2011 inns iesnn 1 michigan state university a computational introduction to the brain-mind juyang...

2011 INNS IESNN 1Michigan State University

A Computational Introduction to the Brain-Mind

Juyang (John) Weng

Michigan State University

East Lansing, MI 49924 USA

[email protected]


Human Physical and Mental Development

Studies on the adult brain

Studies on how the brain develops


Machine Mental Development


Totipotency

Stem cells and somatic cells Genomic equivalence:

All cells are totipotent: whose genome is sufficient to guide the development from a single cell to the entire adult body

Consequence: the developmental program is cell-centered


Genomic Equivalence

Each somatic cell carries the complete genome in its nucleus

Evidence: cloning (e.g., sheep Dolly) Consequences:

Genome is cell centered, directing individual cell to develop in cell’s environment

No genome is dedicated to more than one cell Cell learning is “in place”: Each neuron does not

have an extra-celluer learner: cell learning must be fully accomplished by each cell itself while it interacts with its cell’s environment


How to Measure Problems in AI

Time and space complexity? High or low “level”? Tasks that look intelligent when a machine

does it? Rational or irrational? Handling uncertainty? …


Task Muddiness

Independent of problem domain Independent of technology level Independent of the performer: machines or animals Can be quantified Help us to understand why AI is difficult Help us to see essence of intelligence Can be used to evaluate intelligent machines Help to appreciate human intelligence


Task Muddiness

Agent independent Categories only Each category can be extended Categories adopted to model task muddiness:

Environment Input Output Internal state Goal


Environmental Muddiness

Measure Clean Muddy Awareness Known Unknown Complexity Simple Complex Controllability Controlled Uncontrolled Naturalness Artificial Natural Variation Fixed Changing Foreseeability Foreseeable Nonforeseeable


Task Executor

Human agent:the human is the sole executor

Machine agent:Dual task executor A task is given to a

human The human programs

an machine agent The agent executes

QuickTime™ and a decompressor

are needed to see this picture.


A Partial List of Input Muddiness

Measure Clean MuddyRawness Symbolic Real sensorSize Small LargeBackground None ComplexVariation Simple ComplexOcclusion None SevereActiveness Passive ActiveModality Simple ComplexMultimodality Single Multiple


A Partial List of Other Muddiness

Category Measure Clean Muddy Size Small Large Representation Given Not given Observability Observable Unobservable Imposability Imposable Nonimposable

State

Time coverage Simple Complex Terminalness Low High Size Small Large Modality Simple Complex

Output

Multimodality Single Multiple Richness Low High Variability Fixed Variable Availability Given Unknown

Goal

Conveying mode Simple Complex


2-D Muddiness Frame

Size ofinput

Rawnessof input

Languagetranslation

Computerchess

Visualrecognition

Sonar-basednavigation


Composite Muddiness

m = m1 m2 m3 … mn


Autonomous Mental Development (AMD)


Traditional Manual Development

A = H(Ec , T)A: agentH: humanEc: Ecological conditionT: Task


New Autonomous Development

A = H(Ec )Autonomous inside the skullA: agentH: humanEc: Ecological condition


Mode of Development: AA-Learning

AA-learning: Automated animal-like learning

Unbiased Sensors

biased Sensors

Effectors

Closed brain

World


Existing Machine Learning Types

Supervised learningClass labels (or actions) are given in training

Unsupervised learningClass labels (or actions) are not given in training

Reinforcement learningClass labels (or actions) are not given in training but reinforcement (score) is given


New Classification for Machine Learning

Need for considering state imposability after the task is given

3-tuple (s, e, b):symbolic internal representation, effector, biased sensor State: state imposable after the task is given Biased sensor: whether the biased sensor is used Effector: whether the effector is imposed


8 Types of Machine LearningLearning type 0-7 is based on 3-tuple (s, e, b):

Symbolic internal (s=1), effector-imposed (e=1), biased sensors used (b=1)

Type Internal Effector Biased 0 (000) emergent autonomous Communicative 1 (001) emergent autonomous Reinforcement 2 (010) emergent imposed Communicative 3 (011) emergent imposed Reinforcement 4 (100) symbolic autonomous Communicative 5 (101) symbolic autonomous Reinforcement 6 (110) symbolic imposed Communicative 7 (111) symbolic imposed Reinforcement


The Developmental Approach

Enable a machine to perform autonomous mental development (AMD)

Impractical to faithfully duplicate biological AMD Hardware: Embodiment (a robot) Software: A developmental program

Task nonspecific AA-learning mode, from the “birth” time through the

“life” span


Comparison of Approaches

Approaches SpeciesArchitecture

World Knowledge System behavior Task-specific

Knowledge-based Programming Manual modeling Manual modeling Yes

Behavior-based Programming Avoid modeling Manual modeling Yes

Learning-based Programming Models withparameters

Models withparameters

Yes

Evolutionary Genetic search Models withparameters

Models withparameters

Yes

Developmental Programming Avoid modeling Avoid modeling No


Developmental Program vs Traditional Learning

Properties of program Traditional programs

Developmental programs

Sensor-specific and Effector-specific

Yes Yes

Program is task-non-specific No Yes Tasks are unknown at programming time

No Yes

Generate representation automatically [1]

No Yes

Animal-like online learning No Yes Open-ended learning for more new tasks

No Yes

[1] For tasks unknown at the programming time.


Motives of Research for Development

Developmental mechanisms are easier to program:lower level, more systematic, task-independent, clearly understandable

Relieve humans from intractable programming tasks: vision, speech, language, complex behaviors, consciousness

User-friendly machines and robots:humans issue high-level commands to machines

Highly adaptive manufacturing systems (e.g., self-trainable, reconfigurable machining systems)

Help to understand human intelligence


Task Nonspecificity

A program is not task specific means: Open to muddy environment Tasks are unknown at programming time “The brain” is closed after the birth Learn an open number of muddy tasks after birth

Avoid trivial cases: A thermostat A robot that does task A when temperature is high and

does task B when temperature is low A robot that does simple reinforcement learning


8 Requirements for Practical AMD

Eight necessary operational requirements:1. Environmental openness: muddy environments2. High dimensional sensing3. Completeness in internal representation for each age group4. Online5. Real time speed6. Incremental:

for each fraction of second (e.g., 10-30Hz)7. Perform while learning8. Scale up to large memory

Existing works (other than SAIL) aimed at some, but not all. SAIL deals with the 8 requirements altogether


Definition of AA-Learning

A machine M conducts AA-learning if the operation mode is as follows:

For t = t0, t1, t2, ... , the brain program f recursively updates the brain B, sensory input-ouput x and effector input-output z


The Central Nervous System

The forebrain The midbrain

and hindbrain The spinal cord

Kandel, Schwartz and Jessell 2000


Brodmann Areas (1909)

QuickTime™ and aTIFF (Uncompressed) decompressor


Kandel, Schwartz and Jessell 2000


Sensory and Motor Pathways

Adapted from Kandel, Schwartz and Jessell 2000

My hypothesis:Brain has complex networksthat emerge largely shapedby signal statistics (Weng IJCNN 2010)


Multimodal Integration




Weng IJCNN 2010

The brain has only two exposed endsto interact with the environment:

Brain’s Vision System




Triple Loops

Weng IJCNN 2010


Solving the Feature Binding Problem

Weng IJCNN 2010


Area as A Building Block



Weng IJCNN 2010


Neurons as Feature Detectors: The Lobe Component Model

Biologically motivated: Hebbian learning lateral inhibition

Partition the input space into c regions X = R1 U R2 U ... U Rc

Lobe component i: the principal component of the region Ri

Weng et al. WCCI 2006


Different Normalizations




Dual Optimality of CCI LCA Spatial optimality leads to the best target:

Given the number of neurons (limited resource), the target of the synaptic weight vectors minimizes the representation error based on “observation” x:

Temporal optimality leads to the best runner to the target: Given limited experience up to time t, find the best direction and step size for each t based on “observation” u = r x





Weng & Luciw TAMD vol. 1, no. 1, 2009


CCI LCA Algorithm (1)












CCI LCA Algorithm (2)










Plasticity Schedule

t1 t2

t

2

(t)

r = 10000


Natural Images


IC from Natural Images


Temporal Architectures




Based on FA Ideas




From FA to ED network

FA: sn = f(sl,am) s: state; a: symbol input ED:

The internal area learns:yi = fy (sl, am)

The motor area learns: sn = fz (yi)

s: a numeric pattern of z, a sample of Z spacea: a numeric pattern of x, a sample of X spacey: a numeric pattern of y, a sample of Y space


Training and Tests

Luciw & Weng IJCNN 2010


Performance


Three Types of Information Flow

Different directions for different intents

Mixed modes are possible

There is no “if-then-else” type of switches


For any FA there is an ED network

ED: Epigenetic Developer

FS: Finite Automaton

Relation: An ED network can learn any FA

Marvin Minsky at MIT criticized ANNs

Weng IJCNN 2010


Almost Perfect Disjoint TestUsing Temporal Context



Luciw, Weng & Zeng ICDL 2008

More Views, Better Confidence

Externally sensed Internally generated context


For any FA there is an ED network

ED: Epigenetic Developer

FS: Finite Automaton

Relation: An ED network can learn any FA

Marvin Minsky at MIT criticized ANNs

Weng IJCNN 2010


From FA to ED network

FA: sn = f(sl,am) s: state; a: symbol input ED:

The internal area learns:yi = fy (sl, am)

The motor area learns: sn = fz (yi)

s: a numeric pattern of z, a sample of Z spacea: a numeric pattern of x, a sample of X spacey: a numeric pattern of y, a sample of Y space


Complex text processing

New sentence problem Recognize new sentences from synonyms

Word sense disambiguation problem Temporal context

Part of speech tagging problem Label words according to part of speech

Chunking problem Grouping sequences of words and classify them by syntactic labels

Weng, Zhang, Chi & Xue ICDL 2009


Recent Events on AMD

ICDL series: http://cogsci.ucsd.edu/~triesch/icdl/ Workshop on Development and Learning (WDL) 2000, MSU, MI USA 2nd International Conf. on Development and Learning (ICDL’02): MIT, MA USA 3rd ICDL (2004): San Diego, CA USA 4th ICDL (2005): Osaka, Japan 5th ICDL (2006): Bloomington IN, USA 6th ICDL (2007): London, UK 7th ICDL (2008): Monterey, CA, USA 8th ICDL (2009): Shanghai, China 9th ICDL (2010): An Arbor, Michigan USA 10th ICDL (2011), Frankfurt, Germany

EpiRob workshop series, 01, 02, 03, 04, 05, 06, 07, 08, 09, 10 AMD Technical Committee of IEEE Computational Intelligence Society

http://www.ieee-cis.org/AMD/ AMD Newsletters

http:///www.cse.msu.edu/amdtc/amdnl/ IEEE Transactions on Autonomous Mental Development

http://www.ieee-cis.org/pubs/tamd/


Now and Future Now (not many people agree):

Humans start to know roughly how the brain-mind works Future (not too far):

Systematic breakthroughs in artificial intelligence along all fronts: Vision Speech Natural language Robotics Creative intelligence

A new industry: New type of software industry Cloud computing for brain-scale applications Service robots and smart toys entering homes Robots widely used in public environments

2011 inns iesnn 1 michigan state university a computational introduction to the brain-mind juyang...

Documents

inns iesnn

task slide

internal state

cellcentered slide

goal slide

human intelligence slide

cells environment slide

human agent