Transcript Slide

“CHINA BRAIN” PROJECT
Building China’s First Artificial Brain
Prof. Dr. Hugo de GARIS
Director, Artificial Brain Lab,
Cognitive Science Department,
School of Information Science and Technology,
Xiamen University, Xiamen, CHINA.
[email protected]
Prof. Dr. Hugo de Garis, Director of China’s first Artificial Brain Lab,
Cognitive Science Dept, Xiamen University, Xiamen, CHINA
Artificial Brains – the Context
The next decade or so will be exciting because :a) By 2020, one bit of information will be stored on a single atom.
b) That means we will have “Avogadro Machines” (with 1024 parts)
c) Neuroscience is now learning the secrets of the micro-circuitry
of the rat brain, e.g. the revolutionary work of Henri Markram.
d) This micro-circuitry knowledge can be put into Avogadro Machines
and speeded up a million times (electronic vs. chemical speeds).
e) This marriage will provide the technologies for building artificial
brains (e.g. to control home robots).
f) Home robots (AB controlled), world’s biggest industry by 2030.
Markram’s simulated rat brain cortical column. The rat has a
few 1000s of such columns. Humans have about a million.
Markram’s work creates “e-NS” (neuroscience in a computer).
“CHINA BRAIN” PROJECT
BASIC IDEA
Evolve 10,000s of neural net modules (in real time?) to control the
100s of behaviors of a NAO (French, Robocup standard) robot.
So an artificial brain is defined here to be a “network of (evolved
neural) network modules”.
Each module performs some simple task (e.g. recognize someone’s
face, detect the color of an object, detect the direction of motion, …..),
rather like Minsky’s “agents” in his “Society of Mind”.
After the modules are evolved they are connected to make an artificial
brain (according to the designs of human “BAs” (Brain Architects)),
whose neural signaling is performed by the (IMSI) operating system.
Six Basic Components to this Project
a) Parcone (Partially Connected Neural Evolutionary) model
software used to evolve neural net pattern recognition modules
b) NVidia PC Supercomputer, Tesla S1070, used to evolve
neural net modules in (hopefully) real time, i.e. in about one second
c) Operating system (IMSI, Inter Module Signaling Interface), used
to perform the neural signaling of all the modules in the brain
d) NAO robot, to be controlled by the artificial brain
e) Language component, NAO robot talks, listens, understands,
obeys simple commands, answers questions
f) Consciousness component, self knowledge of own body, memory
a) PARCONE
An essential ingredient in the China Brain Project is the software
that is used to evolve neural net modules, called Parcone.
The artificial neurons evolved by this software are partially connected.
This allows users to specify large numbers of input, middle and
output neurons (as parameters), e.g. an image of 60*60 (RGB) pixels.
This implies 60*60*3 = 10,800 input neurons. (If the neural net were
fully connected => 100,000,000s of connections, too large!)
Automated Evolution of Positive and Negative Images:
Users supply M positive images (of the pattern P to be learned) and
N negative images (i.e. not P). The module is evolved to give strong
positive output signals to Ms, and strong negative signals to Ns.
Population
Chrm is a
pointer to a
population of
chromosomes
NList
NListPtr
Chrm
A chromosome is a
pointer to NList
NList is a list of
pointers to each
neuron’s list of
neurons pointed to.
HashPtr
Hash
Neuron ID
NeuronWt
struct
A Hash is a hash table
of pointers to a NeuronWt
struct
Weight Bits
Weight Value
Data Structures of the “Parcone” Neural Net Model
Evolution Set
Test Set
60*60 (pixel)*3 (RGB) = 10800 pixel values
= 10800 input neurons to a Parcone module
Parcone Experimental Results :
Face recognition (90%-95% accuracy).
Color detection
Motion detection
Hubel-Wiesel (slanted moving line of light) detection
Keys, shoes, ….
Counting (“3-ness”, i.e. if 3 objects, the module signals)
etc, etc….
Parcone is a powerful pattern detector. We plan to evolve 1000s
of pattern detector modules for our artificial brain.
b) The NVidia PC Supercomputer, Tesla, S1070
This PC supercomputer contain 960 processors, all working in
parallel, at a total speed of 4 teraflops, at the remarkable price
of only $10,000, a PC revolution!
We are currently learning how to program it (with a version of
the C language, called “CUDA”).
4 teraflops is about 4000 times faster than an ordinary PC.
It takes about 30 minutes to evolve one Parcone module on a PC.
We therefore hope using the Tesla, that we can do the same in
about one second, i.e. real time evolution.
NVidia’s PC Supercomputer, Tesla S1070
960 processors, 4 teraflops, $10,000
We hope to get real time evolution from it!
Real Time Percept Learning
If the PC Supercomputer truly allows us to evolve a neural net
pattern recognizer module in real time (i.e. about 1 sec.)
then real time percept learning becomes possible.
We imagine that the NAO robot can behave like a human baby,
learning constantly.
When a new object is shown to the robot, it checks all its previous
pattern recognition circuits. If none of them recognize the object,
i.e. all the pattern detector modules give weak output signals, then,
a new pattern detector module is evolved and stored in memory.
The robot brain will be constantly learning!
Exciting prospect!
c) Operating System (IMSI)
“IMSI” (Inter Module Signaling Interface), is the software
operating system that performs the neural signaling of the
whole artificial brain (of typically 10,000s of modules in real time)
IMSI performs 4 major functions
i) Specifies connections between modules
ii) Calculates the neural output signals of all neurons in the brain
iii) Interfaces with NOA’s “Choregraphe” motion control software
iv) Executes simple functions with ordinary programming routines.
i) Specify connections between modules
IMSI prompts users (i.e. “BAs” (Brain Architects)) to specify how
the evolved modules are to be connected to build an artificial brain.
Each evolved neural net module has a unique integer identifier (ID).
So too does each input neuron and output neuron of a module.
e.g. a user can connect the 4th output neuron of module 48234 to
the 3rd input neuron of module 29458, with :(48234, 4) => (29458, 3)
These connection data are stored in look up tables in the IMSI, that
are consulted when IMSI calculates the output signals of all neurons.
ii) Calculate the strengths of the neural output signals
IMSI’s main job is to calculate the strengths of the neural output
signals of all neurons in all the neural net modules in the artificial
brain. To do this it uses the hash tables of each neuron, as well as
the look up tables of the connections between modules.
iii) Interfacing NAO robot’s “Choregraphe” motion control software
The 100s of motions of the NAO robot are programmed by special
software provided by “Aldebaran” (the French company that
manufactures the NAO robots) called “Choregraphe”. Choregraphe
generates (programmable) time dependent angle vectors, for the 25
angles of the robot’s motors.
iv) Simple-function programming modules
Simple functions (e.g. “AND”) are executed in ordinary code.
d) NAO Robot
Noone will “see” the artificial brain. It’s just a set of weights and
connection numbers in a computer’s memory.
The intelligence and usefulness of the artificial brain will be judged
by the behaviors of the robot it controls.
The robot we chose for the artificial brain to control was
France’s “NAO” (Chinese for “brain” – coincidence? marketing?)
robot, waist high, 2 legs, 2 arms, two fingers and thumb, 2 camera
eyes, voice and microphone.
The NAO robot (made by Aldebaran, Paris, France) has its own
software, called “Choregraphe” to control the time dependent angle
vectors of its 25 motors.
e) Language Processing
Prof Ben Goertzel is working closely with Prof Hugo de Garis
on the China Brain Project.
Ben Goertzel’s role is to supervise two of the components of the
project, namely language processing and consciousness
Language :
The NAO robot will be capable of NLP (Natural Language
Processing), listening, understanding, obeying commands,
answering questions, talking.
e.g. “Go the door”, “Who is this?”, “It is Mr. X”, “Who are you?”
“I am NAO, an artificial brain controlled robot”. “Where is the
green chair?” “In the corner that I’m pointing towards”.
f) Consciousness
The NAO robot is to be given some measure of consciousness or
self knowledge,
e.g. recognizing parts of its own body, its reflection in a mirror,
meta knowledge of its own state, etc.
Hence, Prof de Garis’s “territory”, is the low level domains of
perception, motion control, operating system design, etc
Prof Ben Goertzel’s is the higher level domains of language,
consciousness, reasoning, logic, etc.
The NAO robot will be simulated so that anyone on line can
teach the NAO simulated robot. This knowledge can then be
used to control the real world NAO robot, => world project !
Other Developments
a) Special Issue on “Artificial Brains” for the Neurocomputing
journal, to appear 2010.
Guest editors : Prof. Dr. Hugo de GARIS
Prof. Dr. John TAYLOR
Prof. Dr. Ben GOERTZEL
b) Book (contracted by World Scientific) on
“Artificial Brains : An Evolved Neural Net Module Approach”
to appear 2010.
Author :
Prof. Dr. Hugo de GARIS
c) Dr. Ben Goertzel made a guest professor at Xiamen University
d) Chinese NSF Funding Proposal : “A Humanoid Robot that
Learns via Imitation and Reinforcement”
Proposers : Prof. Dr. Hugo de GARIS, Prof. Dr. Ben GOERTZEL
“CABA”
CABA (i.e. Chinese Artificial Brain Administration),
(for USA) NABA (National Artificial Brain Administration)
Similar to America’s NASA, i.e. a government run administration
with 1000s of scientists and engineers aimed at building
Artificial Brains for the country’s artificial brain industry,
especially for the home robot industry.
Prof de Garis is pushing the Beijing government to invest heavily
in the creation of a CABA. There are strong reasons for this.
a)
b)
c)
d)
e)
One of the world’s biggest industries by 2030 (economics)
Fascinating problem (science)
Dominating this industry => national pride (psychology)
International rivalry (China-US-EU-etc) (politics)
“Species dominance debate” (philosophy)