A discussion of the future of AI (artificial intelligence) programs in lens design. The weaknesses of deep learning AI programs compared to knowledge based AI programs are illustrated with many examples from lens design and also chess.
1. The future of AI in Optical Design
Dave Shafer
David Shafer Optical Design
Fairfield, CT 06824
203-259-1431
shaferlens@sbcglobal.net
2. “Deep learning” neural network AI programs are being used in facial
recognition and other pattern recognition activities and have had very much
better success as a chess playing program than old style knowledge based AI
programs. Because of this I think it is very likely that in the future automated lens
design will be handled by neural network AI programs and knowledge based
programs AI may become obsolete and abandoned.
But oddly this will make human lens designers more valuable, at least in the
near term, and I have many design examples that will show this. If the only input
to a neural network program are the rules of chess, or Snell's law and ray tracing
codes for lens design as well as the specs on a desired design, then that same
input becomes very important in a way that it is not in the old knowledge based
programs. Some examples here will show this – how an incomplete rule set as
well as a needlessly restrictive set can cause deep learning AI programs to give
odd and/or disappointing results. They may always need human guidance.
3. Before looking at AI in lens design let us first
consider AI in chess. In a different universe
Homer Simpson might be a math wizard. But
in our universe he is an idiot. How would he
do playing chess against a computer program,
specifically a deep learning AI one?
4. Imagine a robot connected to a powerful deep learning AI chess playing
program that we have just created.
5. How our chess playing robot does against Homer
Simpson will depend strongly on the set of rules of the
game given to it and the possibilities for play that
follow from that. This is also true in lens design.
The key to the outcome is how complete the rule set
is. For example, Homer would be used to a very
important rule, from having played other games, that
the computer program has to be explicitly told.
That rule is that the chess
players can only move their
own pieces, not also those
of their opponent. So very
obvious that it might not
be in the rule book given to
the program.
6. Here is a classic and brilliant chess
puzzle from 1974. The problem is for
White to checkmate in just one move.
A quick study will show that this is
impossible. A long study will show the
same thing. An AI deep learning
computer program will almost certainly
be stumped by this. The solution
depends on a very minor chess rule that
you would probably never think of
applying to this situation.
The board and pieces as shown here
are a completely legitimate arrangement
but it is shown in a very deceptive way,
which is the key to the answer.
7. In setting up the initial board it is
required that the bottom right hand
square space be white.
right wrong
Board is
rotated 90
degrees from
correct position
8. Here is the board shown
earlier. The lower right hand
square is black, not white, so
this view is very misleading, but
still legitimate.
Here is the “correct” (i.e., not
misleading) view, with a 90 degree
rotation. Now the bottom right
hand square is white.
9. With this view everything is different. The pawn can be advanced by one square
and promoted to a knight, which gives an immediate checkmate in just one move.
The puzzle
depends on
showing the
board mid-
game, when
you would
probably not
notice the 90
degree board
orientation.
10. But a human, even a cartoon human,
working on this puzzle might decide that
there is something fishy going on here,
where it is easy to show that there is no
solution possible and yet we are told to
find a solution. A computer would not
think that way or be suspicious unless it
had been explicitly told to check the rule
book when all else fails.
Most test situations are not a contest
between you and the test material (here
a chess puzzle). They are instead a
contest between you and the person
who created the test – here a very
sneaky person.
Most (or maybe all) AI programs are
unaware of this particular human
element. A sneaky diabolical puzzle
creator is a higher level concept outside
of the AI program’s rigid rule book
orientation.
11. Here is a simple Turing test, based on human psychology, to
distinguish between a computer and a person. It is based on an
awareness of how people think. The next slide shows a whole
bunch of computer generated 5 digit random numbers. But
suppose instead these random numbers were all generated by
a person. How can you easily and quickly tell which it is?
Answer – you quickly scan down through the columns of
numbers looking for some that are mostly zeros, like 90003 or
00006, or repeated numbers, like 77773. These stand out
visually from the background and are easy to spot. If you see
none or very very few of these then you know the list was
generated by a person. Few people will include a number like
90003 or 10001 in a list they generate because they don’t think
these numbers “look random”. So here is a Turing test that no
AI program would ever “think” of generating, because it is
based on human psychology. It is a knowledge based test idea.
Turing
14. Homer has tunnel vision and cannot think
of anything outside of his present situation
and that limits what he can achieve.
Computers have a very hard time “thinking”
outside the box if they are too constrained
by a set of rules or alternately are working
with an incomplete set of rules. Let us
move now over to lens design and see this.
There are many situations in modern lens design programs where how the merit
function is defined and how the initial configuration data is specified will
needlessly restrict what designs can be found by the optimization program. This
is where a human lens designer or a very good knowledge based AI program can
find some designs that a deep learning program can never find. Let’s look at how
a deep learning AI program will fail when it lacks at least some basic knowledge.
15. A monochromatic design with a 90 degree field, f/1.0, 5.0 mm f.l. Many
aspherics. Diffraction-limited over the field at .55u Well corrected for distortion
and no vignetting. The key to this kind of design is many aspherics, very long
length compared to focal length and – most important of all – an intermediate
image.
There is no equivalent high performance design without an intermediate
image that can cover a 90 degree field at f/1.0 - that is a key part of this amazing
design type.
16. Suppose you gave an AI program the task of finding a good design with the same
specs, of 90 degree field and f/1.0 with no vignetting, and some focal length value. You
might choose 10 mm focal length. No matter how long the AI program searches it can
never find a design of this type. Why? Because with an intermediate image the focal
length sign changes and you should have asked for a -10 mm focal length. Few
designers would think of that and a deep learning AI program certainly would not. The
best plan is to correct for the square of the focal length or the absolute value, to cover
all + or – possibilities. Then AI could find an intermediate image design.
17. Spherical aberration between the
doublets redistributes the energy
Collimates the
aberrated rays
Intermediate focus gives a design with much
better energy capture and intensity uniformity
Gaussian beam
to flat top design
18. Spherical aberration between the
doublets redistributes the energy
Collimates the
aberrated rays
In controlling the output ray distribution during optimization, the signs of the ray heights
are flipped in the bottom design compared to the top design, because of the intermediate
image. An AI program will never find the very good bottom design if given the ray distribution
targets of the top design. All the target value signs would need to be switched.
19. Here is a good
design from 1934 by
Schmidt, where the
usual Schmidt
aspheric plate is
replaced by an
afocal triplet
without an aspheric
Reference design, f/1.0, 10 degree full field, for comparisons
BK7 lenses
Spherical
mirror
20. It turns out that there is a
variety of three lens
solutions, all the same
glass, and all spherical
surfaces. This one has
much better performance
than our reference design.
Suppose you give an AI
program the task of looking
for other designs with
three front lenses and a
spherical mirror, looking for
the best solution.
21. The best design of all,
with three lenses, is this
one here – where 2 of the
3 lenses are seen in double
pass. It is related to our
original reference design,
pioneered by Schmidt.
Despite the similarity to our reference design (positive, negative,
positive lenses, no meniscus lenses, no lens thickness sensitivity), no
“automatic design” program is ever going to find this new solution – of
using the double-pass idea - from the reference design starting point.
22. From the computer’s
point of view, and also
that of the AI program,
this is a design with 5
lenses, not 3 – because
that is what the light
rays see. But physically
it is just 3 lenses. An AI
program will never find
this high performance
solution if it only
understands 3 lenses to
mean what the light rays
see.
23. A beamspliiter used in converging light has aberrations – spherical aberration, coma, and
astigmatism – which degrade the transmitted image quality. Shenker and Rayces both
independently came up with this very clever insight – if you just choose the coordinate system
differently you will see that the aberrated rays are just a small off-center piece of a very fast
speed on-axis ray cone, which only has spherical aberration – in that coordinate system.
Very fast speed
axial ray cone
Piece of the
axial cone
Design with
coordinate shift
24. That same spherical
aberration can be
corrected by adding a very
weak power spherical
mirror either before or
after the beamsplitter. No
AI program would find this
clever solution. Instead it
probably would do what
human designers have
done in published designs
– added expensive cylinder,
wedge and asphericity to
the beamsplitter element
itself to fix the aberrations.
Spherical
mirror
Spherical
mirror
25. The Fulcher design is a
simple extremely high
performance focusing lens
with just 4 weak power
spherical lenses. Here it is
as a 100 mm focal length
design at f/0.7 and a
wavefront error of .006
waves r.m.s. at .55u
It turns out that there are several other good solutions with 4 lenses, including
an alternate all positive lens design and several where one of the 4 lenses is
negative. The Fulcher design is the best one by far and is not easy to find.
4 spherical lenses
with n = 1.6
26. To know about Fulcher’s OSA article from
the 1950s you would need to have a very
good memory (like this guy here) or have
read the very few articles since then that
refer to it.
A knowledge based AI program might
have a huge amount of design literature
and patents loaded into it, but deep
learning AI based on neural networks
seems to be the wave of the future – not
knowledge based AI. There is no guarantee
that a deep learning AI program would find
the superior Fulcher solution among the
other alternate design solutions.
27. A catadioptric immersion
design for lithography, with
two intermediate images.
An alternate catadioptric
immersion design for
lithography, with two
intermediate images
Both designs are state-of-
the-art in performance
28. It is highly unlikely that an AI design
program could find either of these
designs because of the unusual
configuration of lenses and mirrors and
the extremely high performance levels.
But if AI could find the bottom design,
with equivalent performance to the top
design, it would not “realize” that it is a
worthless design – and neither would
some designers. The reason? – The top
design has two reflections and the
bottom design has three, so the right/left
image orientation is reversed in the two
designs. Only the top design is
compatible with the already existing very
expensive lithographic circuit masks.
29. A conventional design configuration like this for a complex lithography lens
is ideal for a deep learning AI program, or a knowledge based program.
Lenses can be automatically added until a good solution is found. But designs
that combine lenses and mirrors, like the previous slide, are much more
complex and have to deal with ray/mirror clearances and interferences, fold
angles, odd geometries, etc. that would be much harder for an AI program to
“think” of and deal with.
30. Fresnel lenses that deviate light by large
angles, as in the outer parts of a lighthouse
lens, use either tiny mirrors or catadioptric
prism elements. They result in large ray
deviation angles with no color effects.
Sometimes the
simplest ideas would be
the hardest for AI to
discover. Here it is the
transitioning change
between the center and
outer parts to avoid
excessive color, as well
as what is shown next.
31. Dispersion effects in Fresnel elements
Lots of color some color little or no color effects
In principle you could make low or zero
dispersion sheets of plastic Fresnel prisms
No reflection One reflection
Two reflections
32. Cognitive scientist Roger
Shepard’s 1990 table top
illusion. The table tops
are exactly the same size
and shape! Yes they are!
It is an illusion to think that an
AI program that only uses deep
learning with no knowledge
based input can match human
designers.
33. So it looks like the future of AI
in lens design is one where
human knowledge and insights
will greatly expand the useful
terrain that AI programs explore
– by people directing AI away
from known dead ends and
towards regions not covered by
the “rule book” – such as
allowing a double pass of light
through a lens.Homer will probably not, however, be
one of the people best suited for this
human collaboration with AI programs.
34. One thing is certain - in the future AI will substantially reduce the
employment options in many fields. Even today young people are
becoming very concerned about that, as my next slide shows.