Building a More Brain-like Computer

Building a More Brain-like Computer


The human brain is the most powerful
supercomputer in the world. All right, let’s see this electrical headquarters
of yours in operation. It helps us navigate our environment by carrying out about one thousand trillion
logical operations per second. It’s compact, uses less power than a lightbulb
and has potentially endless storage. The human brain is really one of the most
complex systems that we can imagine. We have a fundamental lack in our understanding
of the way the components in the brain interact. But it is this very interaction that generates
cognition and consciousness. All these mind-boggling intricacies have driven
our fascination with the brain and for centuries we’ve been trying to map
and understand it. And most recently – replicate it. The brain is certainly a computer that has been evolving
for nearly 4 billion years. And the more we learn about the brain, the
more we’re able to incorporate the smart ways that it does computation
into our artificial devices. Scientists are beginning to agree that to
realize our technological dreams, we need to build computers that work like our brains. One day these computers could in turn help
us unlock more secrets of cognition. The brain is packed with neuron cells that
constantly communicate with each other through electrical pulses, known as spikes. Each neuron releases molecules that act as
messengers and control if the electrical pulse is passed along the chain. This relay race is happening simultaneously
throughout billions of neurons. Much like the zeros and ones of the computer
world, this is the basic language of the brain. But understanding all of this isn’t enough. We’ve still only scratched the surface when
it comes to figuring out how the brain works. The more I’m working on the brain,
the more I understand how complex it is, how difficult it is. Many relatively easy cognitive functions cannot
really be understood at the level of cells. The human brain
is one of the biggest secrets and mysteries that we have, despite many years of intensive work. Katrin Amunts is at the helm of
the Human Brain Project, a 10 year long attempt at studying the brain. With researchers collaborating across 100 universities, the project is expected to cost around €1 billion. Professor Amunts and her team are working
on one part of it, a 3D digital brain atlas. They are creating three different high resolution
maps – one of neurons; one of their connections – which uses different
colors to indicate the orientation of neurons’ branches; and one map of the receptors
for the messenger molecules. When we think about an atlas of the world,
we can map all the different countries. But then we can also see there are maps illustrating
the level above the sea or the temperature. And it’s a little bit like what we have
in the human brain. There are different aspects. We want to understand where the cells are located. We want to understand where certain areas
are located, how they are connected, what is their molecular profile, what is their gene expression that
is important for function. So there is not one single aspect that can
explain everything in the human brain. So that means we need different types of maps
that reflect different aspects of brain organization. To create the maps, the team is scanning slices
of post-mortem brains. We get brains from body donors and process them, embed them in paraffin, and then cut them
into 20-micrometer-thick sections. 20 micrometers, this is approximately like
thickness of one hair so this is very thin. One brain has approximately 7000 sections. These sections can then be analyzed under
the microscope and we can then reconstruct the areas in 3D. Much like a fingerprint, every brain is unique,
so to account for these differences, the team scan 10 brains for each of their maps. This generates petabytes of data that’s
analyzed with the help of AI and used to run brain simulations on supercomputers but even the supercomputers struggle. So to further our understanding of the brain,
we need better machines. We cannot make our chips much faster
without them melting, unless we designed completely new architectures. We cannot make our components much smaller
because then we reach component sizes where quantum effects take over. So the computation becomes too imprecise
to be practical. We need to find better solutions in order
to increase our computational power. Mihai Petrovici, like many other scientists
in the field, thinks that modeling computer hardware
on our brains is the way to go. It will not only increase the speed and efficiency
of future machines, but also help build better AI. There are certainly things that computers
do much better than brains, such as adding or multiplying big numbers,
because this is what they were designed for. Intricate problems in mathematics are accurately
solved in the minute fraction of the time required for a human calculation. There is no evolutionary pressure for us to
be able to multiply big numbers. Otherwise, certainly our brains would be able to do it. However, there is a strong evolutionary pressure
to recognize your surroundings, to be able to build an internal model
of your surroundings. When you hear a noise in the bushes, to be
able to imagine that maybe there’s a predator there. To be able to recognize faces in order to
live in a society where people can actually communicate and cooperate. And this is what evolution has made our brains excel at. This ability to build an internal model of
the world, to have, sort of, the world inside your heads, to imagine what is happening around you
even if you don’t see it, this is of critical importance for a true
artificial intelligence. AI like Google image recognition, Alexa or
the autopilot in a self-driving car all work thanks to neural networks, software which already tries to imitate
the way our brain recognizes patterns. One thing that today’s artificial intelligence
needs in order to be able to perform whatever task it was designed for, is a lot of examples. So in order for Google, for example, to be
able to show you pictures of cats, whenever you type in cat, it needs to have seen millions of images of cats. That is certainly not how we humans operate and learn. When you show a child, for example, a cat
or whatever other new thing, it just needs to see it a couple of times
in order to quickly grasp the main features that are specific for that animal and then recognize it whenever it sees another
individual of the species. The scientists at Heidelberg University are
working on a different part of the Human Brain Project. They’re using the brain maps developed by
Professor Amunts’ team to build computer hardware they hope will help AI
learn like our brains do. This new hardware is called neuromorphic which
means formed like neurons or like the brain. Actually, none of what you see here on the
outside is really neuromorphic. You might be tempted to think that this is
more or less like the machine that you have at home on or under your desk. This would be true for the outside components but at the heart of the system, there lies
a piece of hardware that is fundamentally radically different from the chips in your computer, and that is the neuromorphic heart of the system. The microchips on these wafers look nothing
like the entangled web of neurons that we have in our heads. But each component communicates like an individual
neuron by sending along spikes of electricity
to their many partners. This design immensely increases the operating speed. neuromorphic hardware generates results
10 million times faster than conventional hardware. We certainly believe that this will become
a big thing, we will see many applications of these systems for everyday tasks. One of them would be face recognition, pattern
recognition in general, speech recognition, the ability to read texts. The ultimate goal, of course, is to create
true artificial intelligence. But it’s really hard to say by when we will be able to actually copy the brain in an artificial substrate. What we can certainly do and what we are doing
right now is – understand particular aspects of computation in the brain. The 4 million artificial neurons packed into
this neuromorphic computer are just a tiny fraction of
the 86 billion neurons in the human brain. Still, it’s a big step forward for the machines. Even though our knowledge of the brain has
increased over the last few decades, it’s still fragmented. If the Human Brain Project is successful,
it could bring this knowledge together and encourage research and collaboration across
different scientific fields. And so this effort could be just the beginning
of the journey. Better understanding the human brain, is really
one of the challenges of the 21st century. We have an increasing amount of people suffering
from neurodegenerative diseases, suffering from major depression,
other psychiatric diseases. We need to have new tools to diagnose and
have better therapies for these brain diseases. And since we are living in an aging population, these diseases, of course,
play a major role in the future.