Why Computers Don't Care - SRC
Transcript of Why Computers Don't Care - SRC
P. Read Montague
Baylor College of Medicine, Houston, TX
Why Computers Don't Care
The Origins of Meaning in Efficient Computation
Part 2. Energy, computation, and meaning
Part 3. Pursuing goals with valuation machinery
Part 1. Structural themes in the nervous system
Why can ideas veto biological instincts for survival?
Theme I: Put data near the processing capacity -maps and integrative regions
Vision
Touch
Hearing
IntegrationIntegration
Theme II: Miniaturization, self-repair, re-configuration
1 mm
1 mm
1 mm
~1-2 billion connections
~1 mile axons (output from neuron)
~2 miles dendrites (branches of neuron)
100 μm20 nm
-7
-6
-5
-4
-3
-2
-1
0
Log time (sec)
Log size (m)
single cell recording
patch clamp
Many physical probes of living neural tissue
Where does fMRI fit?
Optical
dyes
light microscope
brain
synapse
neuron
networks
map
radioactivity
invasive (needles!)
PET scan
fMRIMEG & EEG
no radioactivity
non-invasive
Important aspects of cognition
across these time scales
0 1 2 3 4 5 6 7-1-2-3
functional Magnetic Resonance Imaging (fMRI)
Four antecedents to modern fMRI
Physics and Chemistry
magnetic properties of
atoms & molecules
Physiology & Anatomy
Control of cerebral blood flow
Neurobiology & Models
Psychology models of learningComputing
Cost, Speed, Algorithms
Modern fMRI
MRI
Exploit density differences of water in different tissues
Produce anatomical snapshots
MRI
Use blood flow sensitive measurement
Produce movie of microscopic blood flow
changes
fMRI
Despite the slowness, noisiness, and imprecision of neural computations…
The neural myth
…the brain - usually through the ‘miracle’ of parallelism -performs inexplicable feats of perception and action.
Alan Turing
121
0 01 0$
*0
Turing Machine
Hardware (physical interactions)
Patterns of information processing
Computation Computational Theory of Mind (CTOM)
Turing – mind is not equivalent to our brain, but is equivalent to information processing supported by our brain.
‘Mind-like’ stuff from ‘Stuff-like’ stuff
Blurs distinction between device and algorithm
Meaning is missing
Computational Theory of Mind (CTOM)
Symbol list 1 Symbol list 2
When does meaning get added back?
What is the value of available choices?
Two questions face every mobile creature
What does each choice cost?
Recovering meaning
Organisms must assign meaning to the processes that comprise them.
They run on batteries and the future is uncertain.
Recharge or die
run tumble
A principle for value-based computation
Organisms distribute their finite resources to the computations that sustain them.
Their willingness to do so should scale with the expected long-term returns from the computation.
How is one computation assigned more or less meaning than another?
Where’s the evidence for efficientcomputation in the brain?
32 nodes8 dedicated chess processors per node (total 256)
1.4 tons (lots of AC)200,000,000 positions per second
DEEP BLUE
Kasparov stayed warm to the touch
DDeep Blue was wildly inefficient
Kasparov’s brain is freakishly efficient
~100 watts brain uses ~20 watts
What about evidence fromdetailed neural properties?
Drain batteries slowly
Have goals
Save Space
Save bandwidth
neurons compute slowly
Set and pursue goals
Small parts, sparse connectivity
Stay off the lines, noisy codes
Principles of efficient computation
MYTH is now debunked
What happened?
Why are modern computers so wasteful?
Why is meaning missing from modern model of computation?
Bletchley Park Mansion
Enigma
World War II happened
Computer programs don’t care (enough)
Flexible goals should guide their behavior (multi-level)
Value-based Computation
client client
Hyperscan server and database
hyperscan fMRI (h-fMRI)
Houston (BCM)Houston (BCM)California (Caltech)
New Jersey (Princeton)Georgia (Emory)
Germany (Uni-Ulm)Hong Kong (HKUST)
Houston (BCM)
information: www.hnl.bcm.tmc.edu/hyperscan
software download: www.hnl.bcm.tmc.edu/nemo
Sustain goal
Select goal
Pursue goal
Goal-directed choice
Actual experience Counterfactual experience‘what could have been’
guidancesignals
Sharks don’t go on hunger strikes
But humans do
Ideas gain behavioral potency of primary rewards like food and water
How?
Goal pursuit requires guidance signals dopamine
Computationalmodels
fMRIexperiments
Valuation
Dopamine neurons
Midbrain dopamine neurons
burst
pause
R timenaive
R timeAfter learning
timeAfter learning(catch trial)
Pause, burst, and ‘no change’ responses representreward prediction errors
Red Light becomes a proxy for the value of the juice
Midbrain dopamine neurons
Constant emission of reward prediction errors
ERROR SIGNAL = current reward + γ next prediction - current prediction
Neural evidence suggests that system can be re-deployed for abstractly defined rewards ideas
“I want to solve Fermat’s last theorem”
Induce reward prediction error in a passive conditioning task measure brain response
6sJ train
6s 4sJ test
y = 3
McClure et al., Neuron, 2003
Juice not expected but delivered
Juice expected but not delivered
0.4
0.8
0.0
-0.4
% s
igna
l cha
nge
2 6 10 12 sec
+-
Hemodynamic response in ventral putamen
6s 4sJ
“BETTER THAN EXPECTED”
“WORSE THAN EXPECTED”
What about something more abstract like trust?
TRUST
I love you
Let’s do lunch
BA
A BAB
Trust
Most basic form is about model-building
Requires risk – no trust needed without risk
Cooperation
How can we study trust in a social exchange?
Use simple economic game between two players
Synchronously image both interacting brains
X 3
Investor Trustee
$20
A dynamic version of the Trust game (10 rounds)
Simplifying and quantifying Trust
Trust is the amount of money a sender sends to a receiver without external enforcement.
Agent 2Agent 1I
R
What is Fair?
i
i
i
Split the profit
Split ‘loaf’Split the total
10 + i each
i
Investor Trustee
20-i
collective ownership? common goods?
client client
Hyperscan server and database
Hyperscanning: Scan all brains during the interaction
Houston (BCM)Houston (BCM)California (Caltech)
New Jersey (Princeton)Georgia (Emory)
Germany (Uni-Ulm)Hong Kong (HKUST)
Houston (BCM)
information: www.hnl.bcm.tmc.edu/hyperscan
software download: www.hnl.bcm.tmc.edu/nemo
investperiod
8 s 8 s10 s 10 s 10 s8 s8 s4 s4 s
cue to invest
repaymentrevealed toboth brains
Gave
13
Kept
29
investmentrevealed toboth brains
Kept Gave
6 14
repayperiod
cue to repay
delayperiod
delayperiod
delayperiod
delayperiod
investment phase repayment phase
Totals
2919
totalsrevealed toboth brains
inter-rounddelay period
investperiod
8 s 8 s10 s 10 s 10 s8 s8 s4 s4 s
cue to invest
repaymentrevealed toboth brains
Gave
13
Kept
29
Gave
13
Kept
29
investmentrevealed toboth brains
Kept Gave
6 14
Kept Gave
6 14
repayperiod
cue to repay
delayperiod
delayperiod
delayperiod
delayperiod
investment phase repayment phase
Totals
2919
Totals
2919
totalsrevealed toboth brains
inter-rounddelay period
Structure of a round
decisionsubmittedby investor
decisionsubmittedby trustee
investor trustee
decisionrevealed toboth brains
decisionrevealed toboth brains
investor trustee
time
What we generally see on this hyperscanned trust exchange
What is the behavioral signal that most strongly influences changes in trust (money sent) ?
Agent 2Agent 1I
R
Reciprocity = TIT-FOR-TAT
Activity related specifically to benevolent reciprocity
Caudate nucleus – action choice, reward processing, …
Trustee brain
*
time (sec)-8 0 10
-0.2
0
0.2
Trustee ‘intention to increase trust’shifts with reputation building
Reputation develops
Trustee will increasetrust on next move
Trustee will decreasetrust on next move
submit reveal *
increases or decreasesin future trust by trustee
-0.2
0
0.2
time (sec)
Signal is nowanticipating the
outcome
Temporal shift resembles value transfer in reward learning experiments
burst
pause
R timenaive
R timeAfter learning
timeAfter learning(catch trial)
rounds
.1
.2
.3
.4
3 4 5 6 7 8 9 10
fractionof
highlyaccurateguesses
Investorevents
Trusteeevents
How do we know a reputation is forming?
8 s 10 s4 s. . .
investperiod
cue to invest
investmentrevealed toboth brains
delayperiod
guessperiod
cue to guess
Investment phase
synchronized onsubmission time of both
partners
Summary
Finite resources force distribution decisions on life formsBiological computations always had to assign meaning
Efficient computational systems must have goals
Goal pursuing machinery identified and modeled
Ideas gain behavioral potency of food and sex
Re-deployment of reward harvesting machinery