USC researcher discusses building artificial neurons

USC researcher Dr. Alice Parker and her team at USC have created a rudimentary neural circuit composed of carbon nanotubes. This circuit can emulate certain aspects of a biological synapse. Although it is vastly simpler than a biological neuron, this nanotube circuit represents the first step towards the eventual creation of a fully functional artificial neuron. In an interview with Sander Olson, Dr. Parker discusses the future of artificial neural circuits and “brainlike” computers.

Alice Parker
Question: Your team at USC has developed a neural circuit. Tell us about that.
The circuit is composed of carbon nanotubes, and essentially converts one analogue waveform signal to another. These output signals are lower in amplitude than the input signals and longer in duration. This circuit is designed to roughly approximate certain aspects of the behavior of a neuron, the synapse, but it is vastly simpler and less capable than an actual biological synapse.

Question: How long have you been working on this?
I’ve been working on this for the past five years. I was originally working on integrated circuit CAD design software. But six years ago my team and I engaged in the DARPA grand challenge for autonomous vehicles. It was clear to me that we needed more brainlike computers. I expected this to be a small endeavor but it turned out to be hugely popular with my students and has grown much bigger than I ever anticipated.
Question: How many people are collaborating with you on this?
I have been collaborating closely with Prof. Chongwu Zhou on the carbon nanotube implementation. I have seven PhD students working on the entire BioRC project. Two of my PhD students along with two of Chongwu Zhou’s students were able to create the carbon nanotube synapse. I loosely collaborate with other faculty both inside and outside USC. There are about 15 Masters students who are working on the project. I even have a couple of bright undergrads working on this. So many students have shown interest in this that I’ve actually had to turn students away.
Question: At what speeds can this circuit operate? To what extent can it be improved?
The nanotubes themselves are capable of extraordinarily fast rise and fall times – several orders of magnitude faster than current silicon CMOS transistors. So we are talking about picosecond rise and fall times. It should in theory be possible to compensate for the relative simplicity of our devices by increasing switching speeds.
Question: How accurately does this circuit mimic an actual neuron?
The circuit mimics the input – the function performed by the synapse. This is only a portion of the functioning of an actual biological neuron. To create a device that could perform all of the actions of an actual biological neuron would require many more circuits. So this is a baby step, the first of many required steps, towards an artificial brain.
Question: So an actual biological neuron is an order of magnitude more complex than your neural circuit?
An actual biological neuron would be orders of magnitude more complex than this circuit. In a neuron in the cortex, there are probably on the order of 10,000 synapses, each synapse contains memory and learning mechanisms, and each synaptic input is combined with the others in a complicated, nonlinear fashion. By comparison, our neural circuit is a primitive, crude, and simple device.
Question: So if this neural circuit is only a first step towards creating an artificial brain, what is the second step?
The second step is to use a second nanotube transistor to increase our control of the synapse. We would like to create a neuron that summed multiple inputs. This would be a rather difficult task and would probably require a few years. We will need to create large numbers of carbon nanotubes and then arrange these nanotubes to create transistors. It is currently difficult to build a circuit with even a single nanotube transistor or a few transistors.
Question: How many other researchers are trying to create neural circuits?
There are groups both in the U.S. and in Europe that are working in this general area. In the nanotechnology area, HP is doing work with memristors, which are two-terminal devices, and DARPA is funding some efforts. This entire field of nanoelectronics for neural circuits is fairly new and didn’t exist a decade ago, but it is growing rapidly.
Question: How difficult will it be to create an integrated circuit using thousands or millions of these neural circuits?
We envision fabricating substantially more complex circuits within the next five years. But in order to make true brainlike circuits we will have to incorporate some form of 3-d manufacturing. Creating an artificial brain by using 2-D CMOS manufacturing simply won’t work. There is a new flip-chip technology which is very dense, and it might be dense enough to do some neural circuits. The way to truly solve this problem is by going to 3-D.
Question: Have you thought about how these circuits could be interconnected?
Yes, the interconnection problem is intractable. We will need to create some kind of a switching network that has the necessary speed, flexibility and robustness. We will eventually create self-assembling, self-modifying circuits using nanotechnology that can continually adapt and form new connections.
Question: Will it ever be possible to create a general purpose digital computer using these neural circuits?
Although it is possible to use these neural circuits to perform straightforward digital computations, that is not efficient. It would require substantial overhead to perform a task that silicon CMOS already does quickly and inexpensively. But I can foresee more “brainlike” computers that are hybrid machines incorporating classical computing paradigms along with neural circuits.
Question: When do you think that these first “brainlike” computers may emerge?
We could see the first experimental systems emerge within the next five years. The first commercially available hybrid computer could become available within a decade. But these computers probably won’t be PCs. Rather, they will be in embedded systems that perform some specific function.
Question: What is the best way to program these neural circuits?
Programming is quite tricky. One can’t directly build an intelligent neural network, one can only build a neural net that will learn and through learning become intelligent. But every time we think that we understand biological neurons, the neuroscientists discover another mechanism essential to their operation. We will need to discover quite a bit more about how the brain works before we can design neural circuits of equivalent sophistication.
Question: Have you done any research with graphene?
No, but several colleagues of mine are working on graphene. Graphene seems to be somewhat easier to work with than nanotubes. I am an agnostic when it comes to nanotechnology – I am willing to work on any given nanotechnology. I’ve done some work on memristors, which are limited because they are 2-terminal devices. I prefer working with nanotubes because I can envision building 3-d devices from them. I don’t think that quantum dots are appropriate for this, although might be used for classical digital computations.
Question: Could silicon CMOS ever be used to create a neural circuit?
Yes, and that concept has obvious advantages due to the fact that it is currently the dominant computing paradigm for hardware. But it has issues with connectivity and plasticity. Creating structural plasticity using CMOS circuits is possible but problematic because the CMOS circuits are fixed and poorly suited to forming new connections. So nanotubes appear the best solution at this point.
Question: How long will it be before the first synthetic brain emerges?
I think that the first synthetic brain could emerge within 50 years. There are others, such as Ray Kurzweil, who are considerably more optimistic than I am, and who are confident that we will have something like a synthetic brain within the next two decades. That optimism will push us forward, since it leads people to accomplish tasks that they otherwise wouldn’t. But I am more pessimistic because what biology accomplishes is breathtakingly difficult. So the truth may lie somewhere in between.
Question: How much progress do you anticipate in nanotube electronics by 2021?
By 2021 we could have some very sophisticated prosthetic devices. USC researchers are doing groundbreaking research on connecting artificial devices to living tissue, so we could see prosthetics which are almost as good as an actual limb. By 2021 brainlike computers should be available, and this should enable autonomous vehicles and even driverless cars. I wouldn’t be surprised if we see platoons of cars powered by synthetic brains, driving in sync. Although it will take much longer than ten years to perfect this technology, we should see the first commercial applications emerge within a decade or less.

If you liked this article, please give it a quick review on ycombinator or StumbleUpon. Thanks

USC researcher discusses building artificial neurons

USC researcher Dr. Alice Parker and her team at USC have created a rudimentary neural circuit composed of carbon nanotubes. This circuit can emulate certain aspects of a biological synapse. Although it is vastly simpler than a biological neuron, this nanotube circuit represents the first step towards the eventual creation of a fully functional artificial neuron. In an interview with Sander Olson, Dr. Parker discusses the future of artificial neural circuits and “brainlike” computers.

Alice Parker
Question: Your team at USC has developed a neural circuit. Tell us about that.
The circuit is composed of carbon nanotubes, and essentially converts one analogue waveform signal to another. These output signals are lower in amplitude than the input signals and longer in duration. This circuit is designed to roughly approximate certain aspects of the behavior of a neuron, the synapse, but it is vastly simpler and less capable than an actual biological synapse.

Question: How long have you been working on this?
I’ve been working on this for the past five years. I was originally working on integrated circuit CAD design software. But six years ago my team and I engaged in the DARPA grand challenge for autonomous vehicles. It was clear to me that we needed more brainlike computers. I expected this to be a small endeavor but it turned out to be hugely popular with my students and has grown much bigger than I ever anticipated.
Question: How many people are collaborating with you on this?
I have been collaborating closely with Prof. Chongwu Zhou on the carbon nanotube implementation. I have seven PhD students working on the entire BioRC project. Two of my PhD students along with two of Chongwu Zhou’s students were able to create the carbon nanotube synapse. I loosely collaborate with other faculty both inside and outside USC. There are about 15 Masters students who are working on the project. I even have a couple of bright undergrads working on this. So many students have shown interest in this that I’ve actually had to turn students away.
Question: At what speeds can this circuit operate? To what extent can it be improved?
The nanotubes themselves are capable of extraordinarily fast rise and fall times – several orders of magnitude faster than current silicon CMOS transistors. So we are talking about picosecond rise and fall times. It should in theory be possible to compensate for the relative simplicity of our devices by increasing switching speeds.
Question: How accurately does this circuit mimic an actual neuron?
The circuit mimics the input – the function performed by the synapse. This is only a portion of the functioning of an actual biological neuron. To create a device that could perform all of the actions of an actual biological neuron would require many more circuits. So this is a baby step, the first of many required steps, towards an artificial brain.
Question: So an actual biological neuron is an order of magnitude more complex than your neural circuit?
An actual biological neuron would be orders of magnitude more complex than this circuit. In a neuron in the cortex, there are probably on the order of 10,000 synapses, each synapse contains memory and learning mechanisms, and each synaptic input is combined with the others in a complicated, nonlinear fashion. By comparison, our neural circuit is a primitive, crude, and simple device.
Question: So if this neural circuit is only a first step towards creating an artificial brain, what is the second step?
The second step is to use a second nanotube transistor to increase our control of the synapse. We would like to create a neuron that summed multiple inputs. This would be a rather difficult task and would probably require a few years. We will need to create large numbers of carbon nanotubes and then arrange these nanotubes to create transistors. It is currently difficult to build a circuit with even a single nanotube transistor or a few transistors.
Question: How many other researchers are trying to create neural circuits?
There are groups both in the U.S. and in Europe that are working in this general area. In the nanotechnology area, HP is doing work with memristors, which are two-terminal devices, and DARPA is funding some efforts. This entire field of nanoelectronics for neural circuits is fairly new and didn’t exist a decade ago, but it is growing rapidly.
Question: How difficult will it be to create an integrated circuit using thousands or millions of these neural circuits?
We envision fabricating substantially more complex circuits within the next five years. But in order to make true brainlike circuits we will have to incorporate some form of 3-d manufacturing. Creating an artificial brain by using 2-D CMOS manufacturing simply won’t work. There is a new flip-chip technology which is very dense, and it might be dense enough to do some neural circuits. The way to truly solve this problem is by going to 3-D.
Question: Have you thought about how these circuits could be interconnected?
Yes, the interconnection problem is intractable. We will need to create some kind of a switching network that has the necessary speed, flexibility and robustness. We will eventually create self-assembling, self-modifying circuits using nanotechnology that can continually adapt and form new connections.
Question: Will it ever be possible to create a general purpose digital computer using these neural circuits?
Although it is possible to use these neural circuits to perform straightforward digital computations, that is not efficient. It would require substantial overhead to perform a task that silicon CMOS already does quickly and inexpensively. But I can foresee more “brainlike” computers that are hybrid machines incorporating classical computing paradigms along with neural circuits.
Question: When do you think that these first “brainlike” computers may emerge?
We could see the first experimental systems emerge within the next five years. The first commercially available hybrid computer could become available within a decade. But these computers probably won’t be PCs. Rather, they will be in embedded systems that perform some specific function.
Question: What is the best way to program these neural circuits?
Programming is quite tricky. One can’t directly build an intelligent neural network, one can only build a neural net that will learn and through learning become intelligent. But every time we think that we understand biological neurons, the neuroscientists discover another mechanism essential to their operation. We will need to discover quite a bit more about how the brain works before we can design neural circuits of equivalent sophistication.
Question: Have you done any research with graphene?
No, but several colleagues of mine are working on graphene. Graphene seems to be somewhat easier to work with than nanotubes. I am an agnostic when it comes to nanotechnology – I am willing to work on any given nanotechnology. I’ve done some work on memristors, which are limited because they are 2-terminal devices. I prefer working with nanotubes because I can envision building 3-d devices from them. I don’t think that quantum dots are appropriate for this, although might be used for classical digital computations.
Question: Could silicon CMOS ever be used to create a neural circuit?
Yes, and that concept has obvious advantages due to the fact that it is currently the dominant computing paradigm for hardware. But it has issues with connectivity and plasticity. Creating structural plasticity using CMOS circuits is possible but problematic because the CMOS circuits are fixed and poorly suited to forming new connections. So nanotubes appear the best solution at this point.
Question: How long will it be before the first synthetic brain emerges?
I think that the first synthetic brain could emerge within 50 years. There are others, such as Ray Kurzweil, who are considerably more optimistic than I am, and who are confident that we will have something like a synthetic brain within the next two decades. That optimism will push us forward, since it leads people to accomplish tasks that they otherwise wouldn’t. But I am more pessimistic because what biology accomplishes is breathtakingly difficult. So the truth may lie somewhere in between.
Question: How much progress do you anticipate in nanotube electronics by 2021?
By 2021 we could have some very sophisticated prosthetic devices. USC researchers are doing groundbreaking research on connecting artificial devices to living tissue, so we could see prosthetics which are almost as good as an actual limb. By 2021 brainlike computers should be available, and this should enable autonomous vehicles and even driverless cars. I wouldn’t be surprised if we see platoons of cars powered by synthetic brains, driving in sync. Although it will take much longer than ten years to perfect this technology, we should see the first commercial applications emerge within a decade or less.

If you liked this article, please give it a quick review on ycombinator or StumbleUpon. Thanks