Transcript MIT lecture

Computing Architectures
The human brain
as computing system
Based on presentation from
http://www.stanford.edu/class/symbsys100/ and
http://www.willamette.edu/~gorr/classes/cs449/brain.html
Plan
From symbols to meat
Meet the brain
Brains vs. digital computers
Bio-inspired computers
Reasoning module: concluding
discussion
Motto
Human cognition is based on a very specific
computing system, with specific limits, inherent
trade-offs, etc. that are not necessarily the same
as for digital computers
It is therefore worth looking at the "mind's
implementation" in order to learn more about
the limits of our mind/cognition
Plan
From symbols to meat
Meet the brain
Brains vs. digital computers
Bio-inspired computers
Reasoning module: concluding
discussion
The brain – just 2 pounds of
meat?
The cortex
1.3-1.4kg (2% of the body weight) … [13,14]
2,500 cm2 (rat: 6 cm2, elephant: 6,300 cm2) [14]
1,300-1,500 cm3
2 hemispheres connected by corpus callossum
(250 mill. nerve fibers)
Inputs:
spinal cord
optic nerve (1.2 mill.)
cranial nerves (12)
auditory system, …
The lobes
4 lobes: occipital, parietal, temporal,
frontal
Occipital: vision
Parietal: touch, pressure, temperature, pain
Temporal: auditory information, long term
memory
Frontal: short term memory, planning,
emotion, movement…
Biggest difference from our closest evolutionary
ancestors
Taken from
http://www.sciencebob.com/lab/bodyzone/brain.html
Neurons [14]
 100 billion neurons (children)
300 million – octapus;
18,000 – sea slug Aplysia; 350 - leech
 Diameter: 4 – 100 microns
 Weight: 10-6 grams
 Length: <1 mm – 4 feet (in the leg) [15]
Length of Giraffe primary afferent axon:
15 feet
 Loss of neurons: ~1/sec  31 million/year
 an octapus/10 years
~5,400 at the end of this lecture (sorry!)
How do we know?
Non-invasive (1mm3  ~ 6-7*104 neurons)
EEG (Electroencephalogram),
ERP (Early receptor potential)
fMRI (blood flow; ~1mm; secs-mins)
MEG (Magnetoencephalogram with ERP: ~1.5mm; msecs-secs)
PET (imaging technique blood flow; 1mm; >mins)
Invasive methods: electrodes (1 neuron; msecs)
Lesions
Permanent: injury, disease
Temporary: specific drugs, TMS (<1mm; <secs)
All methods have trade-offs (spatial, temporal resolution)
The brain as a computational
system
The brain is
biological
de-central (plasticity)
non-digital
highly parallel
What does this mean?
The brain: a biological CS
not manufactured from scratch with a
certain intention in mind, but subject to
evolution
Co-adaptation; its parts must have been of use
Not made out of copper or lightconduction cables ....  slow
Signal speed: MAX=120m/s, AV.=6.5m/sec
(1.2 - 250mph) [14]
Signal frequency: up to 1000Hz (activ./sec)
Non-digital
At least to some degree, the brain is nondigital
On the lowest level (i.e. within the neuron):
quasi-digital
this creates an analog signal travelling
along the neuron
at the synapses this is converted into a
chemical signal, which in turn triggers an
elecrical signal.
The brain: a highly parallel CS
Some neurons have up to 150,000 connections
(others as low as 2)
 average: 1,000-10,000 [14]
different brain regions are highly
interconnected
human can manage many tasks at the same
time (sitting, listening to the lecture, doodling)
however, there are also parts of the brain which
are involved in a lot of tasks  "narrow
passages" for computation
Plan
From symbols to meat
Meet the brain
Brains vs. digital computers
Bio-inspired computers
Reasoning module: concluding
discussion
Storage capacity of the brain - I
100 billion neurons
1011! hypothetically possible connections
upto 150,000 connections between each
neuron (180,000km of myelinated nerves)
during the first year of life, the child
generates ~ 15,000 connections for each
neuron (during growth: 250,000 per minute!)
“… this program will support more than 130,000 [i.e.
1.3 * 105] neural connections…”
Storage capacity of the brain - II
# bits = # of neurons * # of connections
1 * 1011 * 1.5 * 105 = 1.5 * 1016 bits
The entire Enc. Britannica contains 109
bits of information (Turing 1950)
In 1987, Hideaki Tomoyori memorized
the first 40,000 digits of π
Information processing speed
of the Brain
# bits/sec = # ops/sec* # bits/op
10 ops/sec per synapse (connection) [3,4]
~1.5 * 1017 bits/sec information transfer
Estimates of the brain's computing power
range from 1011 to 1020 bits/sec
Converging evidence for ~ 1015 [2,3,5,9,14]
~100 teraflops (8 bit words); ~ 8 teraflops
(128 bits words)
Brain vs. digital Computers
Fastest computer atm:
40 terra flops (5,000 processors; NEC)
Planned
360 terra flops (130,000 processors; IBM)
~ 3-4 times faster than the human brain
(8 bit words); 40 times faster otherwise.
Plan
From symbols to meat
Meet the brain
Brains vs. digital computers
Bio-inspired computers
Reasoning module: concluding
discussion
Bio-inspired models of
computation
This gives us a motivation to investigate
bio-inspired models of computation
Learn about the brain by modeling it
Take advantage of billions of years of
evolutionary design
Develop robust computational systems
Neural networks
So what?
 “It is true that a discrete-state machine must be different from a
continuous machine. But if we adhere to the conditions of the
imitation game, the interrogator will not be able to take any
advantage of this difference.” Turing (1950:451)
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
Gazzaniga, Ivry & Mangun (1998): Cognitive Neuroscience. The Biology of the
Mind. Norton.
Merkle, Ralph C. (1988): How many bytes in human memory? at
http://www.merkle.com/humanMemory.html
Merkle, Ralph C. (1989): Energy Limits to the Computational Power of the
Human Brain; at http://www.merkle.com/brainLimits.html
Principles of Neural Science, by Eric R. Kandel and James H. Schwartz, 2nd
edition, Elsevier, 1985
http://www.coping.org/earlyin/ruleout/reason.htm
http://www.jsmf.org/zarticles&pap/John/neural_connections.htm
http://ifcsun1.ifisiol.unam.mx/Brain/neuron.htm
http://ifcsun1.ifisiol.unam.mx/Brain/neuron2.htm
http://www.rfreitas.com/Nano/DeusExDigita.htm
http://www.cheshireeng.com/Neuralyst/nrlnds.htm
http://www.top500.org/
http://www.consciousness.arizona.edu/hameroff/
http://www.neurologicalalliance.org.uk/pages/network/answers.asp
http://faculty.washington.edu/chudler/facts.html
http://www.uncc.edu/sspauldi/LECNote/ch02.html