Big Data 2015 1Michigan State University
Introduction to How Brains Deal with Big Data
Juyang WengProfessor
Dept. of Computer Science and Engineering,Cognitive Science Program, and Neuroscience ProgramMichigan State University, East Lansing, Michigan USA
http://www.cse.msu.edu/~weng/
Big Data 2015 2Michigan State University
Brain Myth is Like Blind Men and an Elephant
Brain
Electrical Engr.
Computer Sci.Biology
Neuroscience Psychology
Math.
Big Data 2015 3Michigan State University
“Big Data” is Only a Hype without It
Big Data
Big Data 2015 4Michigan State University
Must Each Brain Learn from Its Body?
SAIL
Flying Fox(AVS)
Crosser
Dav
Big Data 2015 5Michigan State University
Why Active Body? Kitten Carousel Experiment
A classic study by Held & Hein 1963 Kittens raised from birth in total darkness When old enough to walk, placed in
“kitten carousel” for 42 days One kitten harnessed to pull the carousel Another just being carried in a box.
Big Data 2015 6Michigan State University
Quiz: Kitten Carousel Experiment
Quiz: Your predicted important outcome from this experiment
A. The two kittens are different: one is stronger
B. The active kitten refused to work further later
C. Only the passive kitten learned to see because it has time
D. Only the active kitten learned a critical visual meaning
E. Both kittens learned to see
Big Data 2015 7Michigan State University
Quiz: Kitten Carousel Experiment
Quiz: Your predicted important outcome from this experiment
A. The two kittens are different: one is stronger
B. The active kitten refused to work further later
C. Only the passive kitten learned to see because it has time
D. Only the active kitten learned a critical visual meaning
E. Both kittens learned to see
Big Data 2015 8Michigan State University
Visual Cliff Visual cliff:
A transparent platform Visual sharp drop in elevation
Human infants: 6 – 8 months old, a week or two after they
began to crawl all would cross a visual cliff in initial trials They became increasingly reluctant to cross
in later trials, although nothing bad had happened during crossing.
Carousel kittens: Passive one does not fear Active one does
Implication: Vision is very much developed from
experience!
Big Data 2015 9Michigan State University
Brains Take Flood of Data!
Wu, Guo, Wang, Weng, ICBM 2013
Big Data 2015 10Michigan State University
How Did Your Brain Learn to Segment?
Wang, Wu and Weng, IJCNN 2011
(a) Bottom-up input to a neuron. (b) True object contour. (c) Estimated synaptogenic factor
Big Data 2015 11Michigan State University
Neuroscience: Status Quo
“Data rich and theory poor”
Social habit:“We still do not know how the brain works”(wrong)
“Wetware” work:A lot of advances but in tiny pieces
Kandel, Schwartz and Jessell 2000
Big Data 2015 12Michigan State University
V1: Not Really Orientation?
Hubener et al. Journal of Neuroscience 1997
Orientation map
Ocular dominance map
Each area has thebottom-up input part(e.g., from retina) and top-down input part (e.g.,concepts from motor) but this work did not consider the top-down part
White: ipsilateralBlack: contralateral
A: anteriorP: posteriorM: medialL: lateral1mm
Big Data 2015 13Michigan State University
Innate Edge Detectors?
Kittens are exposed to vertical (or horizontal) edges only after birth
Visually blind to horizontal edges (e.g., no startle response)
No V1 neural cells were found to respond to horizontal edges
Blakemore & Cooper, Nature 1970
Big Data 2015 14Michigan State University
Kittens: Eye Closed Early in Life
David H. Hubel and Torsten N. Wiesel, Journal of Physiology, 1970
Normal
Big Data 2015 15Michigan State University
Brain Self-Wires
Sur, Angelucci and Sharm, Nature 1999
Ferrets rewired early in life
“See” using the “sound” zone
Big Data 2015 16Michigan State University
Early Blind Human:Visual Zone for Hearing and Touch
Trans Magnetic Stimulation (TMS) to the occipital area (normal visual area) hampers the early blind for Sound localization Verbal memory Braille identification
Big Data 2015 17Michigan State University
Conclusion: Brains Constantly Wire Themselves as a Whole
Kittens
Ferrets
Humans
Big Data 2015 18Michigan State University
History of Intelligence ResearchNatural 500 BC
Socrates, Plato, Aristotle 1500’s
Leibniz, Locke, Descartes 1800’s
Pavlov, Thomdike 1900’s
Rumelhart, McClelland
Artificial
1900’s Turing
Big Data 2015 19Michigan State University
The “Task Specific” Trap Treat brain as
“monolithic” pattern classification only
Cannot attend an object from cluttered scenes
Hand-crafted control flow in the “brain”
Task-specific: 7 tasks
Cannot “develop”
Eliasmith et al. Nature, 2012
Big Data 2015 20Michigan State University
AI: Symbolic School vs Connectionist School
Symbols are logic and clean.Artificial neural networks are
analogical and scruffy.
- Marvin Minsky, 1991
(Artificial) neural networks do not abstract well.
- Michael Jordan, IJCNN 2011
Big Data 2015 21Michigan State University
Now-Popular Deep Learning Neocognitron: Fukushima 1980 Cresceptron: Weng at al. 1991
Res-Reduction + Max-pooling HMAX: Poggio et al. 1997 Deep convolution nets with back-
prop 1998 - presentLeCun, Hinton, and others
Popularized by Google. Thanks! Brain’s circuits is NOT a cascade of
modules! Why? Computational reasons …
Big Data 2015 22Michigan State University
Not a Cascade
Felleman &Van Essen 1991
Big Data 2015 23Michigan State University
Intelligence: Symbolic vs. Connectionist
Symbolic 1950’s
logic and clean 2000’s
Common-sense knowledge base
2010’s
Connectionist 1950’s
Analogical and sruffy 2000’s
Autonomous Development
2010’s
Unification of the Two SchoolsGreat interests from industries
Big Data 2015 24Michigan State University
Must Auto-Develop Autonomous Mental
Develop (AMD) Task-nonspecific “Genome-like”
Developmental Program(only about 2 pages long)
IEEE ICDL Conferences IEEE Transactions on
AMD WWN-1 through WWN-9
Weng et al. Science 2001
Big Data 2015 25Michigan State University
8 Requirements for Practical Learning
Eight necessary operational requirements:1. Environmental openness: cluttered environments2. High dimensional sensing (e.g., video cameras are necessary)3. 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 methods at ICDL conferences aimed at some, but not all. Our work
SAIL (1998 – 2010) dealt with the 8 requirements altogether DN (2008 – present) is further brain-inspired
Big Data 2015 26Michigan State University
Experiments: Where-What Networks (WWNs) WWN-1 (2008): single object; cluttered scenes, without presegmentation:
from location to type (recognition task) and from type to location (detection task) using the same network for the two tasks
WWN-2 (2010): add to above free viewing WWN-3 (2010): add to above multiple objects WWN-4 (2010): allowing bypass MM and PP WWN-5 (2011): add to above scale WWN-6 (2012): synaptogenic factors enable neurons to self-wire WWN-7 (2013): add to above multiple parts of each object WWN-8 (2013): multi-modality (left-eye right eye), to appear BigData 2015 WWN-9 (2015): relation of objects (A-B group, A plays B, etc.) Texty: natural language and knowledge acquisition from natural text
Big Data 2015 27Michigan State University
NI and AI Unified Unification of
the Symbolic and Connectionist Schools The control of Turing
Machines as Deterministic Finite Automaton
The brain is an emergent Turing Machine
Automatic programming for general purpose
Big Data 2015 28Michigan State University
Brain-Like AI: Theorems Brains are Emergent Finite Automata (FA)
(world: “tape”) The controller of a Turing Machine is an FA DN learns any teacher FA, immediately and error-free Brain Networks are optimal in maximum likelihood Brain Turing Machine:
Self wires, re-wires, and re-wires again … Automatically self-program for general purposes
Future AI: Machines automatically self-program all the time though their “lifetimes”
Big Data 2015 29Michigan State University
Concepts, Abstraction, Invariance
Weng IJCNN 2010
Big Data 2015 30Michigan State University
Dual Optimality: Lobe Component Analysis (LCA)
Weng IJCNN 2010
Big Data 2015 31Michigan State University
Turing Machine (TM)
Control: the transition function of TM
Big Data 2015 32Michigan State University
Universal Turing Machine
The tape has input data
The tape has input program and data
Big Data 2015 33Michigan State University
The Control of TM is an FA
This form is the transition function of an FA Weng IJIS 2015, IJCNN 2015
Big Data 2015 34Michigan State University
FA: Equivalence Classes
consists of all strings that end in state i
All possible strings from letters in
Big Data 2015 35Michigan State University
Construct FA
Typically there are many possible states A humans tries to prune many states, if the
goal is known The problem in the micro-world can be a graph
G FA is a solution to the problem Search algorithms: Dijkstra alg. and A* Search
Big Data 2015 36Michigan State University
How Does an FA Emerge?
Teacher FA: Accept the language
Big Data 2015 37Michigan State University
Traditional FA: Symbolic Look-Up Table
Big Data 2015 38Michigan State University
Brain-Like Emergent FA: Patterns Only
Big Data 2015 39Michigan State University
Developmental Network (DN): Wire Incrementally
Y Y Y Y
Big Data 2015 40Michigan State University
Basic Theory A brain is a finite automaton, but it emerges The most basic version:
Y area (brain) wires to predict both Z and X Z: concepts, actions, goals, intents, and so on X: expected input: mental images. The EFA thinks! Motivation: important events to speed up learning
Y Y Y Y
Big Data 2015 41Michigan State University
General Vision Problem Solved in Principle:Where-What Network 7
10 objects trained and (disjointly) tested Concept and rules: Location, Type, Scale, Relation
Wu, Guo, Wang, Weng, ICBM 2013
Big Data 2015 42Michigan State University
Ach and NE Transmitters: Self Segmentation
Wang, Wu and Weng, IJCNN 2011
(a) Bottom-up input to a neuron. (b) True object contour. (c) Estimated synaptogenic factor
Big Data 2015 43Michigan State University
Accuracy of the Network
Big Data 2015 44Michigan State University
Relation: Group of Objects Relationship of two animals
A-B Group, A plays B, A gives B to C Knowledge hierarchy be automatically built
Q. Guo, X. Wu & J. Weng, IJCNN 2014
Big Data 2015 45Michigan State University
Invariant Concepts Emerge
Weng IJCNN 2010It emerges with many areas
not a single layer!
Input image
Big Data 2015 46Michigan State University
Earlier Later Neurons Early neurons: X and Z as input Later neurons: Z only as input
Guo, Wu & Weng, IJCNN 2014
Big Data 2015 47Michigan State University
Experiments and Results Training:
individual objects first object groups next
Big Data 2015 48Michigan State University
DN for Stereo VisionINNS Big Data: This Sunday, 11:00am – 1pm
Big Data 2015 49Michigan State University
DN for Vision-Guided NavigationINNS Big Data: This Monday, 3:30pm – 5pm
Big Data 2015 50Michigan State University
Pitfalls in Brain Projects Natural Intelligence: Generate more data, do
not care about basic automata theory Neural Networks as AI: Use traditional
methods to process brain data or “big data” Risk: Much tax-payers money not well used Suggestion: Every brain project must have a
clear plan to educate all researchers for 6 disciplines: biology, psychology, neuroscience, electrical engineering, computer science, and math
Big Data 2015 51Michigan State University
The Major Obstacles of Brain Projects
“More data”: There are sufficient data for humans to understand how the brain works(Very few people see this)
Major obstacles:Habit: Get money and do my own traditional workLack: No degree program for understanding the brain
Suggestions:Fund new computational brain degree programsInvest to get ahead of the upcoming brain tech wave
Big Data 2015 52Michigan State University
Brain-Mind Institute: 2012 - Present
Big Data 2015 53Michigan State University
Brain-Mind Institute Courses BMI 811 Biology for BMR (2012) BMI 821 Neuroscience for BMR (2012) BMI 831 Cognitive Science for BMR (2013, 2014, 2015) BMI 861 Brain Automata (2015) BMI 871 Computational Introduction to Brain-Mind (2012,
2013, 2014, 2015)
BMI Course Packs are now available: including video lectures, ppt files, exams grading, BMI course certificate (if you pass)
Big Data 2015 54Michigan State University
Near-Term Applications Mobile phone / wearable computing based:
Elderly with dementia Children to school Blind and poor vision Drunk walkers
Internet service based Images, video, natural language processing …
Move from lab experiments to development of a new kind of products
Big Data 2015 55Michigan State University
Long-Term Applications Developmental Robots will live Developmental Robots will outlive human
individuals The “brains” of developmental robots will be
copied many times to process various internet data
How many years do we need to wait for that to happen?
Big Data 2015 56Michigan State University
Thank you!Another tutorial by Weng
Introduction to Brainsas Emergent Turing Machines,
9am – noonthe following Thursday, Aug. 13
ICDL-EpiRob 2015, Providence, RI, USA