which can Distinguish Self from non-Self?

5

Who can Fly Better?

6

All These Intelligent Creatures Have Brains

Species' Neurons' Synapses' Nematode( 302( 103( ( Fruit(Fly( 100,000( 107( ( Honeybee( 960,000( 109(

( (

Mouse( 75,000,000( 1011( ( Cat( 1,000,000,000( 1013(

(

( Human( 85,000,000,000( 1015(

(

(

Source(–(hCp://en.wikipedia.org/wiki/List_of_animals_by_number_of_neurons(

7

8

9

10

11

Properties of Neurons Action Potential & Synapse

12

Brain Computations

• Massive parallelism (1011 neurons)
• Massive connectivity (1015 synapses)
• Excellent power-efficiency

– ~ 20 W for 1016 flops

• Low-performance components (~100 Hz)
• Low-speed comm. (~meters/sec)
• Low-precision synaptic connections
• Probabilistic responses and fault-tolerant
• Autonomous learning

13

14

Brain Inspired Robots

• Developing a system that

demonstrates some level of cognitive ability can lead to a better understanding of cognition.

• Building a robot or artifact that

follows a cognitive model could lead to a system that demonstrates capabilities commonly found in the animal kingdom, but rarely found in artificial systems.

15

SyntheNc(methodology( Understanding(through(building(

16

Design Principles for Neurorobots

• Engage in a behavioral task.
• Behavior controlled by a simulated nervous system

that reflects the brain’s architecture and dynamics.

• The world is an unlabelled place.

– Organize the signals from the environment into categories without a priori knowledge or instruction.

• A value system that signals the salience of

environmental cues to the robot’s nervous system.

• Needs to be situated in the real world.
• Behavior and activity of its simulated nervous

system must allow comparisons with empirical data.

Krichmar,(J.L.,(and(Edelman,(G.M.((2005).(Ar#ficial)Life,)Vol.(11,(63W78.(

17

Morris Water Maze: A Test of Spatial Cognition

• Allows comparison of the robot’s behavior with biological cognition.
• Need to demonstrate that cognitive robots can transition to the real

world just as humans or animals do when they are outside a laboratory setting.

18

Krichmar,(et(al.((2005)(( Proc(Natl(Acad(Sci)102,(2111W2116.(

Cognitive robot that transitioned beyond the lab:

The RatSLAM Project: Robot Spatial Navigation

19( Wyeth,(Milford,(Schulz,(&(Wiles(in(Neuromorphic)and)Brain:Based)Robots,)CUP,)2011,)pp.)87:108.(

Morphological Computation and Cheap Design

• The construction and design of agents that are

built to exploit properties of the ecological niche will be much easier or “cheaper”. Pfeifer and Bongard, 2006.

20

Example of Morphological Computation

• Passive Walkers
• Cornell Ranger

– Walked 40.5 miles without being recharged or touched by a person – The coordination of the walking was by the 6

• nboard microprocessors.

– Each step it falls and catches itself in a controlled manner. – Took 186,076 steps while

• nly using 5 cents worth of

electricity.

21

Example of Cheap Design

Segway soccer ball capture mechanism

22

Fleischer,(J.(G.,(et(al.(A(neurally(controlled(robot(competes(and(cooperates( with(humans(in(Segway(soccer.(ICRA(2006.(

Organisms Adapt Their Behavior Through Value Systems

• Non-specific, modulatory signals to the rest of the brain.
• Biases the outcome of synaptic efficacy in the direction

needed to satisfy global needs.

23

Vertebrate'Neuromodulatory'Systems'

Noradrenergic( Cholinergic( Dopaminergic( Serotonergic( Edelman,(G.M.((1993).(Neural(Darwinism:(SelecNon(and(reentrant(signaling(in(higher(brain( funcNon.(Neuron)10,(115W125.(

Neuromodulation as a Robot Controller

24

Cox(&(Krichmar,(IEEE(RoboNcs(&(AutomaNon(Magazine,(September(2009.((

OrienNng(toward(a(posiNve(value(object(–(Dopaminergic(response(

Neuromodulation as a Robot Controller

25

Cox(&(Krichmar,(IEEE(RoboNcs(&(AutomaNon(Magazine,(September(2009.((

Withdrawing(from(a(negaNve(value(object(–(Serotonergic(response(

A Biologically Inspired Action Selection Algorithm Based on Principles of Neuromodulation

• General-purpose control system that ties environmental events to a

wide-range of values.

• Adapts to novel and familiar environments by trading off between being

anxious and curious.

26

Krichmar, J.L. (2012) International Joint Conference on Neural Networks (IJCNN). Krichmar, J.L. (2013) Frontiers in Neurorobotics.

Case Study for Neurorobot Design Navigating a Cluttered Scene Using Vision

hCps://www.youtube.com/watch?v=UEIn8GJIg0E(

Crossing a busy intersection in Ethiopia Walking through a crowd at the San Diego County Fair

27

Visual Motion Pathway

Dorsal Stream (Macaque) Spatial Localization and Action

• Primary visual cortex (V1)

– Tuned to simple attributes of shape, motion, color, texture, depth.

• Middle temporal (MT) area

– Tuned to coherent local motion (retinal flow)

• Medial Superior Temporal (MST) area

and Ventral Intra-Parietal (VIP)

– Tuned to global, complex motion. – Self-motion and object motion. – Multimodal.

(Britten,)2008))

28

V1 and MT Model

• Spatiotemporal-energy model of V1

– Bank of linear space-time oriented filters (rate-based).

• Adapted from Simoncelli & Heeger, 1998.

– Direction-selective cells. – Fully realized in CUDA.

• Two-stage spiking model of MT

– Izhikevich spiking neurons: regular- spiking / fast-spiking

• 153,216 neurons.
• ~40 million synapses.
• Runs in real-time with video.

– Component Direction Selective cells. – Pattern Direction Selective cells:

• Direction pooling + opponent

inhibition.

• Signals the global pattern of motion.
• Solves the aperture problem

Retina spiking LIP

50 Hz 0 Hz

MT

50 Hz 0 Hz 50 Hz 0 Hz

V1 rate-based

50 Hz 0 Hz 100 Hz 0 Hz

... ... ... ... (eqn. 12) (eqn. 13)

29

Model Response to Motion Patterns

Component and Pattern Selectivity

Component6 direc7on6 selec7ve' Pa:ern6 direc7on6 selec7ve'

30

Visually Guided Robot Navigation

Architecture and I/O

ABR server

Image frames (UDP) Servo commands (TCP) WiFi/3G 320x240px ~30fps

ABR client

Cortical model

Obstacle component Goal component

goal

RGB

320x240

V1 LGN

gray 30x80

MT PPCl PPCr

Steering control

Le Carl

ABR(=(Android(Based(Robot( hCp://www.socsci.uci.edu/~jkrichma/ABR(

31

Visually Guided Robot Navigation

Server Control GUI and Results

32

Visually Guided Robot Navigation

By a Spiking Neural Network of Visual Cortex

33

Comparison to Psychophysical Data

• Dotted lines are human trajectories

– Replicated with dynamical system by Fajen & Warren, 2003; 2007. – Comparable to neural simulation by Browning, Grossberg & Mingolla.

• Colored lines are robot trajectories.

34

Visually Guided Robot Navigation

• Demonstrated that large-scale cortical spiking

neural network can:

– Replicate neurophysiological and psychophysical findings. – Control autonomous robot in real time.

• First step toward a complete embedded visual

navigation system.

– Simulated MT responses are sufficient for obstacle avoidance.

• Brain based computing compatible with new

neuromorphic hardware that is smaller, faster, and more efficient than conventional computers.

35

36

Open Issues

• Most cognitive systems

– Exist in sterile, highly controlled laboratory settings. – Rely too much on neural control driving the body and behavior instead of the other way around.

• Failing to make systems that perform more

than one function at a time.

37

Challenge Problem that Requires Full Suite of Cognitive Behaviors

Disaster Relief, Search and Rescue

38

Search & Rescue with Smartphone Robots

• Cognitive mapping
• Navigation
• Object recognition
• Decision making
• Cooperation

39

Team CARL

Hirak( Kashyap( Tiffany( Hwu( Stas( Listopad( Back(row(from(lej(to(right:(Alexis(Craig,(Alex(Wang,(Michael(Beyeler,( Feng(Rong,(Timo(Oess,(Saideep(Gupta.( Front(row(from(lej(to(right:(Emily(Rounds,(Steve(Doubleday,(Jeffrey( Krichmar,(TingWShuo(Chou,(Nikil(DuC.((

40

The Brain is Embodied and the Body is Embedded in the Environment

• Understanding how the brain’s mechanisms give rise to perception,

cognition, emotion, action and social engagement will have a revolutionary impact on science, medicine, economic growth, security, and social wellbeing.

• By developing neural models that follow the architecture and

dynamics of brain networks, we can create a class of robots that

• perate more like humans and other biological organisms.

41

Thank You!!

• More information can be found at:

– http://www.socsci.uci.edu/~jkrichma

42