Rahul Alex Panicker showing an example of computer-art
generated purely from random noise, using a neural network trained on places by the
MIT Computer Science and AI Laboratory
Bobby
Paul George who founded the Kochi Reading Group was extremely keen that we meet with Dr Panicker, who is
related to him and whom he classed as a ‘genius.’ Therefore
forsaking the India vs. Pakistan T-20 Cricket match which was to take place on the same evening,
we attended the AI talk.
Dr
Panicker’s was a rambling general talk, accessible to the
layperson; because some of the points he made were provocatively
optimistic predictions of AI, the audience interposed questions and
participated in the discussion.
Dr
Panicker has nurtured and co-founded a company with a manufacturing
base in India with ideas generated with colleagues of his at Stanford. They make swaddles for infants. Now he is stepping into new
ventures.
Dr
Panicker hails from Mavelikkara in Kerala. After the talk Bobby Paul
George, his host for the evening, arranged dinner at his home
preceded by drinks at the Yacht Club Bar. The evening was
stimulating.
The audience at the talk
You can read more by clicking below.
RP
gave a general talk for the layperson on recent advances in AI. He has a
BTech from IIT-M, followed by an M.S. & PhD from Stanford in Elec
Engg. He is the President and co-founder of Embrace Innovations which
manufactures “easy-to-use, portable infant warmers that do not need
continuous power supply. Designed for hospitals and ambulances, they
are used in NICUs and wards, and for transport, when skin-to-skin
care (a part of Kangaroo Mother Care / KMC) is not possible.” See
He
is now involved with another company dealing in AI and Robotics. Surprisingly, the talk he gave was the same as one he gave at IIT-K two weeks ago.
One imagines IITians would have expected much more
technical detail in a specific area, not a general talk on the
prospects for AI.
RP interacts with the people before his talk
He
said there has been a revolution in ‘Deep Learning.’
MYCIN
was an early expert system (1970s developed at Stanford) that used
artificial intelligence to identify bacteria causing severe
infections, such as bacteremia and meningitis, and to recommend
antibiotics, with the dosage adjusted for patient's body weight —
the name derived from the antibiotics themselves, as many antibiotics
have the suffix. Buchanan and Shortliffe are the names associated
with this
Statistical
techniques and rule-based techniques are both in use.
RP
mentioned Snow’s Two Cultures,
the Sciences and the Humanities, and the communication breakdown
between them. See
Why
this is relevant to AI was not clear, except that RP, like Snow, desired a coming together of the two cultures again — so that
problems may be solved faster?
RP
said that machines can learn by being provided enough examples, e.g.
they can recognise handwriting if enough samples of a person’s
handwriting and the exact transcription are provided. From that point
on by correcting the guesses of the machine, the machine recognition
can become more accurate progressively.
Machines
can also learn to play games like chess (IBM-Deep Blue) and Go
(Google acquired a British company called Deep Mind which developed a program called Alpha Go). Extremely specialised game information and
hardware to process the search among possible moves, and giving a
score to each possibility, and looking many moves deep (far more than even a master player can foresee) enables these machine
to beat the very best human players.
IBM Deep Blue chess-playing computer which beat Kasparov
You can watch this video to learn about the Alpha Go program:
Champion Go player Lee Sedol (right) makes a move in his match against Google's artificial intelligence program, AlphaGo, at the Google DeepMind Challenge Match in Seoul, South Korea
A paper in the scientific journal Nature
describes "the technical details behind the approach to computer Go that combines Monte-Carlo tree search with deep neural networks that have been trained by supervised learning, from human expert games, and by reinforcement learning from games of self-play."
Spam
filters can be considered a simple example of programs that learn by
example from what users classify as spam, and because the resulting
knowledge is held in the cloud it is shared across millions of users,
and the spam filter becomes more and more sharp in its
discrimination.
RP
gave another example: a picture of vegetable vendors in a market. A
program can analyse the pic and come up with a verbal description
that describes its essential features, e.g. “a group of people
shopping in an outdoor market with many vegetables.”
Machines
can read, write, see images, read X-rays, examine pics of tissues and
detect cancers.
Diabetic
retinopathy is another area of expertise in which machines have been
trained to recognise features from photographs of the fundus, and and
they are capable of detecting fine features and counting them and
suggesting a diagnosis. Because these pics are in the
cloud, the diagnoses of many thousands, if not millions of examples
are shared, the machine has an extremely large knowledge base with which to
work; it can come up with its fairly accurate diagnosis quite fast. The AI system's diagnosis is found to be more accurate
than that of a human expert.
Machines
can also navigate the road (the Google self-driving car has done a million
miles with few accidents, none fatal) and recently it was given a
ticket by a cop for moving too slowly. Joe asked if these cars will
only succeed if every other car on the road is similarly disciplined
and automated in its driving, and not composed of Kochi-style
drivers who obey no traffic rules.
RP
gave an example of a brick-laying robot. Not only straight walls, but
the most daring and beautiful curved wall with protuberances can be
laid because the robot works to the design plan it is fed.
The
curved brick wall above, which was built by a robot builder, is a research project with
ETH Zurich and Reto Geiser.
Voice
to text dictation systems have been around for decades and when this
is harnessed to search, it is short step to having voice-enabled Google search
which has become quite commonplace. Siri is an Apple system that
can carry out searches for various things by voice command; here, natural language understanding on an elementary level is coupled to
search to provide easier ways of finding what you want, using
hand-held mobile devices.
RP
raised the question: where are we headed, leaving aside features of
Creativity and Empathy (C&E), which are not particularly strong
features of AI machines at present.
Audience interaction
Well,
short of C&E AI-endowed machines (computers with large banks of
memory, fast processors, parallel processing abilities, and access to large knowledge bases) can still
accomplish a lot of things which are commonplace in modern civilisation, e.g.
drive a car, act as a pathologist, guide a ship, etc.
—>
They can lead to cheap manufactured goods. He gave the example of mines in the extremely hot outback country in Australia that are operated by
machines controlled by women from air-conditioned offices a thousand miles away in
Canberra.
—>
You can manufacture buildings with lower cost and higher quality
—>
You can produce cheaper solar energy. RP mentioned that renewable
sources like wind, geothermal, wave energy, etc are limited in extent
and capacity compared to solar energy. According to him a small
region of USA (the panhandle of Texas - not so small really) can
supply the entire energy of the USA, even extrapolating to the increased
needs 50 years hence, provided it is covered with solar panels and transmission lines to distribute the energy. See Elon
Musk’s talk in this connection, which adds the wonder of Gigapower batteries to store energy at night
—>
You can expand healthcare until it is ubiquitous and affordable (of
course you have to deal with the Pharma companies in USA who are
hell-bent on ripping off patients with exorbitant prices!)
—>
RP held out the futuristic notions of Law Enforcement via AI robots.
He said it need not lead to an Orwellian future necessarily, and these robots can be taught that Black Lives Matter.
—>
Public Distribution System (PDS) in India. He envisaged a system in which
people have little role and therefore it will be (a) fair, (b)
corruption-free. RP has perhaps under-estimated the extreme application of
human intelligence in India to outwit any system!
RP
concluded with a general statement: “Machines can do most things
better than most people.”
“In
2013, Oxford scholars Carl Benedikt Frey and Michael A Osborne warned
that approximately 47% of total US employment was at risk of
computerisation, in an analysis that ranked 702 occupations by their
likelihood of being eliminated.” See
So what about Jobs for ordinary folk? RP proposed decoupling of Jobs from
Income and envisaged a world where everyone is paid a ‘Living Wage’
or Guaranteed Income. About the long history of this idea see
Thus far guaranteed income has only been thought of as an
anti-poverty measure, not as a way of rewarding people well for work
not done!
RP
opined that the oil-rich ME countries already provide this, and in
the world to which we are being inexorably led by AI, machines will obviate
human work, and therefore it will be necessary to tackle the
redundancy of humans. When he said this future is only ten years
away, Joe interjected that it is a wonderful future, but unless you
include Economists and Politicians and convince them the world of
human redundancy is around the corner, and a Guaranteed Middle-class Income is the
solution, it will not come about. It smacks of high socialism and we
know how averse Americans are to that word. Another caveat is that since the advent of steam engines in the mid-nineteenth century people have been saying similar things, but humans are no closer to a toil-free future. For example, here is Keynes forecasting in a 1930 essay:
For the first time since his creation man will be faced with his real, his permanent problem – how to use his freedom from pressing economic cares, how to occupy the leisure, which science and compound interest will have won for him, to live wisely and agreeably and well.
In spite of the highly touted productivity of America Joe thinks there is no advanced country where people work longer hours, for
example his own children. And Bernie Sanders still has trouble
convincing people that healthcare is a Right everyone should have, and
he is ridiculed as a ‘loony’ for proposing that college education should
be tuition-free in public universities and colleges; so there is a
complete disconnect between the world of large-scale human redundancy that will come about through AI and robotics, and its antidote comprising the extensive
socialism RP envisages. Before that utopia can supervene, the present world
of politics has to undergo a revolution of the kind Bernie Sanders is advocating for America.
Moreover, the only countries that RP cited which give their citizens a
guaranteed income are single commodity-based (oil) and that has
become a curse for most of them, who never bothered to develop other
resources, particularly human resources and agriculture. Joe only
knows of Norway with its Sovereign Fund of approximately $900bn that
is managed properly and prudently; other similar funds such as those of Saudi
Arabia and Kuwait are somewhat smaller but are wasting their capital by populist
measures and the funds are depleting on account of the low price that oil
commands at present. Countries like Sweden are able to manage without
oil income because they are first-class at doing a limited number of
things (special steels, timber, hydropower, defence production and
armaments, telecomm, pharmaceuticals, and Sweden has a modern
agricultural economy, besides. The Swedes constantly have to find
their palce in a changing world by being first rate at some things, and being self-sufficient in a large fraction of their essential
needs.
RP
broached the idea that decision-making processes in medical
diagnostics are Convergent, meaning I suppose that in a short finite time
those systems will come up with an answer or possible answers.
He
said Creativity however is not Convergent, although the way he
defined Creativity (“the ability to form novel combinations”)
does make it look rather trite. Where are the inputs for these
combinations? If they already exist then RP is basically positing
that all discoveries made to date by humans by a creative process
involving observations, hypotheses, trial, error, mathematical
deductions from hypothesis, experiment, etc — were already there
from the beginning of time and it just needed combining and sifting.
Anyway that is another debate.
Geoffrey
Hinton at the U of Toronto a few years ago found that Deep Learning
is Convergent. You can read an excellent article linked below. It deals
with IBM’s Watson computer which is now being targeted for use in medical
diagnosis. Watson was the question answering machine designed originally around the game fo Jeopardy, to come up with a precise answer to a question posed in natural language, See:
https://en.wikipedia.org/wiki/Watson_(computer)
A similar use is proposed for IBM Deep Blue, the AI machine which beat Kasparov at chess.
https://en.wikipedia.org/wiki/Watson_(computer)
A similar use is proposed for IBM Deep Blue, the AI machine which beat Kasparov at chess.
This
article speaks of the three recent innovations that have made AI more
profound and speedier to compute with and easier to use.
1.
Cheap
Parallel Computation,
using mass-produced graphical processing units (GPUs), originally
devised to quickly compute and change the screen-images needed for
fast moving computer games.
2.
Big
Data
The presence of large sets of data concerning almost any
problem area so that the training of AI machines to make them
accurate is speeded up.
3.
Better
Algorithms,
specifically the tweak to learning in layered neural nets (neural
nets are a way of recognising patterns using digital simulations of
analog circuits that propagate a signal from one layer to a higher
layer when it detects an elementary feature of the object to be
recognised within that layer). Geoffrey Hinton of the U of Toronto
was able to mathematically optimise results from each layer so that
the learning accumulated faster as it proceeded up the stack of
layers. This is now called Deep Learning.
RP
mentioned the ImageNet 1000 challenge in which a group of images are
submitted to AI machines which then have to classify them. You can
read how a group at Microsoft Beijing Labs solved this with <5%
error using a special set of neural networks
Progress
will be fast, but can it be stopped as we approach a Brave New World?
Even new food recipes have been created by AI which are great to taste but
no one had thought of them before, so unusual were the combinations of
ingredients.
A
question was raised by the senior person, a psychiatrist, Dr George
John, in the audience: a Picasso may be simulated by a future AI
machine, but will a machine develop new literary works comparable to
that of a Dante or a Proust? He doubted it. Even in neurology a
cognitive behaviour can be simulated, but what about the work of a therapist who
has to listen and interact in many different ways depending on the
patient and the context and the previous history; the need for
Empathy at that point is essential. Why did Dr George John think
Picasso was within reach of machines to imitate? Was it because
Picasso was the one who said said: I
do not seek, I find
— an apt counter to Google?
John Outcalt, RP, Dr George John, Shruti
RP
interjected that machines can detect the state of internal feeling of
the subject by observing and having access to many variables — pupil
dilation, wetness of the palms, flush of the skin, heartbeat, BP, etc. and it can detect a smile, perhaps different kinds of smiles.
Sometimes the machines that do useful things for the professionals
(like AI-based bomb-disposal machines) become objects of feeling for
those who work with them. When one such machine blew up in the line
of duty, the minders gave it a ceremonial burial. In the same fashion a Roomba vacuum-cleaning robot had such a claim on the affection of an old lady that she would not part with it when it needed to be taken to the shop for repairs.
A
question from Tanya Kurishingal raised the matter of developing the
right brain (right-brain people are allegedly more creative and intuitive); if this is done systematically in students they may turn out to be more creative in the
sciences, in spite of the fact that right brain activities are
largely classified as artistic activities. In this connection an
anecdote about the famous quantum physicist, Paul Dirac, illustrates
the dichotomy. He spoke to Robert Oppenheimer, the polymath physicist and father of the
A-Bomb, who was well-known for quoting from esoteric books like the Rig
Veda and T.S. Eliot's poetry, and was a dabbler in the arts himself. Dirac tried to
dissuade Oppenheimer from writing poetry with this dictum: “In science
you want to say something that nobody knew before, in words which
everyone can understand. In poetry you are bound to say ... something
that everybody knows already, in words that nobody can understand.”
Finally,
RP came to Consciousness. It has been stated that being self-aware is
a part of consciousness. Cogito,
ergo sum
is merely a statement about being self-ware and therefore extant.
In which connection there is this joke: René Descartes walks into his
favourite bar. The bartender says "Hey, René, gonna have your
usual?"
"I
don't think I am." And he disappears.
RP
stated that Consciousness is like a movie running in our brain all
the while in which we ourselves are an actor and we are constantly
building up and storing images. But this, he said, comes with the
limitation of the storage capacity and processing capacity and speed
of recall of the brain. So with AI we may predict a future in which
you have a brain outside
of
your biological brain with which you are constantly in touch. Will it
then be a heightened sate of consciousness that we will all possess? With
the help of an implant?
We
have to remember that animals too possess consciousness.
To
conclude RP referred to the Ant Man, E.O. Wilson of Harvard
who
asked the question in his book On
Human Nature
(1978): What if we had the ability to change our nature? RP mentioned
something called Crisper, or so I heard, without any explanation; but
what was meant presumably is Genome Editing with CRISPR-Cas9; see the video at
and,
if you are a biologist, look at
Hi Joe! That's a good piece of writing.
ReplyDeleteEnjoyed reading it. Thanks.
K2
Such a detailed account...by JOURNALIST DR.JOE CLEETUS. Hats off to you Joe.
ReplyDeleteRahul is the 2013 recipient of the ECONOMIST MAGAZINE AWARD.
AND he is also a MIT 35..when MIT chooses 35 young achievers under the age of 35..from around the world.