without a trace

Welcome to the final installment of "Brainiacs Week" at the cocktail party! Okay,it was more like a leisurely "Brainiacs Fortnight," thanks to the LHC-related Doomsday distraction, and the usual round of activities at KITP. Some of us need our cookie hour, and that's gonna cut into blogging time, along with that impulsive shoe-shopping expedition with the Spousal Unit over the weekend. I really needed new shoes. I probably didn't need the pricey stiletto black boots with fetching cobweb pattern etched in steel gray, but the Spousal Unit voiced his enthusiastic approval of the purchase, and even offered to assist me in selecting a new outfit to go with them. The selflessness is truly breathtaking.

Anyway, there are far too many interesting talks  on cutting-edge neuroscience going on right now to do justice to all of them in a handful of blog posts; I encourage interested readers to check out the entire collection of online Brainiac talks, not just those mentioned specifically by me. In today's monster post, we're going to talk -- fairly generally -- about brain mapping: specifically, how our ideas about the structure of the brain have evolved in the last 100 years or so.

We've come a long way since the 19th century German physician Franz Joseph Gall made the rounds of polite society, spreading his pseudoscientific notions of phrenology. Phrenology, for those who aren't familiar with it, hypothesized that the brain was divided into dozens of "personality organs", and the shape of the skull showed evidence of bony "bumps" that could be "read" tactically by Gall to determine the personality traits of individuals. It turned out to be nonsense of course, as did the related field of craniometry, which attempted to correlate the overall size and shape of the skull with intelligence and morality. There was some interesting work -- and occasionally icky, yet fascinating early neurosurgery experiments (for those of us with morbid interests) -- being done by anatomists on the physical structure of the brain, but the specifics of how this most central organ actually functioned remained shrouded in mystery.

There were lots of theories, of course, as to how the physical brainy organ and what one might call "the mind" were connected. Those of a Pavlovian bent reduced everything to simple stimulus and response, believing that signals from the sensory organs traveled (one way) to the brain via a simple path, inducing response. Today we know that while the neurons themselves transmit in just one direction, the connections between them are much more complex, and not necessarily one-way. The Pavlovian model lost a bit of credibility in 1929, when Hans Berger showed that there was continuous electrical activity occurring in the mind -- even without much stimulus.The Pavlovian picture was far too simplistic. Behaviorists did their best, but their theories simply weren't sufficient to account for how the brain processes patterns.

Along came Canadian psychologist named Donald Hebb, who penned the following revolutionary words in 1949 in his pioneering work, The Organization of Behavior: "Let us assume that the persistence or repetition of a reverberatory activity (or 'trace') tends to induce lasting cellular changes that add to its stability... When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased." Or, to phrase it more bluntly: "Cells that fire together, wire together." This was the culmination of years of research by Hebb, including the effect of brain surgery and injury on the human brain. For instance, in the 1930s, Hebb was intrigued to observe that the brain of a child could regain partial or full function even if a portion of it were removed, but an adult would be far more adversely affected by similar damage. He concluded that external stimulation played a prominent role in the thought processes of adults. Hebb's understanding was admittedly incomplete, but the theory he cobbled together got many of the brain's most important functions right. He was also correct that learning linked neurons in new ways and thus the neuronal structure could be altered by experience.

Hebb's ideas have influenced many a modern neuroscientist, including Michael Merzenich (who has a blog!), one of the first to argue that the brain was plastic. Merzenich has performed all kinds of experiments with monkeys that would make the PETA crowd apoplectic. For example, he cut a monkey's median nerve and then did multiple mappings of brain activity over several months. Within 144 days, a completely new map of the brain emerged in as much detail as the original ("normal") map.

He has also mapped a normal monkey's brain, then sewed together two of the animal's fingers so they could only move together. After several months, the monkey's brain remapped: the two maps of the original separate fingers had merged, so if experimenters touched any point on either finger, the new single map would light up. A similar effect was observed in the brains of humans people born with their fingers fused (syndactyle or webbed-finger syndrome). Such experiments might seem cruel to the monkeys, but as with most animal testing, they serve a nobler purpose: Merzenich believes there is hope that those born with learning or psychological problems, or who have suffered brain damage from stroke or trauma -- might be able to form new maps/neural connections as well, if neuroscientists can figure out how to mimic the process by which neurons fire together and wire together.

To date, most brain mapping efforts have been on more of a macroscale: identifying which parts of the brain are affiliated with specific functions, for example, or staining single neurons to track them in the mass of brain tissue, or looking at thicker "wiring" that connects different parts of the brain. Ideally, neuroscientists would like to trace the actual "wiring" of the brain: the dendrites and axons that form the synaptic connections between neurons. All the cool kids call this the "connectome." So does MIT's Sebastian Seung, who likens the progression of brain mapping in neuroscience to the mapping of the human genome. Specifically, the discovery of DNA's double helix pattern in the 1950s was a major breakthrough, but still pretty general in terms of information about the specific structure and associated functions of individual genes. It was only when the Human Genome Project succeeded in mapping the entire gene sequence that scientists were able to start really exploring the power of genetics. The field has literally exploded with exciting developments since then.

Neuroscience seems to be at the 1950s level, per Seung: we understand there are is a vast, tangled network of single neurons connected by synapses, but we haven't  yet mapped out each connection separately. (I'm using "we" rhetorically, of course; I have contributed squat to neuroscience.) Only then would we have a complete map of the brain capable of revealing how neural networks actually perform specific functions. More importantly, it could shed some light on the mechanisms of certain brain disorders commonly attributed to "faulty wiring," such as autism or schizophrenia. "A lot of properties of brain function are at the level of the circuit -- information is being integrated, processed, extracted," MIT neuroscientist Elly Nedivi told Technology Review in a recent article. "To understand what that means, you need to be able to see who connects to who."

Unfortunately, just like mapping the human genome, this is a tough nut to crack. First, the thinnest axons are roughly 100 nanometers in diameter, so neuroscientists require nanoscale resolution in their imaging techniques in order to get the detail they need to produce any kind of useful "map." (Complicating matters even further is the fact that you need nanoscale imaging at different length scales: dendritic trees measure in the micron range, while axons can extend to several centimeters.)

Second, it's not like the neurons and their connections are neatly laid out in easy-to-navigate grids; they're a dense, tangled mess, like so many spaghetti strands, so even if you get a good enough high-resolution image, it's still tough to trace the various connections. Most of us have a hard enough time unraveling all the tangled power cords to the plethora of electronic devices now featured in every household. Heck, just the twin cords to the earbuds for my iPod Shuffle are always getting tangled up, necessitating a profanity-laced time-out to fix things before embarking on my workout. It's so much harder to unravel the brain's circuitry: the human brain has 100 billion neurons, and 100 trillion synapses! Winfried Denk of the Max Planck Institute in Heidelberg, Germany, estimates that just producing a wiring diagram for a single cortical column -- a single unit of neurons in the cortex -- would take about 3 billion years using today's conventional methods. After all, C. elegans (the humble nematode) has a paltry 302 neurons, yet it took researchers over 10 years to compile a "circuit diagram" for the little worm -- earning them a well-deserved Nobel Prize in the process.

Denk's lab has made great strides on the nanoscale imaging front in recent years with the development of a new technique called serial block-face scanning electron microscopy. You take a small block of brain tissue (not just a thin slice) and bounce electrons off the top, which produces a cross-sectional image of the nerve fibers in that slice of brain tissue. Then you shave off a thin 30-nanometer slice off the top, and repeat the process. You do this again and again, thousands of times, shaving thin slices off the top and imaging them, slice by slice, to trace the path of each piece of nerve fiber through the  block of brain tissue. You end up with a 3D image of that brain tissue at nanoscale resolution. Researchers at Harvard University have developed a complementary technique that traces neurons in the living brain by labeling them many different colors. Taken together, the two techniques provide two different perspectives on neural circuitry.

Even once you have those nanoscale images, however, your hard work is not yet done. You've got to do the tracing by hand. Denk takes a factory approach to the problem, amassing  a small army of dedicated undergraduate physics students to laboriously pore over the resulting images and assign colors to specific neurons, axons and dendrites in the images. It takes about 40 hours to trace one tiny square, assuming all goes smoothly. (Apparently, neurons are identified on the basis of shape.) Branching complicates the process. Sometimes two students will each be tracing what they think are two different objects, only to find that at some point they merge -- they have actually been tracing the same thing. So they've got to go back and change everything to reflect that unsuspected connection.

I'm pretty sure the students were hoping for a more exciting task when they signed on to work in a leading neuroscience lab. But heck, we all have to pay our dues, regardless of our chosen profession. And there's good news for all those sleep-deprived German undergrads! The days of painstakingly mapping out the brain by hand could well be over, if Seung and his colleagues succeed in their mission to automate the process. Seung's professional background is pretty interesting: he came to neuroscience by way of condensed matter physics theory, working on artificial neural networks (ANNs). That early interest served as a natural segue into neuroscience, and he's drawing on that expertise in his current collaboration, which includes Denk's group in Heidelberg. His goal is nothing less than to transform the field of neuroanatomy into a "high throughput, data-rich field of science," via the creation of automated systems that can take a sample of brain tissue as raw input and generate a complete circuit diagram as output.

Seung describes this latest work as "an image processing problem of unprecedented scale, because a sample of even modest dimensions yields a huge amount of data at nanoscale resolution."  He's not kidding: we're talking terabytes and petabytes of data just to  produce a wiring map of a fruitfly,  never mind a human brain. Achieving the complete fruitfly "connectome" would constitute success beyond most people's wildest dreams, and make it more likely that neuroscientists could achieve, in our lifetime, circuit diagrams for certain critical locations in the human brain: the hippocampus, for example, or the olfactory bulb and retina.

So, Seung is not afraid to aim high, while acknowledging the staggering complexity of the task he's set himself. First, he's taking the tracings created by Denk's undergrads and performing something called binary image restoration: turning analog images into binary images and using those to teach a machine to perform the same task as Denk's students. The resulting algorithm is then used to analyze new chunks of brain tissue.

Basically, Seung is taking the standard learning process and turning it into an optimization problem. Per Seung's Website: "Many types of biological learning can be regarded as optimizations," for example, in operant conditioning, when animals adapt their behavior to maximize their reward. We also have the "practice makes perfect" notion of learning in humans, whereby repetition of complex motor skills -- playing the piano, for example, or learning to surf -- improves those skills with each iteration.

Here's Seung again: "It is widely believed that long-lasting modifications of synaptic connections are responsible at least in part for ... learning. We are interested in the hypothesis that one function of synaptic plasticity is to perform the computations required to optimize neural circuits." Based on this principal, he and his team came up with some working synaptic "rules" that enable artificial synapses to estimate the gradient of the expected reward to achieve maximum reward -- something called stochastic gradient learning by computer scientists. (As a budding calculus student, I believe this means that they are essentially plotting expected reward with degrees of neural activity, then looking at the gradient of the resulting curve to find the optimal reward. Or something like that. Y'all can check out Seung's Website and online talk at the KITP Website for further details.)

So far, Seung and Denk's collaboration has yielded a partial wiring diagram of the rabbit retina, using machine learning algorithms combined with the 2D images generated by Denk's serial block-face SEM to automate the generation of the images. There's a nifty online mini-movie on the Technology Review site that's definitely worth checking out. Go ahead and watch, we'll wait.

Are you back? Good. You just saw a 3D animation of the detailed wiring map of part of the rabbit retina called the inner plexiform layer. That's the little piece of neural tissue at the back of the eye that senses light and sends visual information to the brain. A single neurite appears first, shown in green, followed by a larger subset of neurites in multiple colors. The result is a veritable "brain forest" as the animation traces the "wires" through the dense brain tissue.

However, a human operator still needs to provide oversight to correct the machine's errors (currently in the range of 7% to 8%, according to Seung.) Plus, it's still pretty slow: right now the system has some 34,000 parameters, and even with stochastic gradient learning, it takes a long time to train the computer to recognize that many. They'll need to speed up the automated process by a million times or more to generate larger maps -- like a wiring diagram for the cortical column. If they can do that, they would reduce the time to trace the cortical column from 30 billion years to about two years. (It almost makes the challenge sound manageable. Context is everything.) Also, while Seung acknowledges that the movie looks cool and is "impressive PR," he also points out that it doesn't impart much information that would be considered useful to neuroscientists. Not only are there too many errors, but only fragments of neurons are captured in the sample, making it  very difficult to determine their shape, and hence identify them properly.

Still, 100 years from now, we might know far more about the intricate wiring of the brain than we do today, thanks to those like Seung, Denk, et al who fight the good fight against seemingly impossible odds. And that concludes "Brainiac Fortnight." For those of you who made it this far: thanks for reading. You are troopers!

Comments

[Ellen]

I've done an article on Gall, and phrenology, which you may find interesting:

http://washuu.net/Med-Lec/psycgraf.htm

[chezjake]

I've read the whole neuroscience series here, and I think you've done an excellent job of putting it across to the educated layperson.

While I have no complaints at all about your physics blogging, I think this is an indication that you shouldn't necessarily limit yourself to physics. Go wherever your curiosity and whim take you.