Title: How to Build an 'Artificial Scientist'
Author: Quanta Magazine

Transcript:
physics experiments is one of the most
important way how we can learn something
new from the
universe but the question is did humans
already come up with the best
experiments that we could think about or
are there many experimental techniques
out there that could be built which are
potentially too unintuitive for
humans the possibility of the different
combinations of experiments that one
could theoretically build in a
laboratory is just absolutely amazingly
large and this is where artificial
intelligence comes in how can we get new
scientific understanding using
artificial
intelligence when I started my group
here at Max prank Institute I've decided
to choose the artificial scientist lab
to force myself to think every day about
what is actually the Big Goal we want to
know how to build a artificial
scientist when humans design experiments
they rely on their intuition they rely
on what they have seen in the past let
me give an example in my own work we
wanted to build a new experiment we
failed we couldn't find the experimental
setup and I thought maybe the human
intuition is actually somehow in our way
of finding the right way to build an
experiment so I wrote computer program
that basically just has experimental
equipment that we have in the laboratory
in its own virtual toolbox and and then
it can uh Shuffle around experimental
components and compute its
results so I started this algorithm and
I came back in the
morning and I actually saw exactly the
solution that we as humans couldn't find
for several
[Music]
months the combination of experimental
equipment so it's definitely not that we
could not have thought about it the
surprising thing is that the way how the
machine assembled the components was
highly asymmetric it put in this
unintuitive element which we as a human
just kind of couldn't
[Music]
grasp let's imagine this Bas of all
possible experiments each tiny point in
this diagram would be a different
experiment humans have spent the last
hundreds of years to find some of those
points for instance by building better
microscopes better telescopes better
high energy physics experiments we want
to use artificial intelligence to find
new experiments that we have not
discovered
yet in my group we have a number of
different projects several of them are
building new AI tools for Designing new
experiments what we are using is
scientific knowledge this has all been
published in papers and we use this and
en code this in our algorithms
[Music]
pyto is an algorithm that can design new
Quantum experiments the underlying
process of pyto it goes a step away from
quantum physics and goes to a much more
abstract space that is built just off
graphs graphs are mathematical
construction that have vertices and
edges between
vertices we found in the last years that
Quantum experiment can be translated
into this abstract grp graph and this
abstruct world is way easier to handle
we can use modern AI tools then we can
ask and answer questions in this much
more abstract world and then when we
have a solution we can translate the
solution back into a real Quantum
[Music]
experiment we were interested in
building a specific Quantum state that
is interesting for quantum
computers so ement is one of the
strangest property that quantum physics
predicts where two particles seem to be
correlated over large
distances the normal way how one creates
entanglement is that you produce
entangled particles at one location they
can be generated at completely different
locations if you have the ability to do
something extra called entanglement
swapping you generate two pairs of
entanglement they can be very far
away one of the two particles from each
of the beers you bring to one location
and you perform a specific
measurement the Butner particles have
never been at the same location but
still they can be
entangled so we wanted to have our
algorithm to come up with new ways to
perform entanglement swapping so one of
the students was working on it and he
has coded it up using py toys he got the
first results so I saw this result and I
knew it shouldn't be possible
for creating entanglement swapping the
belief was that you actually need
entanglement to start with our solution
seemed to not need
it it turned out that result is correct
so that means the limit that I thought
needs to be there was wrong there there
is no limit you can overcome it what did
the machine find to overcome this limit
we found out that it uses a completely
different technique using very different
physical principles but is still able to
entangle two particles that have never
been at the same location one of the
reasons why we are excited about this
technique is because it changes how we
think about generating
entanglement we wrote a paper where we
explained the idea that we understood
from the computer algorithm and the cool
thing is the idea came completely
implicitly from a machine and we as
humans we just try to understand what
the machine has actually done this is
definitely a comp opponent of the future
of science that humans will not maybe
discover all things but maybe the
machines discover and humans have to
interpret what the machine has
done a couple of years ago I got the
email from researchers at the Lago
collaboration and the Lego collaboration
is the one collaboration that has built
gravitational wave detectors and
actually observed the gravitational
waves and they were thinking about
whether one could design new
gravitational wave detect s that are
Beyond human
intuition we are trying to use AI
systems to come up with completely
different experimental
setups actually quite a lot of the ideas
that the machine has put in are
completely alien to us they are totally
weird and we are not sure how to really
think about them yet some of those
designs seem to beat the best human
designs in some cases even quite
significantly
now it's a task of the human to
understand the solution that a nonhuman
intelligence has
provided there are different projects in
my group where we actually try to use a
huge amount of scientific literature to
predict what scientists will do and to
hopefully suggest new interesting high
impact research ideas we create
something that's called a Knowledge
Graph this is a very compressed version
of the knowledge that's contained in the
papers you see how scientists behaved in
the past so you can predict in a way
using modern machine learning methods
what scientists will do in the
future one of the most interesting thing
I can think of in the future is building
algorithms that can basically simulate
all of experimental
physics so your AI system has then not
only access to all experimental
equipment in a virtual way but also to a
huge number of open experimental physics
questions it's likely to find some very
unorthodox
Solutions so I think about this as one
of the most exciting large scale
collaboration between the domain of
physics and the domain of artificial
intelligence that would be a dream to
have that
[Music]