by Paul Allen
Over the last decade, I began to think about a “Digital Aristotle,” an easy-to-use, all-encompassing knowledge storehouse. I wasn’t aiming to solve the mystery of human consciousness. I simply wanted to advance the field of artificial intelligence so that computers could do what they do best (organize and analyze information) to help people do what they do best, those inspired leaps of intuition that fuel original ideas and breakthroughs.
That’s why we began Project Halo, a research program that is trying to create a Digital Aristotle. One near-term goal is Halobook, an electronic textbook that can answer typed-in questions with expert-level accuracy. Running on a laptop or tablet, Halobook could serve as a research aide for working scientists or as a tutor for college and high school students, like a personal digital teaching assistant.
In the inaugural Halobook, targeted for release in 2015, we’ll have encoded most of the Advanced Placement biology syllabus, something no one else has yet attempted. After that, we may reach into biochemistry or move to a whole new area, like civil engineering. We might even take on economics or U.S. government. The humanities—philosophy, religion, history, classics—would be much, much tougher. As subject matter shifts from how things work to the values and language that define the human condition (fairness, morality, love), software systems quickly move out of their depth. I recognized this roadblock as early as 1977, in my interview with Microcomputer Interface:
In order to be truly intelligent, computers must understand—that is probably the critical word. It is one thing to feed The Tale of Two Cities into a computer. It’s another to have the computer understand what’s being said. You can’t ask it a question about the theme of a book or why a character does something and get a coherent answer. We haven’t yet reached that level with intelligent computers.
And we still haven’t today, but we’re getting closer. (For a further explanation of Project Halo, see the appendix.)
Ultimately, a Digital Aristotle should make us more inventive and creative. With its steady progress in attacking classic problems like learning, language, and reasoning, I can foresee a time when artificial intelligence could greatly speed our ability to ferret out cures for diseases or help us preserve the environment. As Douglas Engelbart wrote in 1962 in Augmenting Human Intellect: A Conceptual Framework:
Man’s population and gross product are increasing at a considerable rate, but the complexity of his problems grows still faster, and the urgency with which solutions must be found becomes steadily greater. … By “augmenting human intellect” we mean increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems. …
One of the tools that shows the greatest immediate promise is the computer, when it can be harnessed for direct online assistance, integrated with new concepts and methods.
As computing grows increasingly cheaper and more powerful, it is now conceivable that virtually all the world’s data will soon be found online. Organizing it coherently and logically will take a Herculean effort. Given the ever-accelerating expansion of human knowledge, not to mention its breadth and complexity, a final encyclopedia is an elusive goal. But it just might be closer than you think.
IN HIS ESSAY “The Law of Accelerating Returns,” the futurist Ray Kurzweil predicted that the increase in computer processing power will soon lead to a “singularity” of technological change “so rapid and profound it represents a rupture in the fabric of human history.” Kurzweil foresees the imminent arrival of “strong AI,” machines as smart as human beings, the first step in an accelerating progression of smarter and smarter machines—to the point that we’ll be able to download our personalities and self-awareness into computers and gain a sort of digital immortality.
Though I won’t say that a singularity is impossible, I believe that it is centuries away at best. Although Kurzweil credits Moore’s law as an inspiration, Gordon Moore agrees with me, noting that human development “involves a lot more than just the intellectual capability,” and doubting that machines “could overcome that overall gap. …” The sheer complexity of human brain function is daunting in the extreme. It took forty years to develop a computer chess program that could consistently beat the best human players, even though grandmaster-level chess can be achieved with simple sequential logic and brute-force processing. To get a computer to read and understand human language is incomparably harder. We can’t replicate the brain because we’ve barely begun to understand how it works.
There are two basic approaches to artificial intelligence, both of them journeys of thousands of small steps. You can take the Halo approach, in which we’re inventing software to emulate some of the things the brain can do. Or you can try to reverse-engineer the physical brain itself to see how it really functions, which is the story of an institute I founded in Seattle.
CHAPTER 21
MAPPING THE BRAIN
The brain is a never-ending source of fascination for me. It’s the organ that unites us as a species and distinguishes us from one another. It keeps us breathing and upright, makes us elated or anxious, and, not least, harbors our creativity. Here is a truly astonishing piece of evolutionary engineering, one that does so many things so much better than the most advanced computer, and yet we’re just scratching its surface.
The 1990s saw an explosion of new theories in genomics, informatics (the conversion of data into usable information), and molecular neurobiology. In February 2001, eleven years after it began, the Human Genome Project released a first draft of the roughly 3 billion base pairs that comprise our genes. The HGP confirmed that fewer than twenty-five thousand human genes were needed to create the brain’s 100 billion multifaceted nerve cells, all linked in intricate networks totaling a quadrillion neural connections. How could such a small genome serve as the blueprint for such a complex organ? And how might the HGP’s achievement be used to advance neuroscience?
I was meeting around that time with experts in early learning and linguistics, and their research was engaging, but I felt a pull to get inside the human brain, the ultimate machine. To think it through, I met with Jim Watson, the director of the Cold Spring Harbor Laboratory and codiscoverer of the double-helix structure of DNA. Then in his seventies, Watson was a confirmed iconoclast, always ready to venture off the beaten path. He proposed that I found a behavioral research center to focus on gene expression in the brain, the phenomenon of different genes “switching on” in different cells. The cell’s “expressed” genes direct the production of particular mixes of proteins, which in turn differentiate heart cells from skin cells (or tumor cells) and control how they function.
I also met with Steve Friend, the founder of a cutting-edge Seattle genome-analysis company. He, too, thought the time was ripe for a facility at the crossroads of human psychology, genomics, behavioral genetics, and brain biology. Advances in data storage and retrieval would enable us to compile and analyze the masses of new information we gathered.
“The more I read about the brain the more fascinated and interested I am,” I wrote in a December 2000 e-mail that included my first mention of a brain institute. “I am especially interested in how the genetic ‘blueprint’ for building the brain works.”
MOST NEUROSCIENCE RESEARCHERS are highly specialized, pursuing their questions in discrete areas of the brain as though they’re drilling into an orange with a needle. I wanted to cover the entire rind and help scientists locate the most promising spots to drill, to get them probing faster and deeper that much sooner. In March 2002, I invited twenty-one scientists, including four Nobel laureates, to join me at a three-day brainstorming session, or charrette. The scientists assembled at a dock in Nassau in the Bahamas and ferried over to our conference center for the weekend, my yacht, Tatoosh, a serene setting for an intensive discussion.
In addition to Watson and Friend, the guests included Richard Axel, the Nobel Prize–winning neuroscientist who’d advanced the understanding of
our sense of smell; Steven Pinker, the Harvard psychologist and bestselling author of books on linguistics; Marc Tessier-Lavigne, who did pioneering work on the assembly of the embryonic and fetal brain; Lee Hartwell, who won his Nobel for discovering the genes that control cell division; and David Anderson, a Cal Tech neurobiologist who’d play an instrumental role in defining our mission.
I came to the charrette with a rough vision of an institute on the frontier of brain science. One expert suggested that I establish a top-tier research facility, on par with the Rockefeller Institute, to recruit the best and brightest researchers from around the world. The price tag: $1 billion, half to start up and half to endow.
Money aside, I was wary of the traditional academic model for research institutes. Scientists of stature pursue whatever they find most interesting and are not easily steered. I had seen the downside of a loosely defined organizational mission at Interval Research, which Vulcan had closed two years earlier, and I wasn’t eager to repeat it. The alternative was to concentrate on a single large-scale endeavor that might transform the field, a neuroscience equivalent to the Human Genome Project. We’d have concrete milestones en route to tangible results within a few years. I wanted a facility run on an industrial scale, with biotech urgency but without the profit motive.
My guests debated what exactly our institute should address. The discussions were lively, wide-ranging, and often competitive; great scientists are adept at putting forward their proposals. (It’s what they do to get their funding grants renewed.) All sorts of ideas were floated. Was there an underlying genetic basis for happiness, or for love? How could we improve brain-imaging technology? What single disease might be most usefully explored?
By the second day, the conversation kept circling back to the idea that had first surfaced in my talks with Watson and Friend. What neuroscience needed more than anything else, I kept hearing, was something very basic: a better map of the brain.
In existing maps relating to gene expression, the anatomic resolution was too coarse to be of much help in deciphering how the organ really worked. The National Institutes of Health had recently funded a Brain Molecular Anatomy Project, but the work was too fragmented for consistent quality control, and there was only enough funding to look at six hundred genes per year. At that rate, a complete atlas might take half a century. Brain mapping was stuck in a cottage-industry stage, like the one that hobbled genome sequencing before the HGP and Craig Venter’s Celera Genomics made the effort systematic.
By our closing session, the scientists were unanimous. A brain atlas was “an appropriate inaugural project for the Allen Institute because of the incalculable contribution the Atlas can make towards solving basic molecular and genetic questions about human behavior.” The atlas would link genetics and anatomy, with maps of switched-on genes overlying the brain’s three-dimensional structure. It would open new avenues of research into neurological and psychiatric disorders, as well as fundamental questions of brain science. Our initial effort, the scientists agreed, should map the adult mouse brain and focus on healthy specimens. (Most studies were then looking at embryonic brains, and NIH research emphasized diseases.)
A brain atlas of gene expression fit my main criterion, to go where important work lay undone. It was “big science” with obvious real-world utility. From Alzheimer’s and Parkinson’s to schizophrenia, brain disorders afflicted tens of millions of Americans. Once we mapped a normal “reference” brain, we’d be able to isolate the active genes that triggered these ailments. Scientists could begin to find ways to target them therapeutically. The potential was staggering.
And far down the path, I thought, our work might even help uncover the essence of memory, desire, compassion—of what makes us human.
I LOVE TO travel with close friends and family. My mother liked Tahiti and Japan, though she was less fond of Africa after a hippopotamus broke into the compound she was staying in and had to be rope-lifted out of the swimming pool. Her favorite trip was a plantation tour on the Mississippi, where she paused on people’s porches to share iced tea and listen to their stories, just as she’d once lingered with her neighbors on her way home from school in Anadarko. She was still the best listener I’d ever known.
Over the years, I had bought up land around my home on Mercer Island, adding houses for my mother and sister. My mom thrived there amid her fifteen thousand neatly shelved books, most of them bought for a quarter or less at a thrift store. Nothing gave her more pleasure than perusing her stacks, meticulously organized by authors’ last names, and finding an old friend.
My mother used to lead a book club for faculty wives from UW, choosing works by African authors one year, Eastern European novels the next. She took almost as much joy in selecting as in discussing. When she set out to compile a list of 100 favorites for me, she wound up with 165—as she’d often say, “What’s better than a good book?” But her days as a reader were numbered. In a late-night e-mail journal entry on January 21, 2003, I wrote, “My mother is struggling right now with an Alzheimer’s-like condition. (She was diagnosed in the last two weeks.) I’m sick at heart about this.”
Her dementia was subtle at first. One minute she’d finish a crossword puzzle with ease, and the next she’d forget what she’d told me minutes before. She was angry about her memory loss; she’d reached that wrenching phase in which she knew she was slipping but felt powerless to stop it. Then came a long, slow twilight with gathering darkness. I saw the horrors of Alzheimer’s up close, and I was devastated. If there was anything I could do to spare others a similar fate, I was determined to try.
I turned fifty the day I wrote that entry, the point in life where many of us begin to consider what we’ll be leaving behind. In September 2003, I launched the Allen Institute for Brain Science with a $100 million contribution. Its charter was ambitious: “We believe this is a historic opportunity to unite the genome and the brain, and use the data and technology to tackle the challenges of neurodevelopmental, neurodegenerative and psychiatric disease.”
We found a facility in Fremont, a peaceful Seattle neighborhood perched above a ship canal. There was space for our whole staff under one roof: process engineering, molecular biology, anatomy, software development, database creation. As president and chairman of the board, my sister would once again oversee my brainchild. A blue-ribbon group of scientists, including several strong voices from our charrette, would serve on an advisory board.
The highest hurdle for a brain atlas was the sheer amount of data to be collected and organized. Where the HGP’s data consisted of sequences of letters, ours would be high-resolution images, which needed far more storage space. The initial mouse brain atlas would involve 85 million images on 250,000 slides—600 terabytes of data (600,000 gigabytes, or 600 trillion bytes), or more than half as much as the total content of the Internet when we began.
Early on, we confronted a pivotal issue. Should we charge for access to our database? Revenue from users and royalties from commercial work could help expand our operation. On the other hand, the institute’s success had to be measured by the discoveries it sparked. The more widely the atlas was used, the greater the chance of a breakthrough. Charging for access might limit use to elite universities and the largest pharmaceutical firms, while shuting out some talented researcher in Johannesburg or Seoul who couldn’t come up with the fee. We decided to place our data in the public domain, with free Internet access and a powerful, user-friendly interface. No registration would be required.
POSTMORTEM HUMAN BRAINS are all very different. The donors vary in age, genetic backgrounds, and upbringing, all variables that shape the organ’s form and function. And so, like countless human-oriented studies before us, we opted to start with the ideal laboratory mammal, the mouse. The mouse brain is no larger than an almond, no heavier than a teaspoon of sugar. But it’s a terrific template for mapping. It closely resembles our own brain in both form and content, with 90 percent of a mouse’s genes having a human counterpart. Inbre
eding would give us close-to-identical subjects at a uniform age of eight weeks, a near-perfect experimental system.
We chose a state-of-the-art hybridization technique developed at the Max Planck Institute in Germany and later implemented at Baylor College of Medicine. The mouse brains would be sliced north-to-south into hundreds of sections, then dunked into an RNA solution to probe for a specific active gene—one gene per slice, five or six genes per brain. All neurons expressing that gene would be revealed.
The scope of an all-gene atlas demanded an intensively choreographed, high-throughput approach to convert hundreds of thousands of slides into digital data. Modeling our work after Baylor’s, we organized laboratory robots in assembly-line fashion, staining slides around the clock (up to four thousand per week), photographing the sections under microscopes, and channeling the images to our database.
A year after we launched, I promoted Allan Jones, who had overseen the collaboration with Baylor and recruited much of our staff, to run the project. It was a big promotion, but Allan quickly proved up to the task. In December 2004, we released the first installment of the Allen Brain Atlas: visual data from nearly two thousand genes. David Anderson, the advisory board member who first proposed the atlas project, calculated that we’d accomplished over fourteen months what might have taken a solitary scientist seventy-seven years.
A database is only as good as its interactive search functions. If people can’t find and download what they’re looking for, it’s like a vast library with no call numbers. In structuring the mouse brain atlas database, we created software that answers both the “where is” and “what is” questions. Say, for example, you are looking at a gene that increases sensitivity to painkillers like morphine, and want to know just where it is switched on. First you’ll be taken to the high-resolution, two-dimensional data for that gene at the cellular level. Then a viewing application, the 3-D Brain Explorer, will show how the gene is distributed across the “consensus” brain of all mice used in the study. Or, if you prefer, you can type in amygdala (the region of the brain that governs fear and anticipation) and see which color-coded genes are active in that area. Either way, for the very first time, you will have access to gene-expression data at the cellular level.