>> From the Library of
Congress in Washington D.C.
>> Fenella France: Good
afternoon and welcome
to our topic in preservation
series.
I'm Fenella France, chief
of the Preservation Research
and Testing Division
and we're delighted
to have you here today.
The presentation today is The
Digital Restoration Initiative:
Reading the Invisible library.
And our speaker is
Dr. Brent Seales.
Brent is a professor and
chairman of the Department
of Computer Science and
the director of the Center
for Visualization and
Virtual Environments
at the University of Kentucky.
His research centers on computer
vision and visualization applied
to challenges in the
restoration of antiquities,
surgical technology,
and data visualization.
In 2012-13, he was a Google
Visiting Scientist in Paris,
where he continued work
on the virtual unwrapping
of the Herculaneum scrolls.
In 2015, Brent and his team
identified the oldest known
Hebrew copy at the book of
Leviticus, which is carbon dated
to the third century C.E.
And this has been a very
significant discovery
in biblical archeology.
It's particularly
interesting because one
of the things we
find in conservation
and preservation is
integrating new technologies
as a very important part of what
we can do particularly as we try
and do those noninvasively.
So I'd like to hand
over now to Dr. Seales
to give us a presentation.
Please join me in welcoming him.
[ Applause ]
>> W. Brent Seales: Thank
you, Fenella, and thank you
to the Library of Congress
for the invitation to come
and speak to you today.
It really is a pleasure
to be here from Kentucky,
and when people look at the
slide and I begin my talk,
often they wonder, where
is Kentucky exactly?
I'm not going to show you a map.
The other thing people
wonder is,
what is the digital
restoration initiative?
That's what I'm going
to tell you about today.
Before I start my talk, I'd like
to welcome the remote audience
and thank you for joining.
And I also would like to
thank one of my sponsors
at the University of Kentucky,
Lee and Stacie Marksbury
for their kind support.
In fact, the invisible
library, what I'm going to talk
about today is a phrase that
I coined from an article
that was published
in the New Yorker.
And that article was a
retrospective of work
that I had began two decades
ago at the British Library.
And here's a picture from that
era when really we were talking
about the digital library which
was about digitizing collections
and making them available
on something
that was new at the time.
We called it the World Wide Web.
And now we think of the
internet as being the place
where glorious digital editions,
similar to the one
we implemented
in the 1995 Macintosh can
be available for scholars,
researchers and the general
audience to be able to view
and enjoy and do scholarship on.
This was the beginning of my
odyssey into the digital library
which then became for me
the invisible library.
And I'm going to tell
you the story of that
and how I became
interested in some
of the most profoundly damaged
parts of the invisible library.
An example here is shown
from the collection in Naples
in Italy from the Villa of
the Papyri at Herculaneum.
This library discovered several
hundred years ago is the only
classical library ever
discovered in situ
and contained many thousands of
manuscripts that were carbonized
and damaged in the
form that you see here.
The retrospective that was
written in the New Yorker,
I won't ask for a show of hands
of how many of you read it.
Some of you did, I see.
Had a really wonderful review
of-- all things Herculaneum.
And the reporter John Seabrook
took the time to go through many
of the details of Herculaneum.
But I'm going to spoil the
end of the piece for you.
John ended the piece by saying,
by quoting one of my colleagues
who said, "I do not expect this
scroll would be read during
my lifetime.
And he closed the lid of the
small box with both hands,"
that box containing
one of those scrolls,
"his shoulder slumped
in defeat."
I hated this ending.
When I read this
ending, I almost--
I was in the living room with
my family, I almost wept.
And so I had read it and
weeping, I invited John
out to the University of
Kentucky so that I could talk
to him over bourbon
drinks and the racetrack
which is what we do at
Kentucky, and I asked him.
And of course we
had a lovely time
and John was really
just writing the story
about how hard it is actually
to read this material.
And as he followed my team
along and wrote the article,
he did sense that, you
know, we were struggling
because it is a hard problem.
And one of the things
that I want you
to understand today is
the timescale of the quest
that I have been on to be able
to read some of this material
and convert the invisible
library to the visible one.
And I also want you to
understand some of my optimism
about why I think we're on the
verge of being able to do some
of the most interesting
material.
I am fundamentally a--
an optimistic person.
And my mom pulled out from our
archives and you at the Library
of Congress will appreciate
the idea of my mother creating
for me an archive of the
Buffalo Evening News from 1969.
She pulled out the paper
that she saved for me
about the land--
the moon landing.
And I pulled that out just a
few months ago and opened it up
and remembered how I felt, you
know, in that era being inspired
by what we could do, that
we could engineer a way
to actually walk on the moon.
So I'm going to start with a
prologue about inspiration.
When I began this
work in the late 90's,
this idea of converting
the invisible library
into the visible library, and
I hadn't given that words.
Early in my work, I was sent a
letter from an unknown person
who had found evidence of
my work in the literature.
Her name is Cheryl Seacrist
and she lives in this area.
And Cheryl found
that I was doing work
on digital restoration.
And she said to herself, well,
you know, I have something
that needs to be restored.
So, she cold called me and she
actually sent me the artifact.
And I'm showing you a
picture of that artifact.
It was a one-page letter.
And it had been faded so
the text was not legible.
I took the letter to my lab
and this was prior to the time
that we had the sophisticated
equipment that we have now.
And I know Fenella will
understand when we talk
about spectral imaging
and special light sources,
we just didn't have those
things at the time in the lab.
We also didn't have the kinds of
sensors we needed to do imaging.
We did have flatbed
scanners though,
that were incredibly precise
and very high resolution.
And I started to
play with this letter
and found the transformation
in the color spaces
that were available
that allowed me
to highlight exactly the place
in that flatbed scanner scanned
image that would bring out some
of the text, giving me a handle
on a process for restoration.
And so we put a team
together and I had, by hand,
a team of students
walk through the data
and do a complete restoration.
Now what had really intrigued me
and the reason why I bring this
to you now, what really
intrigued me about this is not
that we were being able to
bring text that was invisible
to the visible place but
that the text represented
a narrative.
And I was really interested
in what the narrative was.
I mean I had in my hands
evidence of a story
but I couldn't read the story
until we solved the
technical problem.
"Dearest Cheryl Ann.
Hello, my little darling.
How are you this evening?
Fine I hope."
It turned out that this letter
was from a father who is
in the World War II Pacific
Theater on a landing ship tank.
It was from him to his daughter
and his daughter Cheryl
was the one who ended
up sending me the letter to see
if I could do a restoration.
How wonderful it was
when we read the letter
to realize Frank closes the
letter, "Be a good girl.
Oceans of love and kisses.
Please write soon."
Of course, I think Cheryl
might have been three.
"Love, your daddy Frank."
And, you know, what
a beautiful image.
He is looking at an ocean
and it doesn't really
look so kind, right?
It's the World War II Pacific
Theater that he writes oceans
of love because he's
thinking about his daughter.
So all of this really
resonated with me
and provided the
inspiration, as I now look back,
for the approach to some pretty
profoundly damaged materials
that I now call the
invisible library.
On the left, you see
an ancient manuscript--
medieval manuscript actually
from the cotton collection
at the British Library,
left unrestored, damaged.
Most of the volumes in the
collection were restored
and I'll show you in a
minute what they looked like,
that restoration.
On the right, you
see court documents,
probably Ming Dynasty.
These are in the archives at
the National Palace Museum,
museum in Taipei, Taiwan.
And of course, you have the
iconic Herculaneum scrolls
of which there are probably 2
or 300 now still remaining
unopened, little time capsules
from the classical era
representing the possibility
of who knows what if we
could only read them.
And I will also talk today
about what Fenella mentioned
in the introduction which
is the scroll from En-Gedi,
the scroll having come to me
through no fault of my own
from three years back
and turned out to be one
of our biggest successes.
So the roadmap for my talk
is to first inspire you
with the textual so
that you understand
that not only can we
recover readable text
but we can recover
text to the degree
that we can understand
narrative and do scholarship.
And I'm going to give you
a flavor of the technical
so that you can understand
exactly why this is a hard
problem and what's changed in
the last 10 years to be able
to allow us to solve
this problem.
And then I'll finish my talk
by speaking a little bit
about the collaborative
which turns out to be one
of the most crucial pieces
of the puzzle to be able
to read the invisible library.
So let me start with digital
restoration part 1, pages.
In the mid 2000s, I had
the chance to do a project,
the Marciana Library,
in collaboration
with Harvard University, the
Center of Hellenic Studies
which is right here
in Washington D.C. ran
by Harvard University.
And the project was to
capture a digital copy,
make a digital facsimile
of the "Venetus A"
which is the oldest complete
copy of Homer's Iliad.
And of course you know
the story of Iliad
because we saw the
movie with Brad Pitt.
And those of us who are computer
scientists use that as the way
to solidify our education
of the classics.
Prior to the in facsimile, we
made in 2008 of the "Venetus A",
there was one other facsimile.
It was 1908-- early
1900's, either '01 or '08,
photography done by an Italian
by the name of Comparetti.
And this image you see is the
quality of that facsimile.
When I show you on top of that
the quality of our facsimile,
not only is it better in color
because the original facsimile
was black and white photography,
but it is also better
in resolution.
Both of those features,
color and resolution,
end up providing scholars
with so much more information.
So the facsimile simply as
photography that we were able
to make advanced the field
because scholars could
take both the color
and the high resolution
and the fact
that we could digitally
disseminate this.
Globally, we actually negotiated
a creative commons agreement
for the information
we collected.
Allowed for a never before--
a facsimile that was richer
than had ever been done
before on the "Venetus A".
So, again, here you see a
section of the decoration on one
of the pages, it is a medieval
manuscript and then you see,
as a layer that I fade up
the corresponding result
from our facsimile and you
see how much better it is.
Now, a few surprising things
came out of this work.
Not only was I getting my
bearings and what it meant
to do digital restoration
but on the left,
you see in Comperatti's edition,
there is actually more
information available
and visible than in ours because
in the intervening 100 years,
things actually disappeared.
And we were actually
surprised by that.
I don't know why now that I look
back on it but I just thought,
well, it's better imaging and so
we'll always have better results
but some things disappear.
And so the comparison
between the two,
and this as I play this
fade from one to the other,
will show you as a set of
layers how you can compare the
information that is available
in one and the other.
One thing I learned is that
when an artifact has been imaged
over time, it's really
important not to get rid of any
of those examples but to
try to find a coherent way
to put them all together.
So we began to think
in terms of pages,
not just sequentially
representing a manuscript
but temporally, right,
representing a manuscript
as it has been photographed
over time, and try to find a way
to put those together.
And we developed
some infrastructure
to be able to do that.
Second thing with regard to
pages in that era that we began
to develop was the idea
of digital restoration,
which is to imagine that
after you've taken an image
of something that's
badly damaged,
you can without risk start
to envision the possibility,
and here I'll highlight a
damaged region and I will blow
up this region for you to see.
You can imagine the possibility
of a digital restoration,
maybe by an artist or by a
scholar who really understood
from the data what
really was there.
And there's no risk at all
to the object to be able
to view this as a
layer, as a page.
And it can be very
interesting scholar--
for scholars to be able
to play around this way
and it can also be
very entertaining
to imagine what should have been
there, what maybe was there.
But one thing that it's not is
risky because it's all digital
and the original artifact is set
aside so that's it's protected.
Wouldn't it be nice if in every
case we paid a little more
attention to the physical, you
know, disposition as oppose
to letting, you know,
certain restorers run free.
Sometimes, you know,
it's a lot safer
to do the digital restoration.
So let me just summarize
this by saying
with pure manuscript pages,
I began to develop this idea
of comparisons over time, pages
that have been taken over time,
put together to be compared
and the visualization
of that using some technology
like the registration
of pages so that they align.
And I want to point to before I
move on, one of the key results
from this era by my research
team and others having
to do with registration.
So here's an example of a
photograph of a fragment
from the Dead Sea
Scroll collection
and a second photograph
taken at a different time
with a different camera
from an earlier era.
And if I go back and forth
between these two images,
you'll see that the text
doesn't line up at all.
If you focus on the text,
the scale is different.
Some text is more visible
in this ultraviolet shot,
infrared shot than it is visible
in the visible light shot.
And so we undertook the idea
of registering these together
based on just the text.
And so now I'm going to show
you the algorithm running
to do the registration.
Registration became a core
technology for us and others
in putting together
things that have been taken
at different times with
different devices over time.
So what's happening in this
video is first we're seeking a
rigid transformation that tries
to match the text from one image
onto the text of
the other image.
When that fails, we do
a nonrigid registration
which deforms nonrigidly the
text of one on to the other.
And then when I fade back,
you'll see that the
text matches exactly.
And what we discovered
in using registration
as a core technology is that
if you focus on registration
around things like the text
which is the key thing,
then you can look at how other
things have changed over time
that becomes a really
powerful way
to reconcile the diachronic
part, that part over time
of a collection as
it's been photographed.
OK, so this takes me
to the second part
of reading the invisible
library, which is the fact
that not everything
we photograph is flat.
In fact, it's far from flat.
It's almost impossible
unless you're Comparetti
on the Piazza San Marco in 1908
where you've been given the
permission to slice the binding
from the book, and you've
been given the permission
to press each one of
the pages between glass,
you've been given the permission
to use the bright sunlight
on the plaza so that the
lens can capture the best
possible image.
So in that case,
things are pretty flat.
But in the case for the rest
of us in the modern era,
at the Library of Congress
for example, it's not going
to really be possible to
take a manuscript like this.
And this is one of the
restored manuscripts
from the cotton collection.
And what you see here is the
original parchment, the vellum,
having been inserted
into paper frames.
Paper and the vellum deform
at different rates over time
and they cause this
cockling to occur.
It happens naturally.
And that cockling means
that when you photograph
that page assuming that it's
flat, it creates deformations
in the letter forms, right?
And even in those squares that
were drawn around the boxes,
right, those are
no longer boxes.
It becomes really hard to remove
the ambiguity of whether that's
in the manuscript itself or
whether that deformation is part
of the fact that the
manuscript is wrinkled.
So we began to develop a way to
capture not just the photography
but also the shape of
every manuscript page
as we did the digitization
that shape helping
us do the next step
in digital restoration.
So here you're seeing our
early version of a system
that allowed us to capture
the shape of every page,
reconcile the shape
with the actual image.
And then following
the capture of that,
do what I call digital
flattening.
And the digital flattening
allows us to take the page
as it sat and then
run a simulation
to render what it would like
had we been able to iron it out.
It's sort of digital ironing.
Pioneered this in the mid to
late 2000s as a way to try
to recover the shape problem as
opposed to just the photography
so with layers and
pages, we worked on ways
to amplify the image then
we started working on ways
to recover and restore
the shape.
And of course you see this
is all done without risk
to the real object premised
on the idea of being able
to capture more information
at digitization time.
So flattening allowed us to
get at pages that were more
than just photography
but included the shape
3D representation
and then the visualization
of that either as 3D
or as 3D then flattened so
that you could see the layers
and make the comparison.
Here, let me show you
an example in practice
from the cotton collection
and then ultimately applied
to the Chad Gospels which
you can find online.
If you'd like to talk
to me afterwards,
I can send you the link where
you can try the entire archive
in the creative commons
for noncommercial use
at the internet archive
and at our website as well.
Take a look at the
photography on the left.
You see that because
the page is not flat,
that line at the
bottom is not a line.
And then once it
has been flattened
that becomes a line even
without that being a constraint.
And then if I show you an
off access view and toggle
between that off axis
view and the 3D view,
you can see the result
of the flattening
where the lines no longer
crawl up over the hill
because we've removed
the cockling.
And we've done it
digitally as a restoration.
So what you're seeing here
in my work are examples
of moving toward two things.
A complete framework for
digital restoration and the idea
that when you collect data, you
do more than make a facsimile.
What you're really interested
in is doing data collection.
A facsimile is a
byproduct of that.
And what you collect, say
you collect spectral imaging,
you can make an RGB
image from that.
But it's really the
spectral block,
the data that you're
interested in.
And we noticed this
trend, others did as well,
and we started talking about
beyond digitization, right,
working instead toward
the complete capture
of all the characteristics
of a proxy of an object.
All right, so this brings me to
the place where now we can talk
about complete unwrapping and
how you might approach something
in the invisible library
that looks like this.
OK, I wasn't introduced to
the Herculaneum collection
as a child nor did I
study at university.
History for me as a
computer scientist went back
to the 1950s.
Prior to that, I know
that a volcano did explode
in the bay of Naples.
And I have since of course
learned of the history.
Also I am now not ignorant
of it but, you know,
prior to my mid 2000s work,
I really didn't understand the
significance of the collection
from Herculaneum,
how significant these
manuscripts are
that they were discovered
at all.
That they remained after
a volcanic explosion is--
continues to be a
miracle, but also the fact
that all prior attempts
of physical unwrapping
generate this kind of damage.
This kind of damage
was finally halted
in the mid '80s and early '90s.
And that halting created
an opportunity for me just
as I came on the scene and the
opportunity was might there be a
way to do the unwrapping but
do it completely digitally
as a restoration rather than
assume that there is going
to be a physical
restoration process
that will allow this to happen.
So what I would like to show
you now is the first example
of me ever having tried that
on an example that I concocted.
OK, so having latched on to
this idea after the progression
of taking photographs and
then rubbing out wrinkles
and thinking through well,
how could you do
complete unwrapping.
I became familiar with imaging
technology that was being used
in medicine called
tomography which allowed you
to get a non-invasive
representation
of an object without opening it.
Three-dimensional all the way
through completely non-invasive
and the machines
that were available
at the time were medical.
So I made a proxy and I put
it in the machine at the--
in the basement of
the medical center
at the University of Kentucky.
I guess this is when everyone
else was worried about Y2K
and apparently I was not.
I was thinking about this.
I want to show you the
data in its entirety.
And in playing the data,
I want you to understand
that you're seeing
the axial view.
So I-- this is a piece of
canvas, it's rolled up,
and you're seeing that
canvas played as a movie edge
on with slice near you as
the first frame of that movie
and then slices as you go down
the roll as the movie plays.
So you can think about looking
at the end of a jelly roll
as a deli slice or
takes a piece off.
And that's kind of what
this data looks like.
Let me point out that there are
flashes on the canvas surface,
that's where the ink is.
So now that you've seen all
the data, you should be able
to tell me exactly
what it says, right?
Well, I couldn't for sure,
although we constructed
the object.
But once we used our
software to convert the data
from the way it came
from the scanner
into a complete
three-dimensional model
that we could then unfurl, just
a follow on example of rubbing
out wrinkles digitally.
This simulation allowed us
to be able to take that data
in its original form
and then put it
into a form that was readable.
And we dubbed this
process virtual unwrapping
because what we were doing
is unwrapping something
that was totally in 3D so
that it could be viewed again
as a two-dimensional image.
So now if I asked you what
do you see in this image,
you'd probably will say to
me a bunch of wavy lines
and I have no idea why you put
that on this original artifact.
And, yeah, I don't really
know why I wrote these
symbols either.
We got paint from a paint
store and made this proxy
so that we could understand how
the different paints would show
up in the imagery and if we
could do complete unwrapping.
So first example, we were able
to do a complete virtual
unwrapping from data that came
from a medical grade
tomography machine.
So I began working toward
Herculaneum by building a set
of successively more
challenging examples
and more interesting
text on those examples.
So here is some quotes from
the Iliad, "Two fates bear me
on to the day of death.
If I hold out here
and I lay siege
to Troy my journey home is
gone, but my glory never dies.
If I voyage back to
the fatherland I love,
my pride, my glory dies.
True, but the life that's
left me will be long."
So which do you want?
Well, I want them both.
And we spent a lot of
time in Jessamine County,
Kentucky trying to
add to the mix
of our examples carbonization.
Because I was worried
that carbonization fundamentally
would change things
in our ability to image
and then do unwrapping.
So we burned a number of these
examples until we figured
out how to do a vacuum
evacuated carbonization process.
And once we did, we were able to
run a complete example through--
this is the data from
one of those examples.
You see a lot more spirals,
you see the ink signal
on the surface of the papyrus
available as brighter spots,
and I'll show you an
example from that.
Since we scanned
this in sections
to get higher resolution
than was available
with medical grade scanners
focused on a single letter form,
and that single letter form
after the data is virtually
unwrapped looks like that.
Which inspired us to believe
that even through carbonization,
it's probably going to be
possible to see ink virtually
through this process
of virtual unwrapping.
So the framework was set and
the examples led me to believe
that process of scanning
using tomography,
probably x-ray based, but
really any kind of method
that would provide a
volume noninvasively
that captured the would
be the starting point.
Then this idea of segmentation,
which is a technical term
but it's the idea of finding
exactly where the layer is
that has writing and then
building a representation
around that, and then finally
the visualization of the result
by doing some kind
of unwrapping.
That became the pipeline
and the framework as early
as the mid 2000's for
what we want to do.
Now at this point sometimes
even I am a little confused
with the visualization of
what this data looks like.
So I would like to play
for you with audio,
a video that reenacts an
example so that you can get it
in your mind how
virtual unwrapping,
all of these steps,
actually works.
So, I'd like to play this now.
>> [Background music]
For thousands of years,
people have written on scrolls.
The scrolls can contain
historical records,
religious texts or
stories but many
of them have been
damaged over time.
For example, the En-Gedi scrolls
found near the Dead Sea were
in a building that
had burned down,
leaving this scrolls charred,
blackened and wrinkled.
But since the En-Gedi
scrolls, [inaudible] is opened.
How can we ever unlock
their contents?
There is in fact a way to read
scrolls without damaging them.
Imagine for a moment that a
baker has two types of dough.
The white dough represents
papyrus
and the red dough
represents ink.
Using the red dough, the
baker can create a pi symbol
on the white dough.
When the baker rolls
up the dough,
the pi symbol is
obviously no longer visible,
just like when papyrus
is rolled up.
Imagine the effects of time and
the environment are this oven,
baking the pastry
for a thousand years.
If the pastry was still soft,
the baker could simply unroll
it, but now it will break
and the information
will be lost.
The baker can however slice the
pasty and decode the message
from the traces of red
dough in each slice.
[ Music ]
Using sophisticated
technology, it is now possible
for the scrolls to
be digitally unrolled
without ever damaging
the scroll.
It's important for damaged
scrolls containing unique
ancient information
to both physically
and digitally preserved.
There are many ancient scrolls
but they're too damaged to read
that likely contain historically
enlightening information.
[ Music ]
Digitization and virtual
unwrapping allow us
to view this otherwise
inaccessible information.
[ Music ]
>> W. Brent Seales: I'd like
to specially acknowledge my son
Jessie who did the stop
motion animation as a project
in high school for that.
And that very last sequence
actually took five hours.
So does that help?
Does that help give
an intuition, yeah?
So the x-ray basically gives
you that without having
to do the actual slicing and
then the rest is all software.
So you can imagine how
excited I was to go
to the Institut de
France, make the approach
to Monsieur [foreign
language] who was
at the time the Secretary
of Perpetual of the Academy
of inscriptions et
belles-lettres
which is the academy that is
interested in the classics.
And seated, he granted
me the permission
with my collaborator
Daniel Delattre on the right
and the librarian Madame
Pastoureau on the right,
Daniel on the left, madam
pastor on the right, to go ahead
and scan two intact Herculaneum
scrolls that happened to be
in the collection at
the Institut of France.
They have six papyri from
the Herculaneum site.
They were required by
Napoleon, I believe--
in fact, I'm sure they were a
gift to Napoleon and they ended
up at the Institut of France.
Four of them were unwrapped.
And so they are in fragmentary
state but two were intact.
And we believe that the time
that by scanning them using
the best tomography of the day
and employing one these three
ideas, either a shape change
or perhaps some impurities
in the ink
or maybe even a multi-power
trick
where the x-ray would allow us
to use different power
settings to see contrast.
That using one of those
methods we would be able
to push the data completely
through a virtual
unwrapping pipeline.
And here you see Madame Fabienne
Keru who was the conserver
at the time handling
one of the scrolls.
And as an aside,
I want to tell you
that there were some
surprising things that occurred.
Well, they were surprising
to me at the time
because I I'm naïve.
I continue to be naïve.
And I thought that the hardest
part would be scanning the
scroll and then processing
the data.
But in fact, the hardest part
was protecting the scroll
so that we could scan it at all.
And we did that in the end
by creating a custom-fit case
that the scroll could be
put into because it needed
to be turned on its end to be
able to do a gentle pirouette
in the machine we had
to do the tomography.
So in order to create that case,
we scanned the exterior shape
of the scroll, made a model.
And because we didn't have
3D printing at the time,
we used an artist and the model
to cast a mold and that mold
from plaster to injected
polyurethane cut down
and then seated with the base.
And I'm not really proud
of that plywood base.
That was a prototype.
Ultimately-- And you see one of
the scrolls being formed fitted
into the shape that
we have there.
Ultimately, we created
these nifty cases
that the conservators
agreed were a great way
to not only transport but also
to be able to scan the scrolls.
And that process took six
months, back and forth
with our designers to Paris.
You can imagine how mortified I
was to have to go back to Paris,
each time, time and time again.
But our goal was to capture the
data and we did that in 2009.
This is the first radiograph of
Herculaneum scroll taken in--
at the Institut of
France on site.
And you'll the seam and the
radiograph going vertically.
The dark mass is
the scroll itself.
You see some of the texture of
the polyurethane as it was cast
and then you also see the
sophisticated closure mechanism
on the top of the container
which is the rubber band
holding the case together.
And here is what the
reconstructed slice looked like.
In fact, let me play this
for you as a movie, OK?
So this is no jelly
roll, all right?
This is no piece of dough
with one big pi waiting
for you to decipher it, OK?
I want you to take this in
and imagine me watching this
and imagining how my software
is going to find every one
of those layers automatically
and then do the virtual
unwrapping
of every one of those wraps.
And I'm sure right now
you're counting those, right?
You're counting how
many there are.
You can't tell exactly how
many there are, is that right?
Yeah, so we-- I freaked out, you
know, when-- because in the lab,
I was doing what's
on the lower right.
And the reality of
Herculaneum turned
out to be a lot more
tortured and challenging.
So we had the permission to
capture the data and we did.
We had what I believed
was a complete pipeline
for processing.
And I still believe
that pipeline is
exactly the right way
to approach the problem.
But we were not able at the time
in 2009 to segment the writing.
We were not able to
actually solve this problem.
So this brings me to the second
to the last part of my talk
where I'm going to talk about
the miracle that occurred.
You know, whenever we
get in trouble, right,
we expect an actual miracle to
occur and sometimes it does.
Sometimes it actually
does occur.
I love this quote from one
of my favorite radio programs
so "This American Life" with
Ira Glass who interviewed Teller
of "Penn & Teller",
the math magicians.
And Teller said, "You can't
look at a half-finished piece
of magic and know
whether it's good or not.
It has to be perfect before you
can evaluate whether it's good.
Either it looks like a
miracle or it's stupid."
You know, when I heard this
interview I thought actually
that characterizes
virtual unwrapping for me.
Because until the magic
occurred, I had been walking
around with a half-finished
trick, right?
We hadn't actually
completed the job.
We had data that was good
and we had a pipeline
that we believed would work,
but we hadn't pushed
everything all the way through
and so the trick was incomplete.
And when a trick is incomplete,
yeah, it looks stupid, OK?
So, how stupid did it look?
Well, I don't know how many
of you have Googled your
name and an adjective.
Just pick an adjective
and see what comes up.
If you Googled my
name and stymied,
and then you just hit
I'm Feeling Lucky,
you will find always and
forever probably this article
about how we weren't making
progress on being able
to read these scrolls
and that was true.
I wasn't able to
argue with that.
We were at an impasse until the
scroll from En-Gedi came my way
through no fault of my
own three years ago.
So I want to talk to you about
that scroll and why it was
so significant in helping
me complete the process
of virtual unwrapping.
In 2015, I met Pnina
Shor who is the head
of the Dead Sea Scrolls Project
at the Israel Antiquities
Authority.
And I met her there because
she had agreed to give me
on disc all of the data that
she had scanned of this scroll
that had been found in an
obscure location, the synagogue
on the shore of-- the
western shore of the Dead Sea
in a town named Ein Gedi and
had been found 50 years ago.
It was so badly damaged
that no one was able
to do any physical restoration.
She on her own accord with a
team scanned it and was unable
to make any sense of
the data and so agreed
to give me the data and let
my team take a crack at it.
I'll always be grateful to
Pnina for giving me a shot.
Because in the intervening time,
our software become a
lot more sophisticated.
We worked on the data
after she gave it to me.
And when she called for an
update, I really thought nothing
of going ahead and sending her
what our current progress was,
because as you see being
played in this movie,
our software had matured to
the place where we were dealing
with real data and we were
working daily on algorithms
that would help us
improve our ability
to take really sophisticated
data similar to Herculaneum data
and convert it into
completely unwrapped results.
So I'm going to go ahead
and forward past the rest
of this video and show you the
first image that we obtained
from the En-Gedi data.
And you could see in this
draft image some problems, OK.
On the lower right
you see the smearing
from our incorrect algorithm
at virtual unwrapping,
where we were creating some
smearing and some stretching.
But you can also
see very clearly
that there is obvious writing
and we were able to
see that writing.
And it turns out that Pnina
was able to read that writing.
I sent her the draft
and she said,
in an email to me afterwards,
Brent you won't believe what
groundbreaking discovery
you've made.
It's still all hush
hush because we want
to have a press conference
about it next week.
Which of course that was
when the image I
sent her was a draft.
And then she went on to say what
we've deciphered is the first
chapter of the book
of Leviticus.
After the Dead Sea Scrolls,
this is the earliest
bible ever found.
We already have it carbon dated
to the sixth century which turns
out to be a misreading,
it was third century CE.
And if that's not enough,
it's the first time a bible,
a Hebrew bible was been found
within an excavated synagogue.
So, it turns out that
it was a big discovery.
All they needed was the kernel
of what we had done up until
that point to be able to read
it because it was a known text.
And we did have the press
conference the next week.
So probably the first
time Skype ran for an hour
without actually
crashing, but you see me
on the background being
projecting from my office
at the University of Kentucky
with my team behind me
like little angels whispering
the right answers into my ears.
And you see in front
in Jerusalem Pnina Shor
on the left, Sefi Porath,
the original archeologist
who pulled it with his own
hands from the ground who came
out of retirement for that
moment, and on the right,
David Merkel who did the scan.
So let me tell you a
little bit about this scroll
and show you our result, and
then I'll complete my talk.
On the shore of the Dead Sea
in the early '70s was
discovered an ancient synagogue,
it was Byzantine.
And it is now a national park.
So you can go there and you
could see the synagogue.
It looks like this.
It's a beautiful mosaic
floor covered by a tent.
And in the floor was
found the holy ark.
And in that holy ark is
where the En-Gedi
Scroll was discovered.
It was so badly damaged
that it was not possible
to identify anything about it.
But Sefi, the archeologist
always believed
that it could hold significance.
And so he told me personally
when we did the press
conference,
he actually thanked me.
He said, "I didn't know if in
my lifetime we would ever solve
this mystery."
But let me show you
now the second video
that actually has audio
which will explain
to you the complete unwrapping
process on the En-Gedi Scroll.
>> Virtual unwrapping begins
by acquiring a three-dimensional
volumetric scan
of the damaged manuscript.
This scan produces a set
of cross-sectional images
that show the internal
structure of the scroll.
When viewed as a 3D object,
one can clearly see the
individual layers of the scroll
that any text on the surface
of those layers is
obscured from view.
In order for a readable version
of the scroll to be produced,
these images must be passed
through our virtual
unwrapping pipeline.
First, we capture the 3D shape
of the layers of the scroll
in a process called
segmentation.
On the left side of the
screen, the software moves
through the scroll image
by image tracing the shape
of a single scroll back.
On the right, we see the 3D
model that this produces.
Next, we extract the
ink from the data
in a process called texturing.
Using the 3D shape
generated by segmentation,
our software makes another pass
through the scroll this time
looking for very bright pixels.
Bright pixels indicate
regions of dense material.
In this case, inks
made with iron or lead.
We now have a single
wrap of the scroll
with the text shown
clearly on its surface.
However, because the surface
is curved, it's difficult
to read all of the text
from one viewpoint.
The flattening stage
of a pipeline converts
this textured 3D surface
into a flat plane so that the
text can be more easily read.
To produce the best results,
these three steps must be
performed on one small section
of the scroll at a time.
As a result, we end up
with several texture images
that must be merged together.
This merging process creates
a single consolidated image
that shows the full text.
Using this pipeline, we have
restored and revealed the text
of five complete wraps
of the En-Gedi Scroll.
The two distinct columns
of Hebrew writing
revealed the scroll
to be the book of Leviticus.
This marks the En-Gedi
Scroll as the earliest copy
of Pentateuchal book
ever found in a holy ark,
a significant discovery
in biblical archeology.
>> W. Brent Seales: And I'd
like to give special thanks
to my daughter who
did the voiceover.
The price was right on that deal
and also, you know, I miss her
and so I'm reminded of her
when I hear her voice now.
Somebody after my talk asked
me is your daughter OK thinking
that maybe, you know,
is it memorial
and I said, no, no she's fine.
She just lives in
California, you know?
When they go off to
California, you know.
So here you go.
This was our complete
unwrapping, five wraps,
fully readable text,
known text pushed
through the entire pipeline.
The miracle occurred and now
the magic trick actually looks
like-- it doesn't look
stupid anymore, right?
Because it's complete
and we pushed
through the entire pipeline
and refined each section
of the pipeline to
the point we're now--
we think that we're
actually ready not only
to solve the technical
part of virtual unwrapping
for most materials but also
to provide those materials
as scholarship to scholars
who can then use the quality
of our result to do, for
example, biblical scholarship.
And let me tell you that we did
that with the En-Gedi Scroll
and this is why it's
significant.
This dating in terms of
biblical scholarship occurs
between the period of the Dead
Sea Scroll second temple period,
scroll from En-Gedi at 300 C.E.
and really the next witness
which was 6 or 700 years later
where Cairo Genizah
gives evidence
of the Masoretic
text being settled.
And so there was some
question about how early
that Masorete text goes back
into the unsettled first
and second temple period.
The scholars told
me all of this.
I am not a biblical scholar.
But look at what
we were able to do.
We wrote a biblical
scholarship paper together.
Michael Segal and Emanuel
Tov, giants in the field
of this area, with
me as a co-author,
and the reason why they
allowed me to be a co-author is
because the image they
used does not exist.
The image that they used
is the result of a piece
of software that
we wrote, right?
It's completely generated
virtually.
It's not a photograph, right?
It's not the result of
physically unwrapping anything.
And so we were able to take that
image at a quality high enough
that a paleographer and a
biblical scholar could print it
out and create basically
what they would do
with the photograph
of an open fragment.
And then because they could
see all the letter forms,
do paleography, a
complete transcription
and biblical scholarship.
So from the pages
to the flattening
to the complete 3D unwrapping,
you see the progression
of this work to the place
where now I could talk
about my rousing conclusion,
which is what I'm calling
reference amplification
and the use of what we all
have heard off in the news
in the last several
years machine learning.
Now I want to tell you why
machine learning is really
important because
conventional wisdom,
which is almost always
wrong, which somehow I seem
to always adopt, right,
tells us a few things.
It tells us that well, that
scroll was probably a fluke,
Seales got lucky, that's for
sure, there is obviously metal
in the ink so that you could
see the signal really easily.
So this probably isn't
going to have any relevance
to Herculaneum because we
all know there is no metal
in the ink et cetera, et
cetera, conventional wisdom.
Also conventional wisdom
also tells us, you know,
why couldn't it be Genesis?
You know, why did it have
to be-- just as an aside.
Well, it's true that there
was metal in the ink.
We don't know exactly
the composition
because we analyzed
everything non-invasively.
But if you take a look at
the signal in this figure
from our paper, you
can see really bright
where the Hebrew letters are,
meaning, you know, very dense,
lots denser, right, than
the surrounding material,
the animal skin.
So how do I answer that
conventional wisdom?
I mean for the longest time, I
also believed this statement.
Carbon ink is invisible
in tomography.
I don't know.
Those of you who know
may have heard that.
But, you know, we
were doing some work
at a high energy
physics facility.
And whenever I hang
around bright people,
they raise me up, right?
Does that work for you?
I don't know if I actually get
smarter or I just act smarter
because they're all
smart, right?
But, you know, we
started working together
and I started thinking, really?
Carbon is not visible?
I mean, why is everything
else visible in some way,
shape or form but carbon ink
on carbon is not visible?
I just don't get that.
So we ran some numbers and
I started to think about it.
So I went back to my
kitchen and we started
to do some experiments.
I got out the iron gall, the--
not the iron gall,
the carbon black.
I ruined a few utensils.
And I made some patterns.
And, you know, I really gobbed
it on because I thought,
you know, I'm going to
see what's going on here.
And then we got a little
bit more sophisticated
with some test patterns
in papyrus
and some other real materials.
And we started to do a
systematic study instead
of just relying on
conventional wisdom.
And guess what we found.
Of course carbon is
visible in tomography, OK?
There it is.
Carbon on paper.
On the left, you
see that the carbon
and the paper looked
very similar in density
because they're about
the same at that energy.
And on the right, you see that
the paper is much brighter
than the carbon at
the other energy.
And we can actually look at
this, different energies,
different contrasts
between the two.
Then I did an experiment
with some students,
where I asked them to write up
some-- make up some proxies.
And this student decided--
Allie decided that she was going
to put carbon black on papyrus--
excuse me, on parchment.
We hadn't really
done that before,
carbon black on parchment.
Most of the ink we've seen
on parchment are in Gaul.
So she did that.
And she scanned it.
And she ran our tool
to unwrap it virtually
and then she showed
us the result.
And there was a delta in Carbon.
Why can you see it?
It's supposed to be invisible
in tomography, right?
Well, we did a little
bit of work and found
that what's actually going on
is that you can see the carbon
but you have to have the
right resolution and you have
to tease it out from
the underlying pattern
of whatever the structure
is that the carbon is on.
If it's on papyrus, you end
up with this fiber pattern
that is really hard
for your eye, you know,
to see the difference when
there's carbon on there.
Papyrus has that fiber
structure but parchment doesn't.
And the reason that Allie was
able to see the carbon ink
on the parchment was because
the parchment was really smooth,
right, and then you get
that carbon on there,
you could see it straight away.
So we actually started
to do some experiments
where we took carbon
ink on papyrus
and we developed a framework
for being able to tease
out what you can't
see with the eye.
And we used machine
learning to do it.
And I'll be showing you one
of the first examples
and then I'll finish.
On the left you see
a test pattern.
Six, 5, 4 at the top is six
codes, five codes, four codes.
Real scientific, right?
On the right, you see the
tomography to give you a sense
of what you can see with the
naked eye and what you can't.
On the left column with six
coats, you can actually see,
there's enough density
that the carbon is visible.
But as you get down to only four
coats, you can't see it anymore.
What we hypothesized was that we
could train a machine learning
system to understand the
patterns in the carbon
that you wouldn't be able
to see with the naked eye.
And here is the result from
that first experiment showing
that the machine learning system
is able to amplify the ink
in that fourth column even
though you can't pick it
out with you eye.
And here is the triptych
that shows you on the left,
the original pattern for
column 4, the training sample,
and then the amplification
on the right.
So I call this reference
amplified computed tomography.
And the idea is that you
use a set of examples
of what the ink looks like
in tomography to train
up a huge scale machine
learning system
that it can detect
subtleties in the tomography
that you cannot see
with your human eye.
Reference amplified computed
tomography is actually now
for me a wholesale change
in the way I believe
tomography should be treated.
In fact tomography is not
for you, it's not for me,
it's first for the
machine learning system
which amplifies what you want
and then the result
is for you and for me.
Here is an example
on a real fragment.
The photograph on the left
shows you a lunate sigma
from Herculaneum.
It's real carbon ink
from Herculaneum.
And on the right you see a
volume render of the tomography.
And I don't think that you
can probably see the ink.
Because we looked at that
and thought we'd failed, OK?
Now this is what
the data looks like.
And as the bar goes
left to right,
you see on the right what
the slices look like.
It turns out you can't
really see the ink
in the slices either.
You know, you'd like to be
able to see those bright spots
like I showed you in
my early examples,
but we just don't see that.
But what we do have are
lots and lots of fragments
from Herculaneum that are open
and they have visible
text on the front.
That visible text we believe can
form for us a reference library.
The reference library from
machine learning so that
when we scan, what those
look like in tomography,
we can leverage that to
build a neural network,
evolutional neural network,
to be able to amplify what
the tomography actually looks
like for ink.
Here's the example.
The lunate sigma
with the green box,
the green box being right here,
shows the size of the test
that we give to the
machine learning system.
That little section we
say, take a look at that
in the tomography, that's
where we think ink is.
And then we also move that
box around and say, OK,
here's a spot where
there's no ink.
That's where we know
there's no ink.
So, we have a training
system that says here's ink,
here's no ink, learn
the difference.
We actually do that on
a tenfold experiment
which is the lower right.
Ten sections, we drop one out,
train on nine, classify
onto one.
Because we don't have a
full reference library yet,
this is the tenfold experiment
that simulates being able
to have a bigger
reference library.
Let me show you the result
of having done the machine
learning to classify ink.
Here's the original data
and here is an amplification
of the ink using the
neural network we trained,
pure carbon ink from
Herculaneum.
OK, now as we turn on the
result of the classification,
you see that the classifier
at about 80% precision
recall is able
to get a pretty good
estimate of that lunate sigma.
OK. So here is the triptych that
shows you the reference photo,
the amplified result, and
the original tomography.
With machine learning being
the instrument that we use
to defeat conventional
wisdom and pull
out very subtle signals
in tomography.
So I have numbers here
and if you'd like to talk
with me afterwards
about precision recall,
the way we structured
the network,
how many test examples
there were,
convolutional neural networks
on 3D, we can talk about that.
But what I want to do is
to show you a quick example
that I worked up
over a larger text
to show you what it would
look like if we could get 80%
or 90% precision recall on
carbon ink in a real text.
So there is the text
in its glory
and here is what it
would look like at 50%,
60%, 70%, 80% and 90%.
Now if we push that back into
a volume and I've blown this
up a little bit so
that you can see
that the fibers really do
confuse the ability to read.
Fifty, 60, 70, 80 and 90%
and then the reference image.
We believe that if the
weakened take open fragments
and get a neural network to
do detection at between 80%
and 90% precision recall, we'll
be able to read any kind of ink.
Are there plenty of things
that fall under this category?
Well, yes.
Let me say that the invisible
library is extensive.
It includes all kinds
of surprising thing
that you may never
have heard of even
if you're not a computer
scientist.
For example, there's a
manuscript from Jack London
in the Huntington Library in
California that was badly burned
in the earthquake,
1906 in San Francisco.
And now it's carbonized but it
was probably written in an ink,
in a variety of inks that
would respond really well.
For example, the
Franklin papers.
Manuscript that was
discovered on a cadaver
from the Franklin Expedition,
search for the Northwest
Passage, all sailors lost.
Cartonnage.
Lord knows there's
cartonnage everywhere, right?
And inside that cartonnage there
may be original manuscripts.
And of course the
enigmatic Herculaneum scrolls
which I hope someday
we will read.
So now to my epilogue
and to conclude my talk,
I'd like to say that
collaboration is the key
and the glue to making
all of this work.
I've talked about the textual
and I've talked about
the technical.
So now let me just say a
word about collaboration.
Our most recent project
at the Morgan Library was
to take tomography and
apply it to this manuscript,
the Morgan M910 which is an
early Egyptian Coptic Acts
of the Apostles.
It's unable to be
opened or restored
because the damage
inside is profound.
And let me say that in
approaching this project,
you see the manuscript
in the center.
I'd like to identify the people
involved to make this happen.
We have conservators
led by Maria Fredericks.
We have a manuscript
scholar Paul Dilley
at the University of Iowa.
We have engineers
from Microphotonics
who graciously donated
to my team the scan time
that we needed to
do the project.
We have a research specialist
Christy Chapman who's
in the audience today.
We have a computer scientist too
as my staff member, Seth Parker.
And we have of course
the media who share
with us the ability
to tell stories.
And let me tell you that
telling stories is a big part
of what it's about.
It was the beginning of my
motivation because the narrative
of the story is the thing
they were actually all
after in the end, isn't it?
So if I put that on a diagram
for you and in leave it
in French so that we all
understand how collaborative we
have to be.
Yes, I still try
to learn French.
I would like to conclude my talk
by thanking you for listening
and asking if you would like
to ask me any questions.
Thank you very much.
[ Applause ]
>> Fenella France: So we're
going to take some questions.
I'm going to ask
Brent and that he
or I will repeat your questions
so that the external viewers
can actually hear the question.
So we will rephrase it to
how he would like to answer.
>> Even though we're in
Washington, I'd be happy
to rephrase the question
into one
that I can answer so, questions.
>> So, just to clarify a
point that I'm not sure
if I completely understood.
The machine learning--
the computer is learning to
identify ink from non-ink or--
>> W. Brent Seales: Yup.
>> -- or character
from the non--
>> W. Brent Seales:
So the question is what exactly
is the machine learning doing
in this case.
And my answer is your
instinct which is, yes,
the machine learning is helping
us classify ink from no ink.
Now other applications
of machine learning help you
recognize higher level things.
There is work being done to
recognize the letter forms,
do OCR, that kind of thing.
What we're interested
in tomography is purely
amplifying a very weak signal.
And so we're doing
machine learning
at the level of ink, no ink.
Yes.
>> Have you and the machine
[inaudible] machine learning
exercise, have you just tried
filtering the background given
that the paper or the substrate
has particular frequencies
to as-- presume what the
machines is learning is doing,
is filtering out certain kinds
of features, characteristic
of the background but you
could directly filter--
>> W. Brent Seales:
Yeah, absolutely.
>> Have you tried to--
>> W. Brent Seales: So let
me rephrase the question.
The question is why don't you
just use conventional techniques
and filter out the background
which would then give you
the foreground signal.
And what we're doing is
using machine learning
as the mechanism for
doing exactly that.
And the reason why we need
so sophisticated a tool is
because the variation
in the fiber structure
of the papyrus masks
the weak signal
of the carbon fairly profoundly.
And we haven't found a
direct way to be able
to capture that signal.
But through millions of examples
in a supervised learning
setting, we have been able
to capture that signal.
And this is something
that other approaches
in machine learning
have discovered.
Ultimately, we may
discover a pathway
for doing what you're
suggesting right now.
The machine learning
framework is our way
to be able to capture that.
Yes.
>> Is it worth considering
to further
or not the many other
layers of En-Gedi?
>> W. Brent Seales:
The question is so what
about En-Gedi is there
are continued work going
on for the other
layers of the scroll.
We worked on or other
artifacts from En-Gedi.
The scroll we worked on
was completely unwrapped.
So five wraps, that
was all we had and all
of the visible text is there.
And that data is
publicly available.
So anyone else who
would love to jump in
and refine our results
is welcome to do it.
They can download
all of the data.
There is other information.
There are other artifacts
from En-Gedi.
And to my knowledge, no one
yet has approached the rest
of the set of artifacts
from En-Gedi
and I'm actually
dying to do that.
One of the barriers,
you might ask,
if I might pose a question
for you is funding.
Because these things are very
old and they're going to be fine
if we don't do anything
so you need funding
to be able to do something.
So when funding appears,
then this work I
believe can continue.
Any other questions?
>> I was wondering if
you could comment on some
of the challenges with the
Herculaneum scroll and the kind
of advances in modern technology
that you think are going
to help you eventually
unwrap that.
It's seems like resolution
I think may be one
of those issues.
>> W. Brent Seales: Yes.
So the question is can I comment
on the Herculaneum question
and what may have changed
or what's necessary in order
to advance that and your
intuition is exactly right.
Resolution is crucial
and we've discovered
that there is a sweet
spot in the resolution
that we think we need to be
able to capture everything.
And too high a resolution
is not good
and to low a resolution
is not good.
We are now capable of being able
to capture the correct
resolution.
But another thing about
Herculaneum is access.
Because despite the fact
that I negotiated access more
than a decade ago to the
collection in France,
it's still extremely difficult
at the four institutions
that hold this material to work
on this material
collaboratively.
For obvious reasons, it's
incredibly sensitive material,
difficult to handle,
there's always risk, right?
So-- And funding
is always an issue
and people are committed
to other projects.
So, all those things play a
role advancing the agenda.
Yes?
>> Can you-- Do you
have any confidence
that the technology can
differentiate a carbon-based ink
from an ink that contains metal
such as copper, iron, or lead,
or others that may be
is sort of possible?
>> W. Brent Seales: So
do I have any confidence
that we can differentiate among
different kinds of inks in,
for example, tomography?
And my answer is, I believe that
reference amplified tomography,
using a classifier like
machine learning will allow us
to create libraries for
everything that we'd like to see
and then we will be able
to classify the difference
between those inks, yes.
And I believe that
there are going
to be other tomography-based
applications
for doing exactly
this kind of thing.
You're going to see an explosion
of it in the next five years.
Yes, you had a question.
>> You were talking
about collaborations
and I remember reading that
there was a, I don't know,
if it's similar but
an unwrapping
of Herculaneum scrolls
project at the Vatican.
Were you involved with that too?
I remember there was an
Italian scientist and actually,
I'm not sure if it was
Herculaneum scrolls
but it was a carbonized scroll
that they were trying to unwrap.
>> W. Brent Seales: Yes.
So the question is what about
other work, perhaps the Vatican
or an Italian team
working on Herculaneum.
So my answer is that, yes, I'm
actually gratified by the fact
that after the two decades
of work I've been building this
miracle or magic trick, right,
others are beginning to
work in the same area.
And there are advances
being made by other teams,
one at Cardiff and
some Italian groups
who have access to the material.
So what I'm hoping is that
as we all advance the field
of reading the invisible
library, right,
what we will do together
is share good information
and build a standard for
how we report things.
I didn't talk about this in
my talk, but let me just say,
when I report to you
something inside an artifact
that you cannot see and you will
never open, who do you trust?
How do you know that the En-Gedi
text is actually Leviticus?
I mean I scanned it, right?
I say that I-- my team
developed it from software.
It's actually really, really
important for the community
and peer review to develop
a standard for the review
of these claims so that we
can be open about what is
and is not a letter form,
what is and is not a text.
OK. Very, very important.
And if you go and read
my paper in science
on the En-Gedi Scroll, you
will see at the end of my paper
that I devote an entire section
and then I released all the
information from our work
for third-party review, to try
to lead by example and say,
you know, it's important, so.
>> I have one offside
question [inaudible], which is,
how is the technology evolving?
>> W. Brent Seales:
Offside question,
how is the technology evolving?
Thank you for that question
and for staying with us.
Yes. So there are three
different ways I see evolution
happening at the scanning level.
We see things that
used to be done only
in high energy physics
facilities,
like phase contrast
tomography now available
in desktop lab units, making
them available to hospitals
as well as libraries
and museums.
So the evolution of scanning
technology continues.
Terahertz imaging is also
another nascent technology
that's moving forward.
Second thing is that we see
machine learning as a tool.
And even though we might
not know how to tease
out a classifier in
a conventional way,
we can use machine learning
and millions of examples
and still make progress without
fully understanding necessarily
exactly what's going on.
And the third piece of
progress is that collaborations
with museums and libraries
are becoming more open
to allowing scientists
to explore the corners
of their collection because
they realized there's value
in the invisible library.
And so I see those as advances.
>> What is just sort of a
ballpark figure of the cost
of imaging a codex this way?
>> W. Brent Seales:
So the class--
The question is what's
the ballpark figure
of imaging a codex?
And that depends on how
you arrange the logistics
of doing that.
If you're willing to take
your codex to a facility
where they are setup, the
cost is the scan time,
which can be under $1000.
Now, if you need to transport
equipment, set it up on site,
hire staff to do
the collaboration
and then the scanning
on site, 10 times that.
Now if you want to consider the
cost for the post processing,
I'm not sure what I would
charge you right now
to use my software
but-- I'm just joking.
We're going to develop
our software to be
in the open source domain so
that anyone will be able to use
that software and so we
want that to be free.
So we can inspire the invisible
library going to scale.
That's what we would
like to see happen.
So we believe that right
now the cost is too high
but it will be driven
down by the fact
that these scanners will
become more available
and we will have software that's
available to do this work.
Last question?
>> Fenella France:
Last question.
[ Inaudible ]
>> W. Brent Seales: Yeah.
In fact, we're-- so the question
is taking differentiation
of inks a step further
and looking
at layers, mixtures, chemistry.
So two things tangled there,
one is x-ray fluorescence, XRF,
gives a complete
characterization
of the elemental composition.
And people-- And instrumentation
is moving toward being able
to do more of that in 3D
almost tomographically.
Using reference amplification,
we believe we can get
at a better discrimination
of things but probably not
to the degree that you can
use something like XRF.
But we have characterized ink
as sort of thin film analysis.
That was actually
our inspiration.
I come from Lexington, Kentucky
with long heritage with Lexmark,
a printer division that
originally started with IBM,
and that printer
division does nothing
but characterize how
ink goes on to paper.
And so you can see some of
that inspiration as well
in our thinking about
ink on papyrus
as thin films being deposited
and understanding it that way.
Thank you very much
for that question.
And thank you very much
for your kind attention.
[ Applause ]
>> This has been a presentation
of the Library of Congress.
Visit us at loc.gov.