Name of Program: Robots – Podcast and Community
Name of Episode: Robots: Brain-Machine Interfaces
URL: http://www.robotspodcast.com/podcast/2009/08/robots-brain-machine-interfaces.html
Date of podcast: August 14, 2009
Interviewer: Sabine Hauert
Interviewee: Prof. Steve M. Potter
Adam: Steve Potter is a professor at the Laboratory for Neuroengineering at Georgia Tech. He
was the first to hook in vitro [cultured] neurons to robots in an effort to study learning and
memory. The resulting stimulating systems are controlled by a brain grown in a petri-dish and
have artificial bodies. Having the brain in the dish makes it easier to get a sneak peek into
what’s going on and even provide treatment ideas for patients with epilepsy.
Sabine: Hi Steve, welcome to Robots.
Steve: Hello Sabine. This is a great honor. I’m a fan of the show and listen to it often so it’s
really wonderful be on it.
Sabine: Excellent. Can you maybe present yourself to our listeners?
Steve: Ok. I’m a professor of Biomedical Engineering in the Laboratory for Neuroengineering
at Georgia Tech; that’s in Atlanta in the United States and my department is also shared with
Emory University’s Medical School, which is nearby Georgia Tech. I am a neuroscientist and an
engineer. My PhD is in Neuroscience, but I have built a lot of things and love to work with my
hands and build things and so my department is ideal for that because I can be both. I study
learning in memory and neuronal dynamics. So, what does that mean? We are interested in
finding out what changes about your brain when you learn something so hopefully your listeners
as they are listening to this are having a few of their neurons altered in some way so that after the
show, their brains will be a bit different. At the present, we don’t really understand how that
works so my lab is trying to chip away at that large stone of ignorance and hopefully answer
some questions about physically what changes about brains when you learn something.
Sabine: We just spoke with Chuck Higgins, who has interfaced robots with fully developed
insect brains and you have been looking at a more bottom-up approach and trying to understand
basically how the brain functions. You’ve been looking at developing a brain or growing a brain
and then interfacing it into a robot. Now, first of all, how do you grow a brain and sustain it?
Steve: Ok, maybe I’ll introduce the concept of in vitro Neuroscience. In vitro means in glass
and it means that we grow brain cells in a petri-dish and the ones that we use are very special
petri-dishes that are instrumented with electrodes that are built into the bottom of the dish; 60
electrodes that we can hook up to the computer. We can use our electronics to either stimulate
the cells in the dish or we can record their activity. So, these are cultured neural networks so
they’re actually living cells that we got from rat embryos. We dissect the cortex and dissociate
the cells and pour them into the dish and they settle down on the surface and start growing, and
forming connections. We make time-lapse movies with our microscopes of this process while
they’re growing. They’re submerged in a bath of liquid that’s basically what’s in your blood
minus the blood cells, its all the vitamins and salts that you need to keep brain cells alive. We
put them in an incubator to keep them warm at body temp, we control the atmosphere carefully
so that the medium they’re in stays at the right pH. So, they’re very touchy cells. Unlike other
cultured cells, neurons don’t divide so whatever cells we but in the dish at the beginning of the
experiment are all the cells we get for the entire experiment, which could last for months.
Sometimes we grow these cells for months. We’ve actually have grown one culture for two
years so we have spent quite a bit of effort preventing these from getting any germs in them. If
that petri-dish gets even one bacterium in it, the cultured neurons will die. So it doesn’t have its
own immune system like a brain in your body does. And, I should say that a cultured network
calling it a brain is overstretching the truth a bit. There are maybe 50,000 brain cells in there but
they don’t have nearly as much organization as an intact brain so that’s a plus or a minus. It’s a
plus because it much simpler and we might understand it more easily than we can understand an
intact brain, but the obvious minus is that whatever we learn may or may not be relevant to intact
brains in humans. So we have to keep on reminding ourselves that in vitro neuroscience has its
limitations.
Sabine: What can you learn from growing this brain?
Steve: Well, I’ll get more specific about some of the things that we’re interested in, for studying
learning and memory. For example, these networks, as they’re growing, the cells start
communicating with each other with small electrical impulses that we pick up with our
electrodes, and those things are called action potentials. Now these are the signals that go down
your nerves from your brain to your muscles that get the muscle to contract or they get your
glands to secrete something. In our dishes, sometimes these signals, when we pick them up, are
used to control robots. So we were the first ones to connect cultured neural networks to robots. I
think that was in year 2000, and we used them [robots] as a visualization tool to see whether or
not the networks are learning. So, if you just have a network in a petri dish, you can’t really tell
what’s going on in there, what’s its thinking. What does this activity represent? But if you
embody this network by hooking it up to a robot so its actual activity can drive the robot around
(say, the K-team’s Khepera robot is one we’ve used) then you can see if there’s any changes in
the activity patterns that change the behavior of the robot. You can also take the sensory signals
from the robot’s sensors – such as infrared proximity sensors, which is what the Khepera has –
you can take that and encode it to stimulation patterns that you deliver to the culture neural
network. So we built software and hardware to make a closed-loop system whereby the cultured
neural network activity moves the robot; it behaves in a laboratory environment. It’s sensors pick
up new signals and those get transmitted back into the dish where hopefully they’ll make some
changes and the robot will learn something. So, one of the advantages of the in vitro
neuroscience approach is that we can take this cultured neural network and put it under the
microscope and look at it in great detail, and see what’s going on in there even while it’s
behaving, which is very hard to do with an animal because animals move, right? So, all that
movement makes it hard to do any kind of microscopic imaging. So if you want to look with a
microscope and see how brains are thinking, it really helps if you have a simplified brain in a
petri-dish.
Sabine: Let’s concentrate a bit on one of the more recent projects you’ve done, which is the
Semi-living Artist. Can you present this project to us?
Steve: Okay, the semi-living artist is named MEART and that stands for Multi Electrode Array –
Art. So, the multi-electrode array is the instrumented petri-dish that we use. And MEART is a
collaboration that we have had since about the year 2002, with an art-science collective at the
University of Western Australia in Perth, and the collective is called SymbioticA. And, the
fellow who we collaborate with there is named Guy Ben-Ary, and he works with a robot artist
named Phil Gamblen. They built this beautiful arm that draws on large sheets of paper and when
I first met those two, they were controlling this arm with some neurons that they got from
goldfish so they called it “fish and chips” because they were trying to grow these goldfish
neurons on silicone chips. So I said, “We could probably control that robot arm also with
neurons in our lab in Atlanta.” We set up the system to do that. It means that my hard working
graduate students figured out how to communicate over the internet clear across the globe
(Atlanta’s pretty much exactly at the opposite side of the globe from Perth) and have the signals
of a cultured neural network drive this thing around and cause it to draw drawings. There’s a
camera there in Perth, or actually this thing has been on tour all over the world, so wherever it
happens to be. This camera is sending signals back to Atlanta through the internet and those
signals get translated into some kind of electrical stimulus that we deliver to the cultured network
and we use that to try to train the network, hopefully, to draw something that we can recognize. I
have to say that it’s still at a toddler scribbling level but it’s very interesting in a philosophical
sense that this is a half-living half-mechanical entity that can express some kind of creativity. In
other words, the activity in the network is what really determines what the drawing looks like.
The kind of thing we have learned by playing around with MEART is that when it is being
exhibited for say, two weeks at some art exhibit, it will be getting stimulation all the time, pretty
much. It will be drawing and looking at its drawings and we’ve observed that after a while that
stimulation tended to calm the network down. So previously, the drawings might have been
rather erratic and there was lots of activity, and as the thing drew it moved less and the network
had less of this spontaneous activity. So, this actually has resulted in a spin off that our lab is
doing; a medical spinoff that we’re now using electrical stimulation to calm epilepsy in
experimental animals. And if it works, we obviously would like to try this out in humans who
have intractable epilepsy. So even though this is a very fanciful project, MEART has some
tangible benefits when you look at the science behind it.
Sabine: Is there a reward mechanism to direct the learning?
Steve: that is a very good question. In the brain, in intact brains, we have very complex circuitry
that does the rewarding and allows certain connections to be strengthened, say, by releasing
neurotransmitters such as dopamine. In our culture dishes, we do not have this complex circuitry.
So one approach we can do is to apply the reward chemically with some sort of miniature
spritzer. Called a ”Pico-spritzer”, actually. We can put dopamine or some chemical like that into
the spritzer. We built this system that allows the network to apply its own dopamine if it has
accomplished something that the robot was supposed to do, something that we say was a good
thing. We weren’t successful in getting this to induce any learning. So we have a lot more work
to do to understand how does dopamine cause these changes in connections.
So you could also say that neurons and neural networks, most of the brain cells in your brain, are
devoted to noticing coincidences. They’re looking at the world around them and the inputs and
trying to find out what is related to what, and when they find that out they make connections
between those two different things. So at the psychological level we call that associations and at
the cellular level we might just call it a synapse. But you could say it is intrinsically rewarding to
these cells to do that because it is something that is very hard to get them not to do; they really
want to form associations. We have to work hard in our culture dishes to prevent this happening
if we wanted to. Basically, if you give a network lots of inputs, it will start to notice associations
between them and strengthen connections between them.
Sabine: So what are the future steps with this semi living artist?
Steve: Ok, so MEART, the robot that draws on the large pieces of paper has actually been retired
to give way to a new thing that the folks in Perth have created called Silent Barrage. This thing
was presented in New York recently and it consists of a room full of pillars that are controlled by
our neurons here in the lab in Atlanta. They are robotic pillars you could say. They have this
mechanism that spins around the pillar and can move up and down the pillar and draw on paper
that has been wrapped around the pillar. So each one of these pillars represents one of the
electrodes in our culture dish and the height of this thing spinning around represents how much
activity that electrode is recording from the brain cells. So, what the effect is as you walk
through this room of pillars, is you have all these things moving up and down and spinning
around and making a lot of racket. You’re experiencing the on-going activity of this culture dish.
There is some kind of feedback system where there are cameras above watching visitors to the
gallery. Where they happen to walk is where the stimulation goes. We pick up the images and
translate those into stimuli that try to excite the network based on where the people are and what
they are doing. There are a lot of technical glitches that still need to be worked out. But you
could say that the first exhibit was a beta test and it is hopefully going to go on exhibit in Europe
soon and they’re going to improve some of those things. This is an interesting art-science
collaboration that has really helped us get scientists talking about art and artists talking about
science and seeing how they can contribute to each other.
Sabine: Do you think this new environment will help with the learning or is it more for the art?
Steve: So, so far, this situation in silent barrage that we have set up does not have a specific goal
that it is trying to learn; you could say that is it just trying to absorb the info about where the
attendees are and respond to it. So it is more of a stimulus response thing. We hoped, however,
that after a while we will see some changes in the stimulus responses, and then we will start to
study some kind of learning effect. So I would say that at this point it is more of an art project
than a science project. Our studies with wheeled mobile robots are more scientific and in that
case, we are trying to get them to solve some very simple navigation tasks by training them with
certain kinds of patterned electrical stimulation.
Sabine: I’m curious where do you now stand with these wheeled robots?
Steve: We have recently published a paper in which we managed to accomplish, after eight years
of trying, getting the network to control a robot to move in a certain direction by electrically
stimulating it. So, we previously were able to change the networks with stimulation but we
weren’t able to predict ahead of time what kind of changes we could induce. So this is the first
example of successful goal-directed learning in a cultured network that controls a robot. It
required us to come up with sort of an evolutionary scheme for training the network with patterns
of electrical stimulation. We start out with a bunch of random patterns and we keep applying
them to the network until the robot does the right thing and then we use the ones that worked to
train it and we keep on applying those more and more often. Eventually, after a few minutes, we
manage to get the network to control the robot correctly. Then we studied “How does it forget?”
and we stopped the training process and see how long it continues to behave properly. So far, it
only seems to remember things for about 90 minutes. Obviously, we have a long way to go in
understanding the brain processes of learning, memory, and forgetting.
Sabine: You raised an important issue before, in that you are trying to have these robots be
creative and show intelligent behavior. To what extent do you think, if this robot displays
creative behavior, for example, you consider it a living being or a machine?
Steve: Yes, well I would say I think of these as semi-living. They consist of actual, real, live
living tissue that we have to feed and take care of. There’s a lot of metabolism going on and if
you look at it under the microscope, you will be stunned by how complicated it all is. So, there is
no doubt in my mind that it is alive. To call it an animal usually is something that most people
are not willing to accept, so we use the word animat. We say that it is a simulated animal of some
kind that is either simulated in the robotic form, or it could just be a simulation on the computer,
that is being controlled by the neurons. So, we call it a semi-living animal or an animat and I’m
perfectly happy to entertain other points of view on how much alive it is or isn’t. As I said
before, it is missing a lot of the crucial circuitry that is in intact brains.
Sabine: Thanks, Steve, for being here with us on Robots.
Steve: It’s been my pleasure. Thank you, Sabine.