In recent years some researchers have given up on instructing computers to see. Why not let a computer teach itself? Collect a huge number of examples of a visual task performed by humans, and then program a computer to imitate these examples. If the computer succeeds, it will have “learned” the task without any explicit instruction. This method is known as machine learning, and it’s an important subfield of computer science. It has yielded the digital cameras that focus on faces, as well as many other successes in AI.
Several laboratories around the world, including my own, are using machine learning to train computers to see neurons. We start by making use of the kind of software that John Fiala and Kristen Harris developed. People manually reconstruct the shapes of neurons, which serve as examples for the computer to emulate. Viren Jain and Srini Turaga, my doctoral students at the time we began the work, have devised methods for numerically “grading” a computer’s performance by measuring its disagreement with humans. The computer learns to see the shapes of neurons by optimizing its “grade” on the examples. Once trained in this way, the computer is given the task of analyzing images that humans have not manually reconstructed. Figure 37 shows a computer reconstruction of retinal neurons. This approach, though still in its beginning stages, has already attained unprecedented accuracy.
Figure 37. Neurons of the retina reconstructed automatically by computer
Even with these improvements, the computer still makes errors. I’m confident that the application of machine learning will continue to reduce the error rate. But as the field of connectomics develops, computers will be called upon to analyze larger and larger images, and the absolute number of errors will remain large, even if the error rate is decreasing. In the foreseeable future, image analysis will never be 100 percent automatic—we will always need some element of human intelligence—but the process will speed up considerably.
***
It was the legendary inventor Doug Engelbart who first developed the idea of interacting with computers through a mouse. The full implications were not realized until the 1980s, when the personal-computer revolution swept the world. But Engelbart invented the mouse back in 1963, while directing a research team at the Stanford Research Institute, a California think tank. That same year, Marvin Minsky co-founded the Artificial Intelligence Laboratory (AI Lab) on the other side of the country, at the Massachusetts Institute of Technology. His researchers were among the first to confront the problem of making computers see.
Old-time computer hackers like to tell the story, perhaps apocryphal, of a meeting between these two great minds. Minsky proudly proclaimed, “We’re going to make machines intelligent! We’re going to make them walk and talk! We’re going to make them conscious!” Engelbart shot back, “You’re going to do that for computers? Well, what are you going to do for people?”
Engelbart laid out his ideas in a manifesto called “Augmenting Human Intellect,” which defined a field he called Intelligence Amplification, or IA. Its goal was subtly different from that of AI. Minsky aimed to make machines smarter; Engelbart wanted machines that made people smarter.
My laboratory’s research on machine learning belonged to the domain of AI, while the software program of Fiala and Harris was a direct descendant of Engelbart’s ideas. It was not AI, as it was not smart enough to see boundaries by itself. Instead it amplified human intelligence, helping humans analyze electron microscopic images more efficiently. The field of IA is becoming increasingly important for science, now that it’s possible to “crowdsource” tasks to large numbers of people over the Internet. The Galaxy Zoo project, for instance, invites members of the public to help astronomers classify galaxies by their appearance in telescope images.
But AI and IA are not actually in competition, because the best approach is to combine them, and that is what my laboratory is currently doing. AI should be part of any IA system. AI should take care of the easy decisions and leave the difficult ones for humans. The best way of making humans effective is to minimize the time they waste on trivial tasks. And an IA system is the perfect platform for collecting examples that can be used to improve AI by machine learning. The marriage of IA and AI leads to a system that gets smarter over time, amplifying human intelligence by a greater and greater factor.
People are sometimes frightened by the prospect of AI, having seen too many science fiction movies in which machines have rendered humans obsolete. And researchers can be distracted by the promise of AI, struggling in vain to fully automate tasks that would be more efficiently accomplished by the cooperation of computers and humans. That’s why we should never forget that the ultimate goal is IA, not AI. Engelbart’s message still comes through loud and clear for the computational challenges of connectomics.
These advances in image analysis are exciting and encouraging, but how quickly can we expect connectomics to progress in the future? We have all experienced incredible technological progress in our own lifetimes, especially in the area of computers. The heart of a desktop computer is a silicon chip called a microprocessor. The first microprocessors, released in 1971, contained just a few thousand transistors. Since then, semiconductor companies have been locked in a race to pack more and more transistors onto a chip. The pace of progress has been breathtaking. The cost per transistor has halved every two years, or—another way of looking at it—the number of transistors in a microprocessor of fixed cost has doubled every two years.
Sustained, regular doubling is an example of a type of growth called “exponential,” after the mathematical function that behaves in this way. Exponential growth in the complexity of computer chips is known as Moore’s Law, because it was foreseen by Gordon Moore in a 1965 article in Electronics magazine. This was three years before he helped found Intel, now the world’s largest manufacturer of microprocessors.
The exponential rate of progress makes the computer business unlike almost any other. Many years after his prediction was confirmed, Moore quipped, “If the automobile industry advanced as rapidly as the semiconductor industry, a Rolls Royce would get half a million miles per gallon, and it would be cheaper to throw it away than to park it.” We are persuaded to throw our computers away every few years and buy new ones. This is usually not because the old computers are broken, but because they have been rendered obsolete.
Interestingly, genomics has also progressed at an exponential rate, more like semiconductors than automobiles. In fact, genomics has bounded ahead even faster than computers. The cost per letter of the DNA sequence has halved even faster than the cost per transistor.
Will connectomics be like genomics, with exponential progress? In the long run, it’s arguable that computational power will be the primary constraint on finding connectomes. After all, the image analysis took up far more time than the image acquisition in the C. elegans project. In other words, connectomics will ride on the back of the computer industry. If Moore’s Law continues to hold, then connectomics will experience exponential growth—but no one knows for sure whether that will happen. On the one hand, growth in the number of transistors in a single microprocessor has started to falter, a sign that Moore’s Law could break down soon. On the other, growth might be maintained—or even accelerated—by the introduction of new computing architectures or nanoelectronics.
If connectomics experiences sustained exponential progress, then finding entire human connectomes will become easy well before the end of the twenty-first century. At present, my colleagues and I are preoccupied with surmounting the technical barriers to seeing connectomes. But what happens when we succeed? What will we do with them? In the next few chapters I’ll explore some of the exciting possibilities, which include creating better maps of the brain, uncovering the secrets of memory, zeroing in on the fundamental causes of brain disorders, and even using connectomes to find new ways of treating them.
10. Carving
One day when I was a boy, my father brought home a globe. I ran my fingers over the raised relief and felt the bumpiness of th
e Himalayas. I clicked the rocker switch on the cord and lay in bed gazing at the globe’s luminous roundness in my darkened room. Later on, I was fascinated by a large folio book, my father’s atlas of the world. I used to smell the leathery cover and leaf through the pages looking at the exotic names of faraway countries and oceans. My schoolteachers taught me and my classmates about the Mercator projection, and we giggled at the grotesque enlargement of Greenland with the same perverse enjoyment that came from a funhouse mirror or a newspaper cartoon on a piece of Silly Putty.
Today maps are practical items for me, not magical objects. As my childhood memories grow faint, I wonder whether my fascination with maps helped me cope with a fear of the world’s vastness. Back then, I never ventured beyond the streets of my neighborhood without my parents; the city beyond seemed frightening. Placing the entire world on a sphere or enclosing it within the pages of a book made it seem finite and harmless.
In ancient times, fear of the world’s vastness was not limited to children. When medieval cartographers drew maps, they did not leave unknown areas blank; they filled them with sea serpents, other imaginary monsters, and the words “Here be dragons.” As the centuries passed, explorers crossed every ocean and climbed every mountain, gradually filling in the blank areas of the map with real lands. Today we marvel at photos of the Earth’s beauty taken from outer space, and our communications networks have created a global village. The world has grown small.
Unlike the world, the brain seemed compact at first, fitting nicely inside the confines of a skull. But the more we know about the brain with its billions of neurons, the more it seems intimidatingly vast. The first neuroscientists carved the brain into regions and gave each one a name or number, as Brodmann did with his map of the cortex. Finding this approach too coarse, Cajal pioneered another one, coping with the immensity of the brain forest by classifying its trees like a botanist. Cajal was a “neuron collector.”
We learned earlier why it’s important to carve the brain into regions. Neurologists interpret the symptoms of brain damage using Brodmann’s map. Every cortical area is associated with a specific mental ability, such as understanding or speaking words, and damage to that area impairs that particular ability. But why is it important to carve the brain more finely, into neuron types? For one thing, neurologists can use such information. It’s less relevant for stroke or other injuries, which tend to affect all neurons at a particular location in the brain. Some diseases of the brain, however, affect certain types of neurons while sparing others.
Parkinson’s disease (PD) starts by impairing the control of movement. Most noticeably, there is a resting tremor, or involuntary quivering of the limbs when the patient is not attempting to move them. As the disease progresses, it can cause intellectual and emotional problems, and even dementia. The cases of Michael J. Fox and Muhammad Ali have raised public awareness of the disease.
Like Alzheimer’s disease, PD involves the degeneration and death of neurons. In the earlier stages, the damage is confined to a region known as the basal ganglia. This hodgepodge of structures is buried deep inside the cerebrum, and is also involved in Huntington’s disease, Tourette syndrome, and obsessive-compulsive disorder. The region’s role in so many diseases suggests that it is very important, even if much smaller than the surrounding cortex.
One part of the basal ganglia, the substantia nigra pars compacta, bears the brunt of degeneration in PD. We can zero in even further to a particular neuron type within this region, which secretes the neurotransmitter dopamine, and is progressively destroyed in PD. There is currently no cure, but the symptoms are managed by therapies that compensate for the reduction of dopamine.
Neuron types are important not only in disease but also in the normal operation of the nervous system. For example, the five broad classes of neurons in the retina—photoreceptors, horizontal cells, bipolar cells, amacrine cells, and ganglion cells—specialize in different functions. Photoreceptors sense light striking the retina and convert it into neural signals. The output of the retina leaves through the axons of ganglion cells, which travel in the optic nerve to the brain.
These five broad classes have been subdivided even further into over fifty types, as shown in Figure 38. Each strip represents a class, and contains the neuron types that belong to the class. The functions of retinal neurons are much simpler than that of the Jennifer Aniston neuron. For example, some spike in response to a light spot on a dark background, or vice versa. Each neuron type studied so far has been shown to possess a distinct function, and the effort to assign functions to all types is ongoing.
Figure 38. Types of neurons in the retina
As I’ll explain in this chapter, it’s not as easy as it sounds to divide the brain into regions and neuron types. Our current methods date back over a century to Brodmann and Cajal, and they are looking increasingly outmoded. One major contribution of connectomics will be new and improved methods for carving up the brain. This in turn will help us understand the pathologies that so often plague it, as well as its normal operation.
A modern map of a monkey brain (see Figure 39) brings back pleasant memories of my father’s atlas. Its colorings bear mysterious acronyms, and sharp corners punctuate its gentle curves. But maps are not always so charming. Let’s not forget that armies have clashed over lines drawn on them. Likewise, neuroanatomists have waged bitter intellectual battles over the boundaries of brain regions.
Figure 39. Map of the rhesus monkey cortex laid out flat
We already encountered Korbinian Brodmann’s map of the cortex. How exactly did he create it? The Golgi stain allowed neuroanatomists to see the branches of neurons clearly. Brodmann used another important stain, invented by the German neuroanatomist Franz Nissl, which spared the branches but made all cell bodies visible in a microscope. The stain reveals that the cortex (Figure 40, right) resembles a layer cake (left). Cell bodies are arranged in parallel layers that run throughout the entire cortical sheet. (The white spaces between cell bodies are filled with entangled neurites, which are not marked by the Nissl stain.) The boundaries in the cortex are not as distinct as those of the cake, but expert neuroanatomists can make out six layers . This piece of cortical cake, less than one millimeter wide, was cut from a particular location in the cortical sheet. In general, pieces from different locations have different layerings. Brodmann peered into his microscope to see these differences, and used them to divide the cortex into forty-three areas. He claimed that the layering was uniform at every location within one of his areas, changing only at the boundaries between areas.
Figure 40. Layers: cake (left) and Brodmann area 17, also known as V1 or primary visual cortex (right)
Brodmann’s map of the cortex may be famous, but it shouldn’t be taken as gospel truth. There have been plenty of other contenders. Brodmann’s colleagues in Berlin, the husband-and-wife team of Oskar and Cécile Vogt, used a different kind of stain to divide the cortex into two hundred areas. Still other maps were proposed by Alfred Campbell working in Liverpool, Sir Grafton Smith in Cairo, and Constantin von Economo and Georg Koskinas in Vienna. Some borders were recognized by all researchers, but others sparked discord. In a 1951 book Percival Bailey and Gerhardt von Bonin erased most of the borders of their predecessors, leaving just a handful of large regions.
Even worse controversies have plagued Cajal’s program to classify neurons into types. He did this based on their appearance, much as a nineteenth-century naturalist would have classified different species of butterflies. One of his favorite neurons was the pyramidal cell. He called it the “psychic cell,” not because he believed in the occult, but because he thought that this type of neuron played an important role in the highest functions of the psyche. In Figure 41, a drawing by Cajal himself, you can see the defining features of this neuron type: the roughly pyramidal shape of its cell body, the thornlike spines protruding from its dendrites, and the long axon traveling far away from the cell body. (The axon is directed downward in this image, des
cending into the brain. The most prominent “apical” dendrite leaves the apex of the pyramid and travels upward, toward the surface of the cortex.)
Figure 41. Drawing of a pyramidal neuron by Cajal
The pyramidal cell is the most common type of neuron in the cortex. Cajal observed other cortical neurons that had shorter axons, and smooth rather than spiny dendrites. The shapes of nonpyramidal neurons were more diverse, so they were divided into more types, which earned picturesque names such as “double bouquet cell.”
Cajal classified neurons into types all over the brain, not just in the cortex. This alternative method of carving up the brain was much more complex than Brodmann’s, because every brain region contains many neuron types. Furthermore, the types in each region are intermingled, like different ethnic groups living in the same country. Cajal could not complete the task in his lifetime, and even today this enterprise has only just begun. We still don’t know how many types there are, though we know the number is large. The brain is more like a tropical rainforest, which contains hundreds of species, than a coniferous forest with perhaps a single species of pine tree. One expert has estimated that there are hundreds of neuron types in the cortex alone. Neuroscientists continue to argue over their classification.
Their disagreements are a sign of a more fundamental problem: It’s not even clear how to properly define the concepts of “brain region” and “neuron type.” In Plato’s dialogue Phaedrus, Socrates recommends “division . . . according to the natural formation, where the joint is, not breaking any part as a bad carver might.” This metaphor vividly compares the intellectual challenge of taxonomy to the more visceral activity of cutting poultry into pieces. Anatomists follow Socrates literally, dividing the body by naming its bones, muscles, organs, and so on. Does his advice also make sense for the brain?
Connectome Page 18