|Robert L. Blum, MD, PhD|
Singularity Summit 2010
Aug 14,15: San Francisco
Notes and Transcripts of Lectures
Saturday and Sunday, August 14th and 15th at the Hyatt Regency in San Francisco.
This conference is a favorite: this has been my fourth.
The leading luminary is futurist Ray Kurzweil, whose 2005 book
The Singularity is Near was the best exposition to date of the nature,
evidence for, and seeming inevitability of a technological singularity:
machines taking over as the key innovators on the planet, for better or for worse.
(This might happen when machines can autonomously design and build
their successors: a closed loop of ever-accelerating intelligence augmentation.)
I have about 48 hours before I jump off the grid completely
for two weeks and lead life as John Muir did in the Sierras
100 years ago (except for the down sleeping bag and freeze-dried food).
Like the Savage in Aldous Huxley's Brave New World, I will bask in the beauty
of a natural world, untouched by humans.But unlike the Savage,
I long to watch the future unfold, perhaps with untold delights:
an extension of human lifespan, freedom from diseases and other bodily constraints,
undreamed marvels of technology, and perhaps unparalleled cooperation and
wisdom among nations. There are, of course, perils aplenty.
Since I will soon depart, I have only had time to post
my detailed lecture notes, seen below. I have analog tapes of these talks
but haven’t done full transcriptions yet. In six months or so videos will be available.
In the meantime, perhaps these will tide you over. They appear in chrono order.
Use these notes in conjunction with the presenters’ abstracts,
which appear here on the web site of SS10. Feel free to disseminate.
JUMPS to my notes below are here:
Saturday August 14th 2010
Sunday, August 15th 2010
(I was unable to attend the talks by Srinivasan, Heber-Katz, Goel, Pepperberg, and Randi)
Sean did introductions for the speakers. These were great infotainment.
Sean reminds me of the character "Guy" in Galaxy Quest, who introduced
the starship crew at a fan conference. Sean, I recommend the "echo style" to you:
"RAY RAY Ray Ray ray ray, KURZWEIL KURZWEIL Kurzweil Kurzweil kurzweil kurzweil."
MICHAEL VASSAR (Singularity Institute) Saturday 9:05 AM
The DARWINIAN METHOD
(Michael is President of the Singularity Institute.)
This talk is on the history of science and contrasts
the earlier, armchair "Scholarly Method" of science with the
evidence-based Enlightenment Method.
Science sometimes makes fruitful use of a synthesis of these.
What is science?
You can't just make Gothic cathedrals without at least some rationality.
So, even if it wasn't science, there was considerable rationality back then.
Knossos had printing.
Indus Valley had running water
pre- 17th century Europe in comparison was a backwater,
but did have really good art.
Archimedes clearly had rational innovations in Classical times.
using the displacement of water as a test for the purity of gold.
Michael, however claims Archimedes was not doing science.
Why wasn't this science? no hypothesis testing, no organized literature and publications.
Etruscans thought that lightning was caused by clouds banging
into one another (rather than the gods); so, at least that shows some progress.
"Scholarly Scientific Method"
Read broadly, it consists of the following:
write abt diverse life experiences, and look for convergence in what people say.
That was the main method in ancient Greece and the Middle Ages.
How to produce great child scientists?
Start with kids who are unusually neotoneous
(this is maintaining child-like curiosity and openness into later years)
Then keep error rates high by exposing them to new material and
new experiences all the time;
Nice slide of cute kid with glasses.
how to create scholars?
get kids that notice subtle patterns and give them diverse experiences;
Why scholarship fails... one reason it fails is because of correlated errors.
wonderful cartoon : Intelligent Design for Dummies
"a reference for the rest of us"... by Dan Dog.
You need separate communities to notice different things and to
eliminate spurious correlations.
There is a reporting bias towards surprising claims.
This is a fundamental bias : surprises get published.
If you report a result saying "Sky is blue" no one will print it.
What changed in 17th century... radical skepticism..
in ancient Greece it simply meant considering both sides
build long chains of reasoning
from seemingly solid foundations and look for surprising conclusions.
Rousseau ... savages are better off than Europeans
(apparently Ben Franklin also shared this belief)
Steve Pinker in his TED Talk clarifies that at present
more than half of males in many primitive tribes DIE from homicide.
If this was true in larger society,
we would have had billions of 20th Century deaths instead of only millions.
Darwin and Wallace... working independently in the 19th Century
arrive at same conclusions; independent confirmation is key to science.
GREG STOCK (CEO of Signum Bioscience, biophysicist and best-selling author:
of the Book of Questions (sold 2.5M copies) and Metaman (a favorite of mine)
EVOLUTION and The POSTHUMAN FUTURE (Saturday:9:45 AM)
(a stunningly good talk: many themes reminiscent of his book Metaman)
He cites "Marooned in Real Time" by Vernor Vinge
(a singularity pioneer, math prof, and sci fi doyen)
an empty landscape devoid of humans. (one poss of the future)
Greg asks "so, where are the wonder drugs?"
Problem is that it takes years to do each clinical trial;
The FDA needs to address this problem.
He (and the rest of us) want to extend our life spans, but will he (in his sixties)
live long enough to be able to take advantage of the "next bridge"
that Kurzweil cites (drugs that implement current genomic discoveries)
Is it likely that the Singularity will lead to love and to a triumph of human values?
It seems more probably that it will lead to human extinction (in some form).
It seems quite doubtful that we will be able to reach thru the singularity in order
to preserve human values/love etc.
We are meat,flesh and blood.
He reads a v poignant quote from Wm Butler Yeats 1926
"Consume my heart away; sick with desire and fastened to a dying animal..."
His company Signum has been targeting Alzheimers;
idea is to preserve something (a functioning brain)
that is worth freezing (cryo preservation).
If u live to age 85... 50% chance of Alzheimers histologic changes
At Signum we look at a key molecule PP2A
It removes phosphates from regular proteins;
It has broad effects on master regulatory proteins.
One of their great findings is that coffee activates PP2A.
fun slide showing fifties housewife drinking coffee:
COFFEE - you can sleep when you're dead!
Has some evidence that coffee consumption decreases Alzheimers
by 50% and also adult diabetes,.
The effect is apparently NOT related to the caffeine content
(don't get the effect from just the caffeine).
Their lead compound is Sig 1012 - a coffee bean extract.
(a serotonin like molecule attached to a lipid).
In a Parkinsonian mouse model, you see improved nesting behavior
(tearing and collecting bits of paper to create a nest).
Should soon be able to move on to human trials, since coffee bean extract
is GRAS (generally regarded as safe by FDA)... coffee bean is a food.
They will test this to prevent Alzheimers.
Will we ourselves be transcended? eg by some combo of carbon + silicon.
Evolution is not static; its methods,are also constantly evolving;
he shows slides of earthworms doing it; Japanese beetles and zebras,
squids, and hover fliesin amorous embrace.
Social evolution is a kind of evolution in which ideas are competing (memes)
Essentials of evolution: reproduction, variation, differential survival.
With humans (unlike Moore's Law) there are no exponentials in improvement;
the generation time is just too damn long.
Transitions and Jumps in organization complexity: blue green algae first...
Warrawona group ... 3.5 B yrs ago.... then on to E Coli then
on to Eukaryotes 2.2 B years ago.
Then abt 1.2 B years ago- multicellularity.
Jelly Fish... then ….
Now aplanet-wide superorganism.: (Metaman… human civilization, writ large)
Great slides of internet pathways,
Earth at night,
Hong Kong.... great picture...
that substantiate the increasing size and complexity of our current superorganism.
Comparison to a termite mound with the queen termite
This is a new global union, a planetary superorganism.... it is beyond metaphor
(This was Greg's theme in Metaman.)
Metaman has tightly orchestrate internal networks; digestive and circ systems,
a cns, it is robust and redundant, it makes use of brutal internal competition;
and benefits from breakthrus in materials...
Example of a breakthru 560 M yrs ago in materials.
New material used in bio orgs... CALCIUM PHOSPHATE
used to create bones, skeletal structure;
essential for the transition to METAZOANS.
And now a new transition that combines complex bio + complex nonbio (materials)
What is the Nature of transitions..?
new materials SUPPLEMENT the old ones
new levels subsume old ones...
get localized homeostatic environments: eg bee hives, termite nests,
and human houses and cities.
As an example, when he travels from his house in Princeton, NJ, to Korea,
a roof is constantly over his head (we have built protected environments).
Observe the power of evolution over just a few thousands of years...
A Grey Wolf.... was evolved into pet dogs...
a huge diversity of dogs: chihuahuas, blood hounds et al
Humans may undergo a similar transformation.
Free Range humans may only remain in tiny little back eddies;
he says we will NOT be free range anymore
In Evolution our components are reused and refashioned.
Although a single neuron in the amygdala network may be part of the
network that mediates fear/ emotion; it itself knows nothing about fear.
May be the same as us in relation to the entirety of human civilization/ metaman.
Look at subprime crisis... we are terrible at predicting future;
the future may get v weird, v quickly.
Does it matter when the Singularity happens? Not really.
It does not matter whether it’s in 20, 50 or 1000 years.
It's an instant in the sweep of time.
(Of course, he and I would like to see it in our lifetimes.)
Human ethics and values are not unique to us.
Our species has kinship groups and extended social ties but that is not unique.
Chimps have it all... sharing, empathy.
slide: Mourning Dorothy's death... chimps watching the
dead body/ funeral of one of their respected older members.
chimp consoling a crying child (or perhaps the chimp's presence is causing the child to cry.)
It's a mistake to think we are paragons of rationality.
We are hugely driven by raw emotion.
Consider that Lebanon was at peace for 2000 yrs then war and destruction;
same as in Bosnia and Zrebreniza.
"Marx was right about socialism... he just had the wrong species" E O Wilson quote,
said abt his favorite species ... the ants.
What might values be in a post-human, cybernetic age:
competition will be extremely strong; sex will be vestigial,
copies will be cheap; death is meaningless.
Uploading (if poss) will disengage us from our bodies.
A slide of an uploading human ... he comments...
v evocative of Christian symbolism and images... Christ-like transcendence;
what would such things have in common with humans... v little...
What will we become?
Nick Bostrom.quote: there will be a gradual elimination
of all forms of beings that we care about.
Carl Shuman... quite inhuman values...
Aren't we kidding ourselves to think that we can possibly intervene
in this process and stop evolution.
Nick Bostrom: a super entity might stop the presses to avoid competition...
the singleton rule... to protect against its own demise by competitors.
the singleton controls the space.
We derive our specialness from our environment (materials, culture, knowledge)
People currently are concerned about clones:
some are exuberant about them, others have angst
But compared to the Singularity,
the whole issue is pathetically unimportant and quaint.
Consider issues of slippery slopes:, losses of identities,
abhorrent values and changes, perverted values..
We should possibly be worried about it;
He shows the Edvard Munch painting: Angst.
Preserving values as we we pass THRU the singularity...
has similar character (angst vs exuberance)
some super AI may manage and control it....
this is perhaps the wish of the Singularity Institute;.
he thinks it’s impossible.
a complex, emergent realm careening toward the future
will go its own way, regardless of what we do or wish.
Final quote: evolution will come from the nimble and the bold.
a 430 BC Thucydides quote ... the bravest will shape the future.
RAY KURZWEIL (Inventor and Leading Luminary of Singularity)
(Ray is on vacation at home in Massachusetts;
he uses video link to give his lecture;
quality of video is excellent, and he can see us as well.)
The MIND and How to Build One (Saturday: 11AM)
Ray begins by addressing one of his critics -
John Horgan, former Editor in Chief of Sci. American.
John claims that the brain is too complicated for humans to figure out and to
emphasize his point he shows a highly complex brain with millions of neurons.
Ray says he located that image and it's actually not a real brain but
a computer-generated brain (eg made by Blue Brain) LOL.
Another early critic was Prof. Tommy Poggio, who is head of machine vision at MIT.
Around 2000 or so,Tommy told Ray - you know, we're not getting any info
from brain studies that can help us design machine vision
Ray predicted that it would happen. When Ray saw Tommy in 2006
at the 50 year anniversary celebration of the Dartmouth AI Conference (Minsky/ McCarthy),
Tommy said ... you were right. Poggio had built a machine vision system based on
the HMAX algorithm which is inspired by the wiring of complex cells in visual V1
to simple cells in V1... it outperforms humans in certain visual tasks (with very brief duration
Can we design artifacts before we fully understand the principles of the brain?
Yes. Example is Bernoulli principle ,,, entire world of aviation is based on it,
but still not fully understood.
Once we understand algorithms of cns we can add them to our AI toolkit.
We seem to have a stable estimate of the compute power of the brain:
10**14 to 10**16 calcs per second. IBM will have a supercomputer that cranks out
10**16 calc/sec within a few months- and will be shipping.
Justin Rattner (CTO of Intel) says that INTEL
has 3d electrical circuits working in the lab.
possible successor to planar CMOS. Then, of course,
cloud computing is almost here ... use vast array of machines sitting idle.
Jaron Lanier; Jaron (at Cal Berkeley) says current machines
are not doing real AI and are not conscious -
but that is really a straw man. Nobody claims that current machines
are consc or doing human level AI.
See Jaron's op-ed in the New York Times.
Info Tech (IT) is exponential; so, 30 doubling means an expansion by a billion...
amazingly, no discontinuities in Moore's Law despite world wars, depressions, biz cycle.
Ray's prediction: in coming decades the issue of whether a machine can be a person,
and whether a machine can be conscious will become compelling issues
and not simply polite dinner conversation.
Singularity is Near came out in a 2005 book .
Ray's researchers update his charts every year
and they've stayed right on track since then.
Most recently updated in 2009.
(BTW (my aside) Bill Dally, former head of CSD at Stanford predicts
300X incr speed in computers by 2020
due to incr in parallelism and multiprocessors.
Bill, as CTO of Nvidia, has NVIDIA working on it now )
Ray's current focus... brain and mind.... cutting edge of Singularity.
He claims... brain uses 6 levels of indirection and thinking...
1 billion pattern recognizers. (I infer that
he's describing the operation of Mountcastle's cortical columns).
Artific. brain will not be restricted in volume to just 1400 cc.
Since 1890 American census, Moore's law is only one example (the fifth paradigm)
it has shown extremely predictable trajectories just like the ideal gas law despite the
vagueries of entrepreneurship, R and D creativity, innovation.
it looks like thermodynamics..
But the law only works in the aggregate; eg
no one could have predicted that Google would be the leader of search.
Even tho bits produced are doubling every couple of years,
our consumption also more than doubles;
Also other IT technologies are growing even faster.
Storage capability (mag storage, etc) and
DNA sequencing costs.
He shows a slide that has two videos playing (DNA sequencing etc)
alongside graphs showing Moore's Law of DNA and mag storage.
Life Extension: in ? 1800 median life expectancy was age 37.
BRAIN and MIND...
huge conceptual gap btwn objective world and subject (hard problem of david chalmers)
he agains shows image of cerebral cortex
(article by John Horgan stating that brain is justtoo complex
to understand but (ironically) John used a simulation of cortex gen. by Blue Brain.
Ray was just on a panel in Israel debating with Henry Markram (of Blue Brain)
in which Ray was the conservative... Henry says ... will pull it off by 2018.
Ray says No; cannot do it til at least 2020's (? human parity in neural emulation ??)
Even if we have brain sim running, it still MUST LEARN in order to be functional...
that is the key to the brain. It must interact with a sim. environment and body.
Consider this: the brain is designed initially by the genome, which only has
800M bytes of info and which can be compressed down to only 50M bytes
using lossless compression. eg the amino acid sequence ALU is repeated
We have partially reverse-engineered the cerebellum; it is critical to
coordinating motion as in catching a fly ball;
(my note: Prof Jennifer Raymond is working specif. on this problem at Stfd...
circuit diagram of cerebellum.)
Cerebral Cortex (CC) is only region of the brain where we can think
in hierarchies. Think of something and give it a symbol -
only CC allows us to do this. Lizards can't do it... no CC.
CC is the size of table napkin in humans...
has one billion units .... pattern recognizers... modules...
can recognize the cross bar in a capital A regardless of shape.
Other units recognize irony or humor (that are higher in the hierarchy of ideas.)
more abstract, but same organization.
So, we can recognize faces, pity, irony.
These units are organized as ordered one dimensional lists.
They are ordered and can go in only one direction; just like reciting
the alphabet - only do it forward ABC;
LISP machine companies of the seventies and eighties used same principles.
Brain also uses recursion, similar to LISP.
Google is really an AI company;
narrow AI already generates tens of billions of revenue.
AGI will generate hundreds of billions and more.
AI winter in the 80's.
turns out that the LISP enthusiasts were right...
each cortical module is like a lisp statement... incredible hierarchy..
we're devleoping a good idea of how these modules work.
how they are wired together.
(an axon cannot grow arbitrarily to link two modules that are an inch or two apart)
Spindle cells are unique to humans and great apes;
We have 88,000 of them and
more of them in the R hemisphere. Each has
100k or more connections. They are really big cells.
They fire in response to emotionally evocative material.
CONSCIOUSNESS (Consc)... hard to discuss because of conceptual gap..
penrose and hameroff ... claim we can explain it...
cellular computing in microtubules, but this is a fantastic leap of faith; consc in tubules? (NO!).
CONSC... Ray has never heard of any falsifiable test for consc by which
you could decide whether a device is consc; eg does it pass the Turing Test.
(that's a great challenge for consc. researchers)
Critics of Ray like John Searle accuse him of being a reductionist but
John favors biolog naturalism, which is also reductionist.
Important issue because our whole moral system and legal system depends on this.
Ray is writing and one chapter is called "you gotta have faith."
He equates consc and soul... people believe this as an article of faith.
he thinks consc is not scientifically penetrable....
he wants a sci falsifiable hypothesis or experiment;
how do you decide whether paramecium or animals are consc or whether it's just reflexes.
His attitude. if it quacks like a duck then it is a duck.
When our AI characters are Turing competent, then we will believe they are consc.
Tommy Poggio... visual stream is broken into 7 different movies.
(Ray uses this to reason that consc is an illusion.)
We have billions of bits streaming into our eyes.
Just as we now have machines that can move mountains;
we will have machines that extend our mental reach.
Question from Jack Sarfatti in re: consc and quantum collapse.
My Question to Ray: What in your view is the level of neuroanatomic detail required
for machines modeling the brain to be able to pull off human parity?
eg is it ion channels, detailed models of post-synaptic membranes, macromolec dynamics?
Ray's A: prob. does not require any of those; just
mid-level functional organization of circuits (algorithms).
Ray: important level is understanding what cortical modules do and how they are wired together.
50k new neurosci papers every year – lots of work moving us toward a solution.
BEN GOERTZEL (CEO of Novamente and Biomind)
AI AGAINST AGING
(Sean McCabe introduces Ben as being the "bad ass of AI.")
Ben says he's glad to know he's graduated from being a pencil-neck geek
to being a bad ass. That's progress.
Usu. he talks abt Novamente and OpenCog - his AI work.
But he's also CEO of BioMind LLC which does narrow AI applied to bioinformatics.
That is his focus today... work done in collabor. with Genescent.
How do you define Life? He is inspired by Bill Clinton,
who answered (in response to sex with Monica Q)
that depends on the meaning of the word IS.
Body can be understood as a machine;
Bio systems are v complex, multi-level, self-organizing.
We humans are poorly suited to analyze these data because we
evolved for hunting food, killing enemies, etc.
Costs about one billion dollars to develop a new drug.
Why do we age? Some issues...
Hayflick limit - only a few dozen replications then cells auto destruct in
a recursive cascade of self-destruction as dna repair goes bad.
Aubrey DeGrey's approach: as stuff goes wrong; just fix the damage.
Biologists are skeptical - they say harder than he claims.
Aubrey wants eg to move mitochondrial dna into the nucleus of the cell,
but this will have unpredictable consequences
(since cytochromes are inside the mitochondria).
Targets: Hayflick limit; dna repair; mitochondrial damage
antagonistic pleotropy: we live well as a 10 yr old, as a 20 yr old etc.
and we have adaptations to optimize us at each difft. phase.
Those strategies are not compatible.
Death rates level off when you are 80 or 90... late life....
annual death rate flattens out but unfort it's v high.
Genescent was founded by Michael Rose of UC Irvine.
They have bred drosophila that live 5X longer than normal- months rather than weeks.
All done by selective breeding. Now they are studying their genetics
to find out why they live so long to
account for this spectacular increase in longevity.
Yes; fruit fly life is longer by 5.5X but this involves
about abt 1200+ genes and multiple complex pathways;
We applied AI principles to this stuff.
That is the driving force behind his company Biomind LLC;
to apply advanced (narrow)
AI for to do postgenomic bioinformatics.
eg, they are trying to work out the genetic basis for chronic fatigue syndrome;
they can do quite accurate diagnosis of genetics of Parkinsons and Alzheimers.
(no supporting evidence presented)
Another bio application - use narrow AI text parsing (Ariadne genomics or genetics)
to extract relationships among published papers in the literature.
Biomind found a few dozen key aging- assoc genes
that seem to have more importance than the others.
Look at hubs in the regulation network.
Some of these genes are modified by
selenium, vit E, estradiol,, valproic acid, quercetin,
resveratrol, zinc, folic acid, isoflavones...
these act on proteins identified by our analyses as being important for longevity.
Seems to suggest that we are barking up the right tree.
We now have the have the snps for the long lived flies;
we are sequencing the Methuselah flies... these are rich, complex regulat networks.
We want to combine number crunching with human conceptual abilities;
OpenCog.org is his AGI work... two of Ben's books are the hidden pattern...
building better minds.... will come out next yr.
STEVE MANN (CS Prof at Univ of Toronto and the original human cyborg)
Extending Ourselves With Technology
(The mechanics of this presentation are totally cool and wow the audience.
Steve was wearing a video camera mounted on his glasses
that is connected to the internet so that 30,000 or so people see what he sees.
He has done this for decades - see photo.
One of the massive screens that we audience members look at
shows a direct feed from his video camera. So, to present a diagram all he does
is pick up a pad of paper and draw the diagram - we see it in real time!)
(I missed first few minutes)
What's important is the big picture over centuries- how will things unfold.
Consider Danny Hillis and the Long Now clock;
Steve loved this when he first heard about it -
a cuckoo comes out once every 1000 yrs. and chirps.
Surveillance is a clear and present danger. Paradoxically, privacy seems to be one
of his major professional concerns (later questioners ask him what happens to his
video camera when he goes into the bathroom.)
We see potential invasions of privacy all around us.
He shows an infrared sensor on a sink that images
your hands to make sure they get washed
and an infrared toilet (not clear what it does).
He pioneered SOUS veillance... watch from below.
A camera on top of a tower watching from above would be SURveillance
(as in the French word). SOUS veillance (his coinage) is human-centric
... watching the watchers (eg watching officials, the police etc)
While he was an MIT grad student, Nick Negroponte and
Marvin Minsky were on his committee.
(Steve Mann was building wearable computers in high school ...
He shows us a video recording by Nick Negroponte.)
Some unnamed MIT prof strongly objected to what
Steve does as being an invasion of privacy.
Tried to get Steve thrown out of MIT.. But Mitch Kapor was consulted.
Mitch said... wait this is of great value.
Steve describes his EyeTap device... both a camera and a display device....
it has 45 degr beam splitter... mediated reality...
laser light reconstructs what u wud be seeing.
He works with Martine Rothblatt... was at a future persons confer.
(It’s great when Steve looks into the audience...
we see ourselves thru his eyes.
He comments on the increase in spatial frequency
of the rows of people as you look toward the rear
of this long narrow ballroom at the Hyatt.)
He worked on Interaxon a bot-controlled display
used in the Ontario Olympics.
Gordon Bell of MSFT makes a similar camera that
Gordon uses on his chest... looks like camera employed by HAL.
HDR Imaging... high dynamic range imaging....
build super hi res construct of reality.... changes gain depending on level of light..
The Undigital Singularity. He wrote an essay called Being Undigital IEEE IST 1995
(this was a counterpoint to Nick N's being digital...
since >1990 or so his video feed has had open access to internet ...
everything he sees 30k people on the internet see and they can also "write to his retina"
He calls this GLOGGING ... blogging from sensors.
flUId is flexible limitless user interface device ... Ryan Janzen... undigital interface.
Ryan is a composer of orchestral music and a cs guy... works with Steve..
he wants to preserve that which is integral to human experience;
Another very cool demo -
The HYDRALOPHONE this is big blue plastic fountain with water spewing out the holes.
Ryan gives this part of the talk..
Steve's Dean at Univ of Toronto loved this instrument
and canceled all his appts to play with it.
patented technology... built into a hot tub.. steve wozniak loved it.
he shows a roof top hydralophone... located at a public landmark architecture site
Ryan plays the hydralophone: a low, ethereal sound – mermerizing.
Here are some of Steve's websites .
eyetap.org/fluid fluid media
He shows the ladder of life... SOUS veillance :
Direct cameras AT high level officials (like CSPN does)...
we need undersight..
keep eye on the officials; congressional oversight; police, army etc.
Q do u turn off your video camera in the bathroom?
Q how abt its physiologic effects on you (as a multiyear user)
Martine Rothblatt... separationof gender specific spaces...
as we age... we are gradually replaced with computer prostheses..
may need help using toilet ... husband wife team use it.
A physiol effects... what effect does it have on me?
clothing is parallel; eg naked spa nudist camp....
San Francisco is cold and miserable all yr around;
Q What do u think of theramins?
A: He's built several theramins. The generic problem is
that theramins have no tactile feedback.
BRIAN LITT BRAIN-COMPUTER Interfacing BCI ... present and future
Assoc Prof of Neurology and Assoc Prof of Bioengr Univ of Pennsylvania
teaches an entire course on BCI
a second site of BCI research is the BrainGate group John Donahue at Brown.
BCI figured prominently in the movie INCEPTION..
Reporter asked me if we cud do AVATAR right now... LOL... short answer is NO.
BCI BMI Neuroprosthetics.
BCI includes motor prosthetics but we need to broaden definition
Invasive stimulation in Parkinsons and in Tourettes is also BCI.
How to Classify instances of BCI:
open or closed loop one way vs two way.
Some systems are minimally invasive...
eg cognitive evoked potentials which use the P300 from scalp electrodes.
next (in terms of invasiveness) would be cochlear implant... then
Deep Brain Stimulation DBS subthalamic nucleus stimulation for Parkinsons.
finer temporal scale and spatial precision.
Applications: treatment of diseases: epilepsy (his specialty),
depression, obesity, Parkinsons.
Compensate for losses of hearing, vision, gait, artificial limbs.
Restore or repair: stroke, spinal cord trauma, peripheral nerve injury;
Augment: consciousness, memory, speed, attention, perception cog processing.
This is already controversial: Olympic committee had to rule on
whether a below the knee amputee with prostheses could compete in regular Olympics.
A was NO... unfair to normals!!!
Levels of Organization: does stimulation or monitoring occur at level of individual neurons or
with cortical columns or with deep nuclei.
Factoid: Arthur Rubinstein could play the Minute Waltz faster than
nerve conduction velocity from motor cortex to his fingers.
Conclusion: execution was done subcortically by over-learned reflex-like paths
that had been well-trained.
Same thing with batter going up against a 100 mph fastball.
Time is too short for the swing.
Batter must begin swing during the pitcher's windup.
Dr. Litt's field is epilepsy control and monitoring.
Reseaerch involves understanding functional. networks in epilepsy (and disrupting them).
Parkinson's disease: stimulate an inhibitory nucleus to upregulate a damaged area in Pksons.
(as an aside he mentions he has 3 teen age boys... it's easier to have pets.)
performance of these systems = product of quality of info, bandwidth, and accessibility.
he shows slides from Nat Geo 2010 article see pic of child with cochlear implants...
Hearing circuits... can interact anywhere in the system
if whole system is intact...
why would you bypass the system to go directly to cerebral cortex?
Don't ! Instead use periph processing... as in cochlear implant.
Feature extraction in cochlear implants.
people try to estimate the formants of speech f1 f2 f3...
Pronounce the word CHOICE : it comes out CH OI ssss;
decompose the word using a bank of filters to extract n channels.
If u have only a few channels in one ear, the speech sounds like Darth Vader.
hard to integrate with speech in opposite ear.
Alan Alda did a Sci American Frontiers program called Growing Up Different in
April 6 2005 PBS Sci Am Frontiers...
Demo'd difference btwn cochlear implants with
1 channel vs 4 channel vs 16 channel vs 22 channel... huge diff as u incr # of channels.
Goes from incomprehensible noise to quite easily understandable.
can even interpret speech with your fingers resting on the speakers lips and throat.
Q why do they cut off at 22 channels
VISUAL Prostheses: many different designs
one design is to wear camera on eye glasses and t
hen use output to stimulate retinal ganglion cells
Kareem Zaghoul (Neurosurg and brilliant bioengr)
will take over functiona neurosurg at NIH .
Kareem Zaghloul made an artificial eye from
neuromorphic chips, working with Kwabena Boahen at Stanford.
Rehab instit of Chicago... a woman with a silicon retina
that reproduces signals in the optic nerve... Kareem Zaghloul..
John Pezarls and Clay Reid (another visual prosthes group)
stimulate lateral geniculate to see patterns.
EFFERENT 2 way BCIs (MOTOR)
Rehab Insitute of Chicago:
woman featured in Nat Geo Article (which is online)
input from brachial plexus output to robotic arm:
she can prepare food inher kitchen... see video... great video for my website.
Must train signal classifiers .... to decide whether patient wants to move arm R or Left.
Georgopolous: built classifier to interpret signals from motor strip.
Andy Schwartz Lab.....
cortical control of a prosthetic arm for self-feeding (see video of chimp eating grapes
using robotic arm)
Meel Velliste et al...
John Donohue group people controlling cursors on screens....
Nicolelis monkey to robot across country.
PLOS One 2009 speech synthesis from cortical input...
Development of algorithms: what features should we extract
biocompatibility of tissue/electrode interface
Brian has a huge lab with dozens of grads and post-docs.
Gordon Baltuch U Penn colleague: puts cortical electrode on surface of CC...
lots of disease targets: schizo depres Pkson tourettes, epilepsy;
Steve Wong is using a clustering algo to tell when u r in the right spot (rather than listening for spikes)
Epilepsy 60 M people 1/3 are resistant to meds...
He works on State of the Art Electrodes
12 million neurons go to 1 electrode; 5mm to 10 mm resolution of electrodes...
Brian Litt and Gordon Baltuch
microseizures r common in peole with epilepsy... small spatial extent;
sz prob come from a cloud of cortical columns...
so, sz may be a disorder of neuroplasticity
What we need is dense (fine spatial) sampling over large areas;
flexible, irreg surface
transistors on a flexible substrate... 20 microns ... 288 channel...
a separate branch of his group is doing CV pacing and monitoring.
720 channel, multiplex down to ... small # of trodes...
high density neural sensor... 360 trodes...
put difft sensors on optical, chemical etc
.Bruce McNaugton David Euston Science 318...
rats replay places they have been;
fastfwd playback of recent memory sequence in prefrontal cortex during sleep.
Human substantia nigra neurons encode unexpected financial rewards
done by Kareem Zaghloul and Justin Blanco..... 13 Mar 2009 in Science.
what is the future of BCI.
Izhak Fried.... idea storage; make transfer of knowledge, feelings, behavior
replay salient events during sleep
direct brain recording fuels cog sci in TICS 2010
local fields vs single units vs eeg
cellular networks underlying human spatial naviagtion.... israeli group
He concludes with portrait of BCI in 2200.
DEMIS HASSABIS (Univ College London neuroscientist; also video game entrepreneur)
Saturday 4:15 PM
Machine Learning is Rapidly Discovering How the Brain Works
(See his PNAS paper on Imagination and the hippocampus... impt research.)
Bio vs nonbio AGI
1) Nonbio approach to AGI
Symbolic AI is the traditional way
(formal logic, logic networks, lambda calc, expt systems)
FLAWS : brittle, time consuming to train, poor generalization.
difficult to acquire new symbols
and symbol grounding problem eg CYC Lenat
most ambitious project was CYC designed by Doug Lenat.
encapsulate all the world's knowledge in a database.
Doug's group has done this for 25 yrs; and total of 600 man yrs
But, when you add 1 new rule , its takes weeks to resolve with existing kb.
2) BIO systems: use brain as blueprint
Consider the searchspace of possible AGI solutions:.
regime 1: that searchspace might me small dense .
could be a small space with a just few stars in it where human brain is one star
relatively close together in design.
regime 2: the space of all poss AGIs could be a large and sparse space...
with solns relatively far apart and brain is only one...
he thinks that regime 2 (this one) is the truth: a large and sparse space.
Evolution has only produced human intel only one time...
evidence that points to regime 2;
ie, it is NOT easy.
Bio approaches to AGI...
consider a spectrum from abstract theory on the Left
to bio implementation on the Right.
At L end of spectrum there are cognitive sci theories like SOAR (Laird/ Newell)
ACT-R (Andersen) and OpenCog Ben Goertzel.
Must shoehorn new modules into each of these diagrams. It is unprincipled and
you have quite a proliferation of these architectures..
how do u prove that yours is best or correct.
At the R end of the spectrum is whole brain emulation WBE
Eg Henry Markram Blue Brain
or Dharmendra Modha with IBM's Synapse project....
big Q is the level of detail relevant to cognitive performance in the emulation...
really 50+ yrs away from doing 1 to 1 mapping of brain.
DEMIS favors a system neurosci appraoch in the middle of the spectrum
figure out the brains algorithms.
David Marr had 3 levels of analysis... the father of computational neurosci.
According to Marr: Complex bio systems have 3 levels.
Top level is computational.... what are the goals of the system
middle level is algorithmic: how does system accomplish those goals:
bottom level is implementation... what is the physical realization
so WBE guys ... want to do the lowest level and
the cogsci guys are at the top/ comput level
he wants to be in the middle and figure out the algos..
there have been rapid advances in neurosci...
new experimental techniques...
2 photon micro, tms, optogenetics, multicell record, imaging
how to find the nuggets of neurosci that are relevant to AGI
from the 50k+ neurosci papers in 2008
how to keep abreast of sea of info..
need full immersions and experience in both areas
it and takes > 5 yrs of immersion.
neurosci provides direction for agi and validation
classic example is computer vision
hubel and wiesel 1959 work on cat cortex... 2 types of cells
simple cells tuned to preferred input stim
eg straight lines oriented at 45 degrees toward L in primary viz cortex.
also, Complex cells so eg C1 has inputs from all the S1 cells
that r oriented 45 degr to the L
so that they fire with a stimulus anywhere in the field.
HMAX model of Poggio is bio-inspired model... HMAX
he did it just right ... extracted algorithm from
hubel and wiesel work.
A second example: PLACE CELLS...
rat is wandering around his cage... trodes stuck into its HC (hippocampus)
and researcher can see the route of the rat because a particular cell
fires say when the rat is in the northeast corner of the cage.
grid cells... Moser et al Nature 2005....
cell firing in a regular hexagonal grid that tesselates personal space.
So, neurosci provides direction for AGI models.
how abt validation...
does an algo constitue a valid component of an AGI system
eg reinforcement learning.
In RL, the only teaching signal is the reward gained from the environment..
RL = reinforcement learning.
TD learning is a method for solving part of the RL problem ....
works by minimizing the error in expected reward.
TD = temporal diffierence learning... sutton.
TD learning implemented by the brain.... schultz dayan montague in science 1997;
1) no prediction and reward occurs anyway
2) reward predicted and reward occurs
The hybrid present in dopamine neurons.
Identical RL algos
implementing RL algos as part of AGI is good... it happens in brain.
These neuroxmitrs are connected with...
dopamine... reward; serotonin mood; ach variability; norepi unexp variability.
hybrid approach: (is DEMIS's approach)
combine the best of agi and neurosci
Example systems for emulation:
model based vs model free systems
theory of mind
top down attn
control conflict resol by the ACC (anterior cingulate cortex) etc.
CONCEPTS are the key... how to acquire knowledge in the brain;
3 dfft levels
symbolic vs conceptual (of a city eg)
perceptual what it looks like vs
symbolic give it aname
symbolic... logic networks , symbolic systems;
perceptual : dbn hmax,,, htm hawkins hmax poggio
conceptual how to do it??
HC-neocort consolidation HC = hypocampus.
HC sits at the apex of the sensory cortex
high level cortex in assn areas.
HC stores the memories of recent memories or episodes and
replays those memories during sleep at sped-up rate
gives high level neocortex samples to learn from
memories selected stochastically for replay
rewarded, emotional or salient memories are replayed more;
circumvents the statistics of the external envirnoment
leads to abstraction.
Feynman: what I cannot build I cannot understand.
TERRY SEJNOWSKI (UCSD Prof and Salk Institute Researcher)
Saturday 5:20 PM
Reverse Engineering Brains Within Reach
He mentions to students that he used to compute
with two sticks of wood....a slide rule!
He talks about reverse-engineering the brain.
Humans have a long history.
First there were marine invertebrates, then marine vertebrates,
then lands plants with reptiles, then
mammals, primates then man.
Bacteria have the same phosphorylation channels as us;
duplication and specialization have worked for millions of years.
slide: great picture of Francis Crick in Indian high-back chair.
Consider that in 1 910 vitalism was prevalent..
With the advent of of xray crystalography, it was for the first time possible
to work out 3d struc of DNA.
Craig Venter took nucleus out of cell and basically created transvestite version of life.
Look at space/ time diagram of neuroscience techniques:
Space on the Y axis starting with Synapse at .0001 mm then
dendrites then neuron then layer then nucleus etc.
Time is on the X axis with .0001 msec and on up to seconds.
Ray K said brain uses 10**16 calcs/sec... he agrees.
In this diagram, a patch clamp is at lower L (ion channels)
and fMRI at upper R (cm resolution and seconds)
also PET scans.
Levels of of investigation: if u model the brain, then what level:
whole cns 1m
systems 10 cm
maps 1 cm
100 micron neuron
synapse 1 micron
molecules 1 angstrom.
When Terry was in college, total # of neurons was 10 billion;
then people found that cerebellum had perhaps 50 billion granule cells.
now people estimate total is 100 billion neurons
b/cause of huge # of granule cells in cerebellum
synapse ... abt 10 **15 synapses... THEY are doing the heavy lifting...
Ray K said 10**16 bits per second bandwidth
bees are the champion learners among the invertebrates....
assoc color, odor, position of flower;
Karl von Frisch... ethologist... found that out;
stimulate with sugar water and bee put outs proboscus.
one partic neuron
VUMmx1 neuron responds directly to the sugar soln...
uses octopamine as neuroxmtr.
by stimualting the VUMmx1 they could condition the bee to put out its proboscus...
temporal difference learning TD learning
montague, dayan sejnowski 1994 ...
P neuron is predicting the future... should you land on the yellow flower or the blue flower.
dopamine is involved.
model could predict in detail the behavior of the bee...
bee prefers the sure thing... 100% prob of sugar
rather than occas flowers with more sugar but some with none.
octopamine in bee is like dopamine in us;
dopamine system originates in substantia nigra or VTA ventral tegmental area
SN projects to nuceus accumbens and to PFC etc.
schult dayan montague ? 1997 temporal differences
TD Gammon : Tesauro, 1994 played against world champions...
it really learned quite well.
see table 1...
TDG 2.1 used 1.5M training games and played against Robertie...
learned v subtle strategies... early blocking...
machine learning (was not taught or designed... TD Gammon learned it;
Deep Blue got all the publicity but TDG was NOT hand programmed. it learned its
strategies: it used a genl purpose architecture.
brain is a collection of special purpose algos.
cognitive dynamics: can make radar more efficient
cognitive radio; cognitive cars
see simon haykin 2005
cognitive power grid...
can make much more efficient grid (plans for future to do smart allocation)
haykin (his colleagues)
his colleagues ? at ucsd.
DENNIS BRAY Prof. at Cambridge Saturday: 5:50 PM
What Cells Can Do that Robots Can't
He is interested in wetware: the computational capacity of every cell.
his loyalty is to the carbon based systems
he studies physiology, development and neurosci at Univ of Cambridge.
Real bio is much more complex than our models
he claims we really do NOT understand how humans do cognitive stuff.
cells are full of complic molecules
E Coli is abt 2 microns long with lipid membrane.
has flagella; receptors sense the envir.
swimming, sensing, metabolism
E Coli contains abt 10**9 molecs made from 4300 genes.
Human cells have abt 10**12 protein molecs and made from 25k genes.
less than 2% of human genes code proteins.
hexokinase binds atp and glucose.. spits out glucose phosphate.
enzymes work like logical devices:
feedback inhibition is universal in bio systems;
membraine voltage + glutamate (and modulation by glycine)
triggers NMDA receptor to open...
this is the underlying mech of learning... the basis for Hebbian synapses...
phosphorylation is a mech of switching an enzyme on and off...
Calmodulin converted to CAM2 kinase...
mech of LTP (long term potentiation)
glycogen synthase is regulated by phosphorylation at 12 sites...
they can oscillate, have memory...
proteins work togerther in small circuits...
eg 3 proteins can generate diurnal rhythms...
Kai A B and C... forms circuit
see RUST et al SCIENCE 318 2007.
small networks in cells process incoming signals
protein nets in cells are highly complex
see Oda et al in Mol Sys Bio 2005...
even more so when we include DNA... eg
sea urchin development 30 hrs... v complex regul paths
Levine and Davidson 2005
living cells contain v complex computing elements....
muscle is a highly organized molec machine;
Consider the insect flight muscle:
actin myosin interaction: slide by cleavage of atp..
muscles are turning over... they are not static...
ELIEZER YUDKOWSKY (Singularity Fellow)
Simplified Humanism and Positive Futurism : 9 AM Sunday
Consider these examples of cognitive dissonance:
Death gives meaning to life... NOT!
Death does NOT give meaning to life...
life gives meaning to life.
He shows Leon Kass (Bush era ethicist) quote ..
.death is important to us humans.
death is a blessing in disguise....
if people had never heard of old age and death and u proposed the following:
let's say u can wrinkle up, feel like crap and then die....
then it will really make life special.
Would anyone actually go for it? No Way.
This is just an example of people pretending to be wise.
He quotes Frank Sulloway.
Ninety-nine per cent of what Darwinian theory says about human behavior is so obviously true
that we don’t give Darwin credit for it. Ironically, psychoanalysis has it over Darwinism
precisely because its predictions are so outlandish and its explanations are so counterintuitive
that we think, Is that really true? How radical! Freud’s ideas are so intriguing
that people are willing to pay for them, while one of the great disadvantages of Darwinism
is that we feel we know it already, because, in a sense, we do.
Is death good? if u say NO then that's advice that no one would be willing to pay for.
Freuds ideas are so intriguing that people are willing to pay for it.
typical (nonsense) : curing disease is good, unless genes are involved
vs simplified version: curing disease is good.
same with life and health up to age 80 is good vs
life and health r always better at any age.
Yoda: Shape matters not....
Personhood theory... is yoda a person?... kids may say no,
Cause Yoda looks different until...
as Elie suggests.... substitute this question... would you EAT YODA??
Good stuff: Life Health Happiness Knowledge (we want more of this)
BAD stuff: death, sickness, pain ignorance (v awful, let's fix this)
The Way of Chuang Tzu translated by Thomas Merton...
In the age when life on Earth was full etc...
lovely sentiment but NOT TRUE... primitive peoples are not peaceable ... a la Steve Pinker
primitive tribes were and are violent... half of men die of homicide.
David Brin.:. most societies have nostaligia for the past golden age...
only in our society do we have positive futurism... our best days lie ahead.
conjunction fallacy is when you assign p(A AND B) > p(A)
adding more detail makes an event LESS probable
even tho it becomes psych more plausible.
that is a branch of the entertainment industry. (that brand of futurism)
is this person telling a story abt the future or predicting the future.
values and goals are not beliefs and predictions.
technophile: tech is our friend, future is good,
lets speed up progress and get there faster.
technophobe... is just the opposite..
technovolatile.. future might be extrem good or extrem bad but isn't likely to wind up anywhere in between.
technovolatile ... that is the position of all of us at Sing Instit ... Peter Thiel, M. Vassar, me.. all
Sir Francis Bacon in The New Atlantis... lovely quote
The end of our foundation is the knowledge of causes and the secret motion of things;
and the enlarging of the bounds of human empire, to effecting of all things possible.
contrast that with Bill McKibben, Enough (one of Bill's books)
"You can't stop progress. but that's not true. We could choose to mature"... NO NO.
they don't have a vision of how to get there from here (folks like Bill McK)
The New New Atlantis
positive futurism... climbing upward with or without a natural slope ...
positive goals (health, longevity)
3 worlds collide (his story online)
Q abt Brave New World (artificially induced, Utopian, designed happiness)
A... logical fallacy from fictional evidence...
RAMEZ NAAM (MoreThanHuman.Org)
The Digital Biome Sunday 9:40AM
We live in the Anthropocene Era... human-induced change.
the planet is warming...
global mean temp is undoubtedly anthropogenic
this is due to CO2 and methane...
will be 3 degr C warming over this century..
if we engage in biz as usu, temp will go up by 6 degr of warming...
will get 3-4 meters of sea rise... if 6 degr C of warming...
(some grumbling from a skeptic in the audience)
This won't lead to the apes taking over (Statue of Liberty on beach:
slide from Planet of the Apes)
Runaway warming is the real problem -
Increased temperature thaws permafrost ... releases methane and CO2...
Get feedback loops..methane is much more powerful greenhouse gas...
and incr temp could lead to huge release.
there is 750 petagrams of carbon in the atmosphere...
plants soils have 2k PG (petagrams)
deep oceans have 40k petagrams
coal, oil, gas have 10k...
we could get dramatic warming...
Younger Dryas .. v rapid warming 10k yrs ago... 15 degr C warming in a few decades...
at times of scarcity, people fight over scarce resources... Wm Calvin (neuro guy.)
CO2 gets into ocean and causes ocean acidification...
corals can't create new shells, calcareous plankton...
2 degr C of warming causes corals to start to die off and
at > 500 ppm of CO2 they are all gone...
he shows data on fish tons per unit of effort...
data from Japanese fishing fleet...
they most go further and they catch less.
fresh water... vital to food supply...
ARAL SEA in 1989 ... vs 2009 almost gone...
Russians redireced rivers for agric and it’s almost gone...
Ogallala Aqifer: in MidWest US: runs from S Dakota to Texas... huge depletion...
2008 spikes in food prices... yields r not going up anymore..
He shows grain yield per hectar- Green Revolution has run its course...
SPECIES LOSS: as a computer scientist, he hates data loss, loss of info:
loss of species.. missed opportunity...
species loss is also a singularity... rapid expansion. not all exponentials are good.
World's Liquid Fuels Supply r also dwindling..
world oil supply is not expanding;
SARS virus.. appeared out of nowhere and then was everywhere...
NDM-1 new superbug in UK... bacteria do lateral gene transfer.
Biology as info tech
Sequencing center.looks like a data center.
George Church... wants to do 100k genomes by 2020;
but he predicts 10 dollar genome by 2020 so prob we will do far more than that.
3 million species on earth... why not sequence all of them
plants 290k species
insects 740k species
animals abt 240k species.
dogs.... he shows a clade circle... beagles, retrievers etc..
a huge radiative adaptation ... all done over few 1000 years with selective breeding.
tools from biology can also be used to create
ENERGY from biology and biofuels
crops compete with food... refining is too expensive...
craig venter... Exxon has paid him 600M bucks to develop orgs to do it..
create orgs that do what we want by direct design... eg algae...
he shows a possible algae plant in nevada desert.... (artist’s drawing)...
DARPA wants fuels created on-site with biofuels eg in Afganhistan...
Tobacco Mosaic Virus... 10 nanometer.. make a spray on PV from them.
incr photosynth of plants to pull CO2 from atmos.
Aquadvantage salmon... grows faster.... close to approva by FDA.
cocolithiphor.. makes calcium shells 30% of all calcification in oceans...
this family does More calcific in high CO2 conditions.
It is a beautiful organism... great slides...
We sequenced SARS in 5 days
350M terawatt hours of energy from sun per year...
how to tap into it...
20% is best efficiency in lab with PV's.
desalinization... takes a lot of energy...
but not that much.
How about food supplies...
C3 production is used to create wheat rice
C4 production is used to create corn
winning the race:
how to maximize the odds of great future...
Market assumes (tragedy of commons) just dump CO2 into atmos...
no cost to you. need to price or regulate the commons
Q gerry orstrom (gave 10k to Sing Instit)...
how do we know CO2 is inducing warming change ?
A:trapping more heat in the atmos...
Q has warming preceded CO2 yes, in some cases
But actually both r occurring...
LANCE BECKER MD ER doc and Prof. at U. Penn
Modifying the Boundary Between Life and Death (Sunday: 10:40 AM)
life used to be simple...
alive vs dead with nothing in btwn...
patients were rarely successfully resuscitated.
irreversible brain damage in 4 mins.
he suggests ... why that may not be true...
he wants to expand that bndry btnw life and death...
how big.. how long is that bndry.
it used to be 4 mins... how abt if it was 10 or 20 or 40 mins...
sudden cardiac arrest... arrhythmia... 95 to 98% will not be resuscitated.
usu causes are strokes heart attacks, brain death...
ischemia and reperfusion... loss of nl oxygen or blood flow...
what is death... we know surprisingly little about it.
when is death...
he went to the lab decades ago to find out....
when do cells actually die...
he was growing heart cells in vitro.
heart cells begin to beat together even on a glass slide.
when do the cells die when deprived of O2?.
heart cells on slide.... deprive the cells of O2 for 1 hour ....
during ischemic period of 1 hour the cells were NOT dying...
ONLY DURING REPERFUSION did they start to die off.
cars... run out of fuels for 1 hr...
imagine this: after 1 hour you put gas in the empty tank of the cars
and then they explode!!!
this is exactly what happens with neurons and many other cell types:
they only die off AFTER the O2 is restored after 1 hr .
What's going on? It appears that cell death is an ACTIVE process
it's not simply cells running out of fuel and dying.
He began to study the mechanisms of cell death.
all roads lead to the mitochonidria...
The mitochondria were key to our becoming muticellular creatures.
Billions of years ago they became incorporated into larger cells that
ate them, then developed a symbiotic reln.
Mitochondria consume O2 and need it to stay happy.
Lack of O22 causes eletrons to pile up inside them.
The reperfusion injury occurs when O2
is reintroduced into the cells and combines
with the excess electrons... that's the signal for cell death.
How to prevent it? He has tried cocktails of
antioxidants, inhibition of signaling, cooling, mitoch inhibitors like HS CO CN
We really need a mitochond meter (to inform us of their state).
Look at the article in Popular Science called COLD RELIEF ....
cold heart can save your brain..
But cooling has a major problem...
it is v time dependent..
He developed a new slurry generation device
to cool someone immed. in the ER.
(Cooling usu takes abt 8 hrs to cool somebody 3 degr C)
He cites rat research by Gerry Buckberg UCLA...
rats subjected to 30 mins of global ischemia...
uncontrolled reperfusion causes injury... compared by Buckberg..
if uncontrolled get 2 deaths from brain herniation (cerebral swelling) in
rat control group.
However when reperfusion is tightly controlled (in treated group of rats)..
Buckberg got 6/6 survived; high brain 02 uptake and
3 rats had full neuro recovery...
Penn Goal... do emergency CV bypass in ER + controlled reperfusion...
he really wants a mitochondrial meter...
This man was dead...
now he's back from the dead
Also see Popular Science.... Frozen Alive article.
SHANE LEGG (University College London, former advisee of Marcus Hutter)
UNIVERSAL MEASURES OF INTELLIGENCE (Sunday: 2 PM)
Marcus Hutter (Swiss) has studied problem of intel for decades.
See Marcus's website on AI and on AIXI, Marcus's genl math theory.
Shane starts with calcs per sec in supercomputers..
He projects out to 2020: get 10**18 calcs per sec
We are currently at 10**15 calcs per sec.
(Ray K's figure was 10**16 now, by IBM)
BUT, rather than just showing Moore's Law in terms of calcs per sec on Y axis,
what we really want is MACHINE INTELLIGENCE on the Y axis.
Is machine intel going up same as Moore's Law?
Well, how do u define or measure intel?
slide: Homer Simpson ... big laugh.
Get def from Russell and Norvig... first few pages
How to approach problem:
he shows a 2 X 2 grid with
Human level vs ideal level and
internal properties vs external behavior
We don't know how to do the internal properties but
we can talk about the bottom R square in the 2 X 2 grid,
ie IDEAL, EXTERNAL BEAHAVIOR.
How do we define intelligence.
He has compiled the world's largest collection of
defs of intelligence (80 definitions).
Def by Gudwin: intel systems r expected to work, and
work well in many difft environments
Def by Simonton: a cluster of cognitive abilities that lead to
successful adaptation to a wide range of environments
the ability of a sys to act approp in an uncertain enviro,
where approp action incr prob of success: Albus
any sys that gens adaptive behavior for a wide variety of goals... Jerry Fodor.
So, to summarize: intel is a property of an agent that interacts with its enviro
to success achieve goals
across a wide range of environs.
Consider the following series 1,3,5,7, ???
maybe the rule is 2n -1
then he shows v complex eqn where it jumps to 33.
NO... Occams Razor... simplest rule... great heuristic.
So, to the above defintions, add in Occam's razor...
Here's what you get.... he shows a small, compact closed expression for intel:
SUM 2**-K(mu) * V from mu to pi
pi is agent and mu ranges over enviros
V = success (reinforcement) = pronounced Nu.
E = wide range of enviros
mu = enviro mu is ranging over a wide range of enviros, E.
Occam's razor is the 2**-K(mu) K is complexity; when complexity goes up
the weighting of that term goes down.
He has formally captured the definition of intelligence;
He shows that this equation captures the essence of many informal defs.
and that it orders simple agents correctly.
Upper limit is AIXI (Hutter's optimal AGI algo)
can prove that it's optimal for all enviros.
It is a continuous measure
intelligence spans a huge range;
Turing test is binary... can't guide research pass or fail.
This eqn is non anthropocentric.
We want want to build the machine intelligence graph (on the Y axis)...
r we doing better or not each year.
AIQ Algorithmic Intelligence Quotient
He applied the equation to a variety of theoretical systems, eg
he gets numbers for
Q(lamdba) + functional approximators
MC-AIXI Monte Carlo AIXI
program can learn poker etc.
if u cannot measure it, then it is not science ... WM Kelvin.
Joel Veness is his colleague.
Marcus Hutter is his PhD thesis advisor (Swiss Natl Instit)
Shane is at Gatsby Comput Neurosci Unit
So, how do you actually apply it to real systems:
extremely complex to apply: its a 50 page paper...v hard to apply
LEARNING is essential to intelligence: must be able to adapt to new enviros
In his testing of the eqn, he used a model system in which
the agent pushes different buttons
to get different rewards. He gets an audience Q: how do u rate humans?
A: don't brush off human:
toddlers r v smart... our algos do not approach them in intel.
cannot do all poss enviros... so they do Monte Carlo sim
of all poss enviros (random sampling).
woman working in psychometrics asks a Q: isn't your test order dependent?
And, aren't some enviros far more challenging than others? How do u rate the enviros?
JOHN TOOBY (Prof of Psychology at UCSB and
Founder and leading light of the field of evolutionary psychology!!)
CAN DISCOVERING THE DESIGN PRINCIPLES OF NATURAL INTELLIGENCE
UNLEASH BREAKTHROUGHS IN AI?
Sean introduces him: "Evol. psych has shown that sex evolved as a
defense against parasites. And here I thought that it evolved just for fun."
His colleague at UCSB is Leda Cosmides.
In regard to future of AI, he thinks that
truth may be really far from what we expect
rather than lying btwn the extremes defined by
the AI skeptics vs the AI true believers.
His goal in evolutionary psych is to reverse engr the code of human nature...
object of study of evol. psych is the
neurally implemented programs that evolved to regulate behavior...
there are two different ways to think abt it ( 2 difft Marr's levels)
1) the physical... Marr's implementation level... neural strucs etc
2) thecomputational (Marr's algorithmic level)
reverse engr the info-processsing ops and data strucs.
why emphasize evolution (the black art)?
natural selectioon acts as a filter on bio struc:
it only builds func struc that solve evolutionarily recurrent adaptive problems.
human brain consists of a collection of problem-solving strucs...
Einstein quote: it is theory that decides what we can observe
(unguided empiricism is too weak and slow);
ancestral world posed a series of problems that
had to be solved to survive and reproduce...
naviagtion, mating, foraging, negotiation in conflicts of interest etc.
the architecture of the adaptive solution (the circuit logic of the program)
reflects ...in detailed ways, the structure of the adaptive problem...
they mesh like a lock and key..
huge amt of adaptive structure... enormous diversity outcomes....
beautiful slide showing the huge diversity of animals.
Here's the EVOL PSYCH 5 step program.
1) ID an enduring adaptive problem faced by our hunter-gatherer ancestors
2) do task analysis to hypthesize abt poss soln programs
3) test hypthesis in lab
4) id the program's neural, developmental and genetic basis
5) test cross culturally
(he uses a field site in Ecuadorian Amazon)
He thought: psych MUST be using evolution...
he took a course in cog psych in 1970 and he was
flabbergasted to find no use of evolution in psychology..
Still, even in 2010 evolution is eft out of cog sci, psych, neurosci.
why would anyone leave it out...
like making a taboo of math.
He quotes Max Planck: Science progresses funeral by funeral!
(leaders of the field have to die off and be replaced)
Example: (Darwin was stumped by altruism... an apparent contradiction to evolution)
So, what are the computation algos for evolved programs for detecting
genetic relatedness, and using that to explain altruism...
selects for avoiding mating with relatives..
2) BUT want to care for and aid close genetic relatives (including offspring)
and to avoid harming them
Kin selection Hamilton's rule... help relatives...
solving these 2 difft probs: ancestral adaptive probs, required the same solutions
You need a kinship index... need that..
needs to affect the motivational system causing sexual attraction...
you can't neces. see relatedness..
what r the cues to relatedness...
domain general learning is not enuf. (need specific expertise)
evolution sifts for those cues which r v reliable generation after generation
Cue: maternal perinatal assoc MPA...
kid observes mother nursing a new baby then baby is related to it.
trying to explain incest aversion..
when hunter-gather bands fission and fuse, nuclear families stay together...
degree of kinship for sibs is related to length of residence.... goes up to age 18...
cue monitoring system.
co-residence monitoring system..
intel is achieve by throwing away irrelevant info...
if u have maternal perinatal assoc MPA but if you're a younger sib and
u lack this then use co-residence monitoring system...
kinship index down-regulates sexual interest... leads to disgust.
these mechanisms are nonconscious: computation of KI (kinship index)
explicit beleiefs abt kinship do not matter
(it is the nonconsc piece that determines their attitudes)
people encode race because it's related to cues abt coalitional alliance...
We really need strong AI to understand what humans want..
must have maps of of human mind ...
human... what is the optimal population for sustainability...
Rather than AGI (general intel), Tooby favors the term BROAD intel...
ie it's a matter of degree .
What are the hacks that we use to accomplish our goals?
intel is computational organization that solves problems.
dedicated intel uses highly target problem sets
(uses nonstandard adaptive logics)
improved intel improvises solns to probs not encountered before...
useful intel is all abt reduction because
all intelligence is too bulky to fit in our brains or our servers...
what r ways we can massively carve away the stuff we do not need.
He states: goal of problem independent solutions has hampered the development of AI.
rather than that: it's better to find pinpoint strategies that solve key problem sets..
color constancy, social xchange, parallax..
native intels r dramatically better... special purpose...
we were faced with narrow probs... predator attack, foraging, social interax.
genl methods r computationally weak, and the more genl, the weaker!!!
Ecological rationality... ecolog struc is enduring...
Goal is to develop an ever larger battery of dedicated intels
bundle together dedicated intels...
food eval, kinematic geolocation, mate eval, face processing, etc.
My public Q: Where can we find out more abt Evol Psych.
A: go to his website at UCSB.
my second Q: How do we know that the answers supplied by
the field of Evol Psych arenot simply post-hoc "just so" stories ...
how can we validate them?
A: many times, when we look for a soln,
what we find (eg with incest taboo or kinship index) is
something unconscious that noone suspected beforehand.
Results are surprising and
then need to be cross-validated across cultures.
DAVID HANSON (CEO of Hanson Robotics)
EMOTIONALLY INTELLIGENT MACHINES
(Sunday: 4:25 PM)
Hanson robotics: we bring robots to life...
example: model of Einstein...
patented skin material with improved facial physiol.
Zeno: a consumer robot product
(Davbid was an Imagineer...at Disney)
We love characters...
characters who understand us and see us.
His main message: development of intelligenct characters are a
key driving factor in AGI (and especially friendly AGI, since they
are forced to interact with humans by design)
Example: Zeno: Hero from Tomorrow.
His Robokids are analogous to trditional character media.
He shows a conversation between Amanda (a human) and Zeno (robot)
this work pushes the frontiers of machine perception,
machine learning, material sci, robotics
He shows an android portrait of Philip K Dick (the great sci fi writer)
VALIS vastus active?
face tracking 2005...
We fall in love with the characters in our favorite novels:
life as a quest... Zeno of Allea...
Computational models of humans:
we r hard-wired for social cognition and face to face interaction.
We want to embody social cognition...
We want our robots to seem like protagonists.
people innately favor agency or characters;
people desire smart, lovable chars. (eg Bugs Bunny)
push friendlier AI into mktplace.
modeling people helps us to understand people better
robotics, software, neurosci r loops
we will discover the answer to the question: what is human?
defining character... means agency... beleiving that there is a being there.
A character has a story with accomplishments and goals.
A moral character: able to get along with people and to find solns.
Our best charcters achieve their goals by honor and by perseverance.
even newborns respond to faces... facial expr r universal ...
Paul Ekman 1972, and 19th cent.Darwin.
The age of character machines is coming upon us:
Sims,, world of warcraft, Club Penguin, Zwinky, Webkinz...
Consider Brazeal/Winston Kismet.
CMU Doc Beardsley
Repliee: Ishiguro Lab
Several innovations from Hanson Labs:
in robots, poor materials are too stiff and the robots are power hungry.
omposs to make mobile, walking robots.
He shows videos of Zeno.
At Hanson, we invented FRUBBER:
lipid bilayer elastomer... lifelike and low power to deform (on the face of a robot)
we want emotional connection.
by deploying these robots as products we get a lot of customer feedback
This work combines many fields: neurosci, materials, art, robotics, psych...
frubber ... better expressiveness , using 23X less power...
extreme pore packing density...
electric field sensing
Character (AI) engine: he shows an interesting slide
(get this from his work at Hanson Robotics)
Speech reco uses Dragon Natl Speaking
Facial recog and expression characterization:
Kino Coursey and Daxtron Labs...
Apollo Mind Initiative... want > human intel by end of decade...
facial express mimcry..
See David's TED talk... audio localization...
in development: sentience engine...
ART is in the loop also.
bold new artforms... reifys AI...
robots can be used for autism research...
want mkt pressure for friendly AI...
robotic character development is a clear path to friendly AI,
JOSE LUIS CORDEIRO ( Director of Millennium Project in Venezuela)
The FUTURE OF ENERGY (Sunday: 5 PM)
He graduated fom MIT.
The largest industry on earth is the energy industy: 8 Trillion dollars per year.
See the report by the millennium-project (online)
endorsed by Sec Genl Ban-Ki Moon.
Cute slide of Ostrich using a jet pack to hover.
He was at Singapore University.
worked on Home Energy Generation:
Larry Page (Google fndr) says this is what I would want to do if I was a student.
Ray K and Larry Page... how to make solar energy economic for humanity...
solar flux is 86k terawatts per year.
hydro power is tiny
peak oil will occur in the 21st century
Sheik Yamani, 2000 Saudi Arabia...
peak whale oil in the 19th century.
sun is approaching a solar maximum;
Club of Rome (Jay Forester) that report had incredible impact but was terribly wrong.
Herman Kahn...claims no limits to growth...
there is plenty of oil he says...
Venezuela... oil for 4 centuries...
The big problem is not that oil is not abundant but that it is dirty: buring it gens CO2.
Different waves of energy use:
wood in 19th cent then coal, then oil, then natl gas, then other...
MIT energy initiative...
Wireless electricity xmission... MIT
Arthur Clarke: any advanced tech is indistinguishable from magic....
Bob Metcalfe (Fndr of 3Com) The ENERNET...
wants entrepreneurs and prosumers to control renewable, distributed systems
Exxon gave 600M bucks to to Craig Venter..
Einstein showed that mass contains the equivalent of 90M megajoules per kgm.
only 15 people in our audience had heard of Kardashev... Russian physicist.
in 1 hr we get as much solar energy as we use in 1 yr....
he wants to reach Kardashev level 1...
JAXA (Japan Space Agency) want to power Tokyo by solar power by 2030 from space...
could use moon-based panels.
funny slide: we SHELL not ARCO...
sunny slide the American view of the world:
(summarizes S America as simply a supplier of cocaine and coffee...)
we may ahve free energy in 30 yrs from solar.
he is moving to Korea to Seoul.
He shows an incredible picture of the new designed city that he is moving to in Korea.
The Yin and the Yang: each has some of the other inside.
New Songdo City, S Korea....
Chinese saying: in crisis there is opportunity.
see his site cordeiro.org
His slides r available there.