Section 8: The Emergence of the Mind
So far, we have wrestled with the structural diversity of proteins and its relationship to the free energy landscape, and we have tried to find some of the unifying and emergent properties of evolution that might explain the diversity of life and the increase in complexity. We have also taken a look at how the biological networks necessary to bind a collection of inanimate objects into a living system emerge. At the highest level lies the greatest mystery of biological physics: the emergence of the mind from a collection of communicating cells.
Figure 28: Ruby-throated hummingbird.
Source: © Steve Maslowski, U.S. Fish and Wildlife Service. More info
We started our discussion of biological physics by considering a chicken egg. Birds lay eggs, so let's consider a bird: the ruby-throated hummingbird presented in all its glory in Figure 28. About 7 cm long and weighing about 5 grams, this bird is capable of some rather amazing biophysical things. Its wings beat about 50 times each second, and they rotate around their central axis through almost 180 degrees, allowing the bird to fly backwards and forwards and hover. A lot of these fascinating mechanical properties can be considered the subject of biological physics.
But there is far more to these hummingbirds than just flying ability. They live for about nine years, spend their summers in the northern parts of North America and their winters in tropical Central America. So, this small animal can navigate in the fall over thousands of miles, including over hundreds of miles of open water, to certain locations and then return in the spring to the region it was born. How is it that a tiny hummingbird can do all this remarkable navigation? The advancement in capabilities of the digital computer over the past 30 years has been truly staggering, yet it pales against what the hummingbird's brain can do. The human brain is far more impressive. Why can a couple of pounds of neurons drawing a few watts of chemical power with an apparent clock speed of maybe a kilohertz at best do certain tasks far better than a machine the size of a large truck running megawatts of power? And at a much more troubling level, why do we speak of the soul of a person when no one at this point would seriously ascribe any sense of self-recognition to one of our biggest computers? We seem to be missing something very fundamental.
Traditional computers vs. biology
We have moved into the computer age via the pathway pioneered by British mathematician Alan Turing, whom we first met in the introduction to this unit. Our modern-day computers all basically use the model described in Figure 29, coupled with the idea that any number is to be presented by bits in a binary representation. We have made things much faster than those early computers, but the basic idea has not changed. Even the quantum computers promised in Unit 7 keep the same basic design, replacing binary bits with more powerful qubits.
Figure 29: Schematic of a modern digital computer.
But this is not how biology has developed its own computers. The basic design has four major flaws as far as biology is concerned:
- The machine must be told in advance, in great error-free detail, the steps needed to perform the algorithm.
- Data must be clean; the potential loss of a single bit can crash the code.
- The hardware must be protected and robust; one broken lead and the machine can crash.
- There is an exact correspondence between a bit of data and a hardware location: The information in the machine is localized.
None of this is any good for a biological system. As far as biology is concerned, our computers are evolutionary dead-ends. We started this unit by considering the fragility of the egg in a large fall. Yet, as the example of Phineas Gage in the side bar shows, our brain can take enormous abuse and remain basically functional. I challenged you initially to drop the possibly cooked egg and see what happens. Now I challenge you to take a 2 cm diameter steel rod and thrust it through your laptop with great force, then try to surf the web.
The brain of a nematode
The human brain is probably the most complex structure in the universe that we know, but not only humans have brains. The adult hermaphrodite of the "lowly" nematode C. elegans consists of only 959 cells; yet when you watch it navigating around on an agar plate, it certainly seems to be computing something based on its sensory input. The creature displays an astonishingly wide range of behavior: locomotion, foraging, feeding, defecation, egg laying, larva formation, and sensory responses to touch, smell, taste, and temperature, as well as some complex behaviors like mating, social behavior, and learning and memory. It would be quite hard to build a digital computer that could do all that, and certainly impossible to pack it into a tube about 1 mm long and a 100 microns in diameter that can reproduce itself.
The C. elegans doesn't have a brain per se, but it does have about 302 information-carrying neurons that form approximately 7,000 synapses. We believe that any real brain capable of making some sort of a computation, as opposed to the collective behavior seen in single-celled organisms, must consist of neurons that transfer information. That information is not transferred to some sort of a central processing unit. Biological computers are systems of interconnected cells that transfer and process information. The network of neurons in C. elegans displays the common feature in interconnectivity: the synaptic connections formed by the neurons.
The brain versus the computer
I want to concentrate on one thing here: how differently the brain, even the pseudo-brain of C. elegans, is "wired" from the computer that you're using to read this web page. Your computer has well-defined regions where critical functions take place: a section of random access memory (RAM) and a central processing unit (CPU). Each part is quite distinct, and buses transfer binary data between the different sections. Take out a single bus line or damage one of the RAM chips, and the system shuts down.
Brains in biology seem to have evolved in a different way. First, they are spatially diffuse. The computer is basically a two-dimensional device. Brains at every level seem to be basically three-dimensional. The interconnection takes place not via a bus, but rather through a vast network of input-output synaptic connections. For example, C. elegans has roughly 20 interconnects per neuron. In the human brain, we believe that the number is on the order of 103. Since the human brain has around 1012 neurons, the number of interconnects is on the order of 1015—a huge number.
Figure 30: Rainbow images showing individual neurons fluorescing in different colors. By tracking the neurons through stacks of slices, we can follow each neuron's complex branching structure to create the treelike structures in the image on the right.
Source: © Jeff Lichtman, Center for Brain Science, Harvard University. More info
It would be a mistake to think that the 1012 neurons in the brain correspond to about 1012 bits of information, or about 100 Gigabytes. The number is much higher, because of the three-dimensional interconnections linking each neuron with about 103 other neurons. Returning to our theme of spin glasses, we can estimate the information capacity by making the simple assumption that each neuron can be like a spin which is either up or down depending on its storage of a bit of information. This means that the total number of differing configurations of the brain is on the order of , an absurdly huge number, far greater than even the number of atoms in the universe. We can only assume that the brains of living organisms emerged as they exploited this immense 3-D information capacity owing to the ability of communities of cells to form neuronal interconnections throughout space.
How does the brain reason?
Given the large information capacity of even a small network of neurons and the fact that the human brain's capacity exceeds our ability to comprehend it, the next question is: How does a brain reason? As usual, we need to start by defining what we're talking about. According to the Oxford English Dictionary, "reasoning" is "find[ing] a solution to a problem by considering possible options." I suppose this dodges the question of the emergent property of consciousness, but I don't see this problem being solved any time soon, although I hope I am wrong.
The hummingbird has a big problem, essentially asking itself: How shall I fly back to a place I was at six months ago that is thousands of miles away from where I am now? Presumably, the bird uses different physics than that of a traditional computer, because the information content that the bird has to sort out would cause it to fail catastrophically. So, we finally have the problem that perhaps physics can attack and clarify in the 21st century: How can a set of interacting neurons with a deep level of interconnects take previously stored information and determine an optimal solution to a problem it has not yet seen?
The hummingbird faces a problem rather reminiscent of the traveling salesman problem, explained in the sidebar. To choose the correct locations to pass through on its springtime journey north, it must consider a number of combinations far beyond the power of any computer system to resolve. How does the hummingbird do it? Is it magic?
Physics shows that it isn't magic. As we have previously discussed, while a protein may fold or a species play with its genome in an almost uncountable number of ways, basic free energy minima schemes lead quite efficiently to a vastly smaller set of combinations that are roughly optimal. Nature doesn’t necessarily find the "best" solution, but it seems able to efficiently find a subset of solutions that works well enough. In the case of the traveling salesman problem, the vast combinatorics interconnects of a neural network of many neurons provides exactly the kind of search over a free energy surface that we need.
The "reasoning" ability of neural networks
We have discussed how landscapes—either fitness landscapes or free energy landscapes—can give rise to vastly complex surfaces with local minima representing some particular desired state. John Hopfield, a theoretical physicist at Princeton University, has explored ways for a system to find these minima. The three basic ideas below highlight how biological computers differ from their electronic counterparts:
- Neural networks are highly interconnected. This interaction network can be characterized by a matrix, which tabulates the interaction between each pair of neurons.
- Neurons interact in a nonlinear analog way. That is, the interconnection interaction is not an "all or nothing" matter, but a graded interaction where the firing rate of neurons varies smoothly with the input potential.
- An "energy function" can be constructed that allows us to understand the collective (or emergent) dynamics of the neuron network as it moves over the information landscapes and finds local minima that represent effective solutions to the problem.
Figure 31: An optimal traveling salesman problem (TSP) tour through Germany's 15 largest cities. It is the shortest among 43,589,145,600 possible tours visiting each city exactly once.
Source: © Wikipedia, Creative Commons Attribution-ShareAlike License. More info
Hopfield and molecular biologist David Tank set out to make an analogy between neural networks and the energy network of a glassy system characterized by a large number of degrees of freedom. Following the three principles outlined above, they used this analogy to write an equation for the free energy of a neural network in terms of the interaction between each pair of neurons, the threshold for each neuron to self-fire, and the potential for each of the neurons in the network. They also recognized that the interaction between pairs of neurons can change with time as the neural network learns.
The solution to a problem such as the traveling salesman problem emerges in the neural network as interaction strengths between the neutrons are adjusted to minimize the free energy equation. The flow of the neuron states during the computation can be mapped onto a flow on a free energy surface, similar to the flow of a spin glass toward its ground state or natural selection on a fitness landscape (but in the opposite direction). Clearly, quite complex and emergent neuronal dynamics can evolve with even the simple system we are considering here.
Hopfield and Tank showed that this neuronal map has quite impressive "reasoning" ability. A set of 900 neurons encoded to solve a 30-city traveling salesman problem was able to find 107 "best" solutions out of the 1030 possible solutions, a rejection ratio of 1023 in just a few clock cycles of the neural network.
Although we clearly are a long way from understanding the emergent nature of consciousness, this example reveals the immense computational power of neural networks. Surely, one of the grand challenges in 21st century physics will be to move from these simple physical models derived from very concrete physics concepts to the vastly more complex terrain of the human brain.