sponsored links
TED2011

Rajesh Rao: A Rosetta Stone for a lost language

March 2, 2011

Rajesh Rao is fascinated by "the mother of all crossword puzzles": how to decipher the 4000-year-old Indus script. He's enlisting modern computation to try to read this lost language, the key to understanding this ancient civilization.

Rajesh Rao - Computational neuroscientist
Rajesh Rao seeks to understand the human brain through computational modeling, on two fronts: developing computer models of our minds, and using tech to decipher the 4,000-year-old lost script of the Indus Valley civilization. Full bio

sponsored links
Double-click the English subtitles below to play the video.
I'd like to begin with a thought experiment.
00:15
Imagine that it's 4,000 years into the future.
00:19
Civilization as we know it
00:22
has ceased to exist --
00:24
no books,
00:26
no electronic devices,
00:28
no Facebook or Twitter.
00:31
All knowledge of the English language and the English alphabet
00:34
has been lost.
00:37
Now imagine archeologists
00:39
digging through the rubble of one of our cities.
00:41
What might they find?
00:43
Well perhaps some rectangular pieces of plastic
00:45
with strange symbols on them.
00:48
Perhaps some circular pieces of metal.
00:51
Maybe some cylindrical containers
00:54
with some symbols on them.
00:56
And perhaps one archeologist becomes an instant celebrity
00:58
when she discovers --
01:01
buried in the hills somewhere in North America --
01:03
massive versions of these same symbols.
01:05
Now let's ask ourselves,
01:10
what could such artifacts say about us
01:12
to people 4,000 years into the future?
01:15
This is no hypothetical question.
01:18
In fact, this is exactly the kind of question we're faced with
01:20
when we try to understand the Indus Valley civilization,
01:23
which existed 4,000 years ago.
01:26
The Indus civilization was roughly contemporaneous
01:28
with the much better known Egyptian and the Mesopotamian civilizations,
01:31
but it was actually much larger than either of these two civilizations.
01:34
It occupied the area
01:37
of approximately one million square kilometers,
01:39
covering what is now Pakistan,
01:41
Northwestern India
01:43
and parts of Afghanistan and Iran.
01:45
Given that it was such a vast civilization,
01:47
you might expect to find really powerful rulers, kings,
01:49
and huge monuments glorifying these powerful kings.
01:53
In fact,
01:56
what archeologists have found is none of that.
01:58
They've found small objects such as these.
02:00
Here's an example of one of these objects.
02:03
Well obviously this is a replica.
02:06
But who is this person?
02:08
A king? A god?
02:11
A priest?
02:13
Or perhaps an ordinary person
02:15
like you or me?
02:17
We don't know.
02:19
But the Indus people also left behind artifacts with writing on them.
02:21
Well no, not pieces of plastic,
02:24
but stone seals, copper tablets,
02:26
pottery and, surprisingly,
02:29
one large sign board,
02:31
which was found buried near the gate of a city.
02:33
Now we don't know if it says Hollywood,
02:35
or even Bollywood for that matter.
02:37
In fact, we don't even know
02:39
what any of these objects say,
02:41
and that's because the Indus script is undeciphered.
02:43
We don't know what any of these symbols mean.
02:46
The symbols are most commonly found on seals.
02:48
So you see up there one such object.
02:51
It's the square object with the unicorn-like animal on it.
02:53
Now that's a magnificent piece of art.
02:56
So how big do you think that is?
02:58
Perhaps that big?
03:00
Or maybe that big?
03:02
Well let me show you.
03:04
Here's a replica of one such seal.
03:07
It's only about one inch by one inch in size --
03:10
pretty tiny.
03:12
So what were these used for?
03:14
We know that these were used for stamping clay tags
03:16
that were attached to bundles of goods that were sent from one place to the other.
03:19
So you know those packing slips you get on your FedEx boxes?
03:22
These were used to make those kinds of packing slips.
03:25
You might wonder what these objects contain
03:28
in terms of their text.
03:31
Perhaps they're the name of the sender
03:33
or some information about the goods
03:35
that are being sent from one place to the other -- we don't know.
03:37
We need to decipher the script to answer that question.
03:40
Deciphering the script
03:42
is not just an intellectual puzzle;
03:44
it's actually become a question
03:46
that's become deeply intertwined
03:48
with the politics and the cultural history of South Asia.
03:50
In fact, the script has become a battleground of sorts
03:53
between three different groups of people.
03:56
First, there's a group of people
03:58
who are very passionate in their belief
04:00
that the Indus script
04:02
does not represent a language at all.
04:04
These people believe that the symbols
04:06
are very similar to the kind of symbols you find on traffic signs
04:08
or the emblems you find on shields.
04:11
There's a second group of people
04:14
who believe that the Indus script represents an Indo-European language.
04:16
If you look at a map of India today,
04:19
you'll see that most of the languages spoken in North India
04:21
belong to the Indo-European language family.
04:24
So some people believe that the Indus script
04:27
represents an ancient Indo-European language such as Sanskrit.
04:29
There's a last group of people
04:32
who believe that the Indus people
04:34
were the ancestors of people living in South India today.
04:37
These people believe that the Indus script
04:40
represents an ancient form
04:42
of the Dravidian language family,
04:44
which is the language family spoken in much of South India today.
04:46
And the proponents of this theory
04:49
point to that small pocket of Dravidian-speaking people in the North,
04:51
actually near Afghanistan,
04:54
and they say that perhaps, sometime in the past,
04:56
Dravidian languages were spoken all over India
04:59
and that this suggests
05:02
that the Indus civilization is perhaps also Dravidian.
05:04
Which of these hypotheses can be true?
05:07
We don't know, but perhaps if you deciphered the script,
05:10
you would be able to answer this question.
05:12
But deciphering the script is a very challenging task.
05:14
First, there's no Rosetta Stone.
05:16
I don't mean the software;
05:18
I mean an ancient artifact
05:20
that contains in the same text
05:22
both a known text and an unknown text.
05:24
We don't have such an artifact for the Indus script.
05:27
And furthermore, we don't even know what language they spoke.
05:30
And to make matters even worse,
05:33
most of the text that we have are extremely short.
05:35
So as I showed you, they're usually found on these seals
05:37
that are very, very tiny.
05:39
And so given these formidable obstacles,
05:41
one might wonder and worry
05:43
whether one will ever be able to decipher the Indus script.
05:45
In the rest of my talk,
05:48
I'd like to tell you about how I learned to stop worrying
05:50
and love the challenge posed by the Indus script.
05:52
I've always been fascinated by the Indus script
05:54
ever since I read about it in a middle school textbook.
05:57
And why was I fascinated?
05:59
Well it's the last major undeciphered script in the ancient world.
06:01
My career path led me to become a computational neuroscientist,
06:05
so in my day job,
06:08
I create computer models of the brain
06:10
to try to understand how the brain makes predictions,
06:12
how the brain makes decisions,
06:15
how the brain learns and so on.
06:17
But in 2007, my path crossed again with the Indus script.
06:19
That's when I was in India,
06:22
and I had the wonderful opportunity
06:24
to meet with some Indian scientists
06:26
who were using computer models to try to analyze the script.
06:28
And so it was then that I realized
06:31
there was an opportunity for me to collaborate with these scientists,
06:33
and so I jumped at that opportunity.
06:36
And I'd like to describe some of the results that we have found.
06:38
Or better yet, let's all collectively decipher.
06:40
Are you ready?
06:43
The first thing that you need to do when you have an undeciphered script
06:45
is try to figure out the direction of writing.
06:48
Here are two texts that contain some symbols on them.
06:50
Can you tell me
06:53
if the direction of writing is right to left or left to right?
06:55
I'll give you a couple of seconds.
06:58
Okay. Right to left, how many? Okay.
07:01
Okay. Left to right?
07:04
Oh, it's almost 50/50. Okay.
07:06
The answer is:
07:08
if you look at the left-hand side of the two texts,
07:10
you'll notice that there's a cramping of signs,
07:12
and it seems like 4,000 years ago,
07:15
when the scribe was writing from right to left,
07:17
they ran out of space.
07:19
And so they had to cram the sign.
07:21
One of the signs is also below the text on the top.
07:23
This suggests the direction of writing
07:25
was probably from right to left,
07:27
and so that's one of the first things we know,
07:29
that directionality is a very key aspect of linguistic scripts.
07:31
And the Indus script now has
07:34
this particular property.
07:36
What other properties of language does the script show?
07:38
Languages contain patterns.
07:40
If I give you the letter Q
07:42
and ask you to predict the next letter, what do you think that would be?
07:44
Most of you said U, which is right.
07:47
Now if I asked you to predict one more letter,
07:49
what do you think that would be?
07:51
Now there's several thoughts. There's E. It could be I. It could be A,
07:53
but certainly not B, C or D, right?
07:56
The Indus script also exhibits similar kinds of patterns.
07:59
There's a lot of text that start with this diamond-shaped symbol.
08:02
And this in turn tends to be followed
08:05
by this quotation marks-like symbol.
08:07
And this is very similar to a Q and U example.
08:09
This symbol can in turn be followed
08:11
by these fish-like symbols and some other signs,
08:13
but never by these other signs at the bottom.
08:16
And furthermore, there's some signs
08:18
that really prefer the end of texts,
08:20
such as this jar-shaped sign,
08:22
and this sign, in fact, happens to be
08:24
the most frequently occurring sign in the script.
08:26
Given such patterns, here was our idea.
08:28
The idea was to use a computer
08:31
to learn these patterns,
08:33
and so we gave the computer the existing texts.
08:35
And the computer learned a statistical model
08:38
of which symbols tend to occur together
08:40
and which symbols tend to follow each other.
08:42
Given the computer model,
08:44
we can test the model by essentially quizzing it.
08:46
So we could deliberately erase some symbols,
08:49
and we can ask it to predict the missing symbols.
08:51
Here are some examples.
08:54
You may regard this
09:00
as perhaps the most ancient game
09:02
of Wheel of Fortune.
09:04
What we found
09:08
was that the computer was successful in 75 percent of the cases
09:10
in predicting the correct symbol.
09:12
In the rest of the cases,
09:14
typically the second best guess or third best guess was the right answer.
09:16
There's also practical use
09:19
for this particular procedure.
09:21
There's a lot of these texts that are damaged.
09:23
Here's an example of one such text.
09:25
And we can use the computer model now to try to complete this text
09:27
and make a best guess prediction.
09:30
Here's an example of a symbol that was predicted.
09:32
And this could be really useful as we try to decipher the script
09:35
by generating more data that we can analyze.
09:37
Now here's one other thing you can do with the computer model.
09:40
So imagine a monkey
09:43
sitting at a keyboard.
09:45
I think you might get a random jumble of letters that looks like this.
09:47
Such a random jumble of letters
09:50
is said to have a very high entropy.
09:52
This is a physics and information theory term.
09:54
But just imagine it's a really random jumble of letters.
09:56
How many of you have ever spilled coffee on a keyboard?
09:59
You might have encountered the stuck-key problem --
10:03
so basically the same symbol being repeated over and over again.
10:05
This kind of a sequence is said to have a very low entropy
10:08
because there's no variation at all.
10:11
Language, on the other hand, has an intermediate level of entropy;
10:13
it's neither too rigid,
10:16
nor is it too random.
10:18
What about the Indus script?
10:20
Here's a graph that plots the entropies of a whole bunch of sequences.
10:22
At the very top you find the uniformly random sequence,
10:26
which is a random jumble of letters --
10:28
and interestingly, we also find
10:30
the DNA sequence from the human genome and instrumental music.
10:32
And both of these are very, very flexible,
10:35
which is why you find them in the very high range.
10:37
At the lower end of the scale,
10:39
you find a rigid sequence, a sequence of all A's,
10:41
and you also find a computer program,
10:43
in this case in the language Fortran,
10:45
which obeys really strict rules.
10:47
Linguistic scripts
10:49
occupy the middle range.
10:51
Now what about the Indus script?
10:53
We found that the Indus script
10:55
actually falls within the range of the linguistic scripts.
10:57
When this result was first published,
10:59
it was highly controversial.
11:01
There were people who raised a hue and cry,
11:04
and these people were the ones who believed
11:07
that the Indus script does not represent language.
11:09
I even started to get some hate mail.
11:12
My students said
11:14
that I should really seriously consider getting some protection.
11:16
Who'd have thought
11:19
that deciphering could be a dangerous profession?
11:21
What does this result really show?
11:23
It shows that the Indus script
11:25
shares an important property of language.
11:27
So, as the old saying goes,
11:29
if it looks like a linguistic script
11:31
and it acts like a linguistic script,
11:33
then perhaps we may have a linguistic script on our hands.
11:35
What other evidence is there
11:38
that the script could actually encode language?
11:40
Well linguistic scripts can actually encode multiple languages.
11:42
So for example, here's the same sentence written in English
11:45
and the same sentence written in Dutch
11:48
using the same letters of the alphabet.
11:50
If you don't know Dutch and you only know English
11:52
and I give you some words in Dutch,
11:55
you'll tell me that these words contain
11:57
some very unusual patterns.
11:59
Some things are not right,
12:01
and you'll say these words are probably not English words.
12:03
The same thing happens in the case of the Indus script.
12:06
The computer found several texts --
12:08
two of them are shown here --
12:10
that have very unusual patterns.
12:12
So for example the first text:
12:14
there's a doubling of this jar-shaped sign.
12:16
This sign is the most frequently-occurring sign
12:19
in the Indus script,
12:21
and it's only in this text
12:23
that it occurs as a doubling pair.
12:25
Why is that the case?
12:27
We went back and looked at where these particular texts were found,
12:29
and it turns out that they were found
12:32
very, very far away from the Indus Valley.
12:34
They were found in present day Iraq and Iran.
12:36
And why were they found there?
12:39
What I haven't told you is that
12:41
the Indus people were very, very enterprising.
12:43
They used to trade with people pretty far away from where they lived,
12:45
and so in this case, they were traveling by sea
12:48
all the way to Mesopotamia, present-day Iraq.
12:51
And what seems to have happened here
12:54
is that the Indus traders, the merchants,
12:56
were using this script to write a foreign language.
12:59
It's just like our English and Dutch example.
13:02
And that would explain why we have these strange patterns
13:04
that are very different from the kinds of patterns you see in the text
13:06
that are found within the Indus Valley.
13:09
This suggests that the same script, the Indus script,
13:12
could be used to write different languages.
13:14
The results we have so far seem to point to the conclusion
13:17
that the Indus script probably does represent language.
13:20
If it does represent language,
13:23
then how do we read the symbols?
13:25
That's our next big challenge.
13:27
So you'll notice that many of the symbols
13:29
look like pictures of humans, of insects,
13:31
of fishes, of birds.
13:33
Most ancient scripts
13:36
use the rebus principle,
13:38
which is, using pictures to represent words.
13:40
So as an example, here's a word.
13:43
Can you write it using pictures?
13:46
I'll give you a couple seconds.
13:48
Got it?
13:50
Okay. Great.
13:52
Here's my solution.
13:54
You could use the picture of a bee followed by a picture of a leaf --
13:56
and that's "belief," right.
13:58
There could be other solutions.
14:00
In the case of the Indus script,
14:02
the problem is the reverse.
14:04
You have to figure out the sounds of each of these pictures
14:06
such that the entire sequence makes sense.
14:09
So this is just like a crossword puzzle,
14:11
except that this is the mother of all crossword puzzles
14:14
because the stakes are so high if you solve it.
14:17
My colleagues, Iravatham Mahadevan and Asko Parpola,
14:21
have been making some headway on this particular problem.
14:24
And I'd like to give you a quick example of Parpola's work.
14:26
Here's a really short text.
14:28
It contains seven vertical strokes followed by this fish-like sign.
14:30
And I want to mention that these seals were used
14:33
for stamping clay tags
14:35
that were attached to bundles of goods,
14:37
so it's quite likely that these tags, at least some of them,
14:39
contain names of merchants.
14:42
And it turns out that in India
14:44
there's a long tradition
14:46
of names being based on horoscopes
14:48
and star constellations present at the time of birth.
14:50
In Dravidian languages,
14:53
the word for fish is "meen"
14:55
which happens to sound just like the word for star.
14:57
And so seven stars
15:00
would stand for "elu meen,"
15:02
which is the Dravidian word
15:04
for the Big Dipper star constellation.
15:06
Similarly, there's another sequence of six stars,
15:08
and that translates to "aru meen,"
15:11
which is the old Dravidian name
15:13
for the star constellation Pleiades.
15:15
And finally, there's other combinations,
15:17
such as this fish sign with something that looks like a roof on top of it.
15:20
And that could be translated into "mey meen,"
15:23
which is the old Dravidian name for the planet Saturn.
15:26
So that was pretty exciting.
15:29
It looks like we're getting somewhere.
15:31
But does this prove
15:33
that these seals contain Dravidian names
15:35
based on planets and star constellations?
15:37
Well not yet.
15:39
So we have no way of validating
15:41
these particular readings,
15:43
but if more and more of these readings start making sense,
15:45
and if longer and longer sequences
15:48
appear to be correct,
15:50
then we know that we are on the right track.
15:52
Today,
15:54
we can write a word such as TED
15:56
in Egyptian hieroglyphics and in cuneiform script,
15:59
because both of these were deciphered
16:02
in the 19th century.
16:04
The decipherment of these two scripts
16:06
enabled these civilizations to speak to us again directly.
16:08
The Mayans
16:11
started speaking to us in the 20th century,
16:13
but the Indus civilization remains silent.
16:15
Why should we care?
16:18
The Indus civilization does not belong
16:20
to just the South Indians or the North Indians
16:22
or the Pakistanis;
16:24
it belongs to all of us.
16:26
These are our ancestors --
16:28
yours and mine.
16:30
They were silenced
16:32
by an unfortunate accident of history.
16:34
If we decipher the script,
16:36
we would enable them to speak to us again.
16:38
What would they tell us?
16:40
What would we find out about them? About us?
16:43
I can't wait to find out.
16:46
Thank you.
16:49
(Applause)
16:51

sponsored links

Rajesh Rao - Computational neuroscientist
Rajesh Rao seeks to understand the human brain through computational modeling, on two fronts: developing computer models of our minds, and using tech to decipher the 4,000-year-old lost script of the Indus Valley civilization.

Why you should listen

Rajesh Rao is looking for the computational principles underlying the brain's remarkable ability to learn, process and store information --  hoping to apply this knowledge to the task of building adaptive robotic systems and artificially intelligent agents.

Some of the questions that motivate his research include: How does the brain learn efficient representations of novel objects and events occurring in the natural environment? What are the algorithms that allow useful sensorimotor routines and behaviors to be learned? What computational mechanisms allow the brain to adapt to changing circumstances and remain fault-tolerant and robust?

By investigating these questions within a computational and probabilistic framework, it is often possible to derive algorithms that not only provide functional interpretations of neurobiological properties but also suggest solutions to difficult problems in computer vision, speech, robotics and artificial intelligence.

The original video is available on TED.com
sponsored links

If you need translations, you can install "Google Translate" extension into your Chrome Browser.
Furthermore, you can change playback rate by installing "Video Speed Controller" extension.

Data provided by TED.

This website is owned and operated by Tokyo English Network.
The developer's blog is here.