Ray Kurzweil on Hierarchies, the Neocortex and Merging with AI

Speaking at MIT's course on artificial general intelligence, Ray Kurzweil recalls becoming excited about AI at age 14 and writing to Marvin Minsky, who spent a whole day with him. He notes that the field had already split into symbolic and connectionist camps, and that Minsky actually invented the neural net in 1953 before turning negative on it amid hype. Kurzweil then visited Frank Rosenblatt at Cornell, whose perceptron recognized printed letters in one typeface but failed on others, and Rosenblatt's answer was to stack more layers.

Rosenblatt died nine years later without trying it, Kurzweil says, though the idea proved prescient. Today's deep learning excitement comes from many-layer nets and the law of accelerating returns, the exponential growth of computing that lets us run massive nets over massive data. A fix for the vanishing and exploding gradient problem is what allowed networks of 200 layers, and he says it underpins the recent gains.

He describes AlphaGo, which first trained on online games and became a decent player, then played itself and soared past the best humans, while AlphaGo Zero started with no human input at all. The catch is that, as the motto goes, life begins at a billion examples. Self-play works when the world can be simulated on a page or two, as in Go, chess and math, but is harder for biology, where simulators are imperfect.

Autonomous vehicles are the middle case: Waymo's few million real-world miles were enough to build a realistic simulator that then produced a billion simulated miles. Kurzweil notes that humans can learn from a handful of examples, which he calls the remaining advantage of humans. He adds that the human brain has no backpropagation and uses a different architecture.

In 1962 Kurzweil wrote a paper on how he thought the brain works, with little neuroscience to go on beyond Vernon Mountcastle's observation that different neocortical regions look the same under autopsy, with the same repeating pattern. Common wisdom held that visual cortex, face recognition and language areas must work on different principles; Mountcastle's view was that neocortex is neocortex. Kurzweil adds an observation from introspection: we recite the alphabet as sequences of sequences, forward only, and cannot do it backward without learning a new sequence.

His proposal was that the neocortex is organized as a hierarchy of modules, each learning a simple pattern, which he developed over a half-century and wrote up in a 2012 book. He says neuroscience evidence now supports it: the European brain reverse-engineering project identified a repeating module of about a hundred neurons, repeated some three hundred million times across roughly 30 billion neurons. Each module's output feeds another module's input, giving a hierarchy defined by connections rather than physical layout.

He says the neocortex is a thin structure, one module thick with six layers of neurons, and that the pattern recognition each module uses is something like a hidden Markov model rather than backpropagation. Markov models handle local variation, as in the speech recognition he built in the 1980s, while long-distance relationships come from the hierarchy. How the neocortex forms the hierarchy is not fully understood, but connections exist from birth like Manhattan's streets and avenues and appear to be activated, which he says can be seen on non-invasive scans.

Kurzweil tells the evolutionary story. The neocortex appeared about 200 million years ago with mammals, whose ability to invent new behaviors, unlike the fixed behaviors of other animals, is a signature of creativity. A mouse whose escape route is blocked can invent a new one, remember it, and spread it virally through the community. For some 135 million years that did not help early mammals much.

Then, 65 million years ago, the Cretaceous extinction event wiped out the dinosaurs and about 75% of animal and plant species, and mammals took over their niches. To anthropomorphize evolution, he says, it decided the neocortex was good stuff and grew it, with a larger fraction of the brain and a curved surface in primates. Stretched out, he notes, the human neocortex is still a flat sheet about the size of a table napkin.

About two million years ago, evolution added perhaps 20% more neocortex, which filled the frontal lobe. That tripled the number of levels in the hierarchy, he argues, and enabled language, art and music. Technology needed another adaptation, the opposable thumb, which let us act on creative ideas, make tools and then use tools to make new tools.

Kurzweil says Larry Page read his book in 2012 and agreed to invest in his young company, saying Google had the data, computers and talent; Kurzweil instead took his first job there, about five years before this talk. Since then he has been applying the hierarchical model to understanding language, which he considers the holy grail of AI. He agrees with Turing that text conversation is a Turing-complete problem, with no simple tricks for passing a valid test.

He describes the progress of paragraph comprehension: his team aced a first-grade test and had average results on the second, but the third grade needed too much common-sense inference. Last week, he says, two systems slightly surpassed average human performance on an adult-level test, which he calls an impressive milestone. In his experience, once an AI reaches average human level it does not take long to soar past it, though he expects language to take longer than Go did.

His team of about 45 people builds a hierarchical model using deep learning in each module, with embeddings for words and sentences and context taken into account, a published approach that powers Gmail's Smart Reply. He thinks deep learning is missing the hierarchical aspect because the world is hierarchical, which is why evolution built a hierarchical brain.

He lists two problems with big deep nets. They need a billion examples, which language has in unannotated form, creating a chicken-and-egg problem, and they do not explain themselves. Moves like AlphaGo's queen-and-bishop sacrifice or an early corner stone in Go were shocking and brilliant but unexplained, whereas in a hierarchy you could inspect each module's content to see what it is doing.

Kurzweil expects AI to hit health and medicine hard and says we are only about a decade from longevity escape velocity, when more time is added to your remaining life expectancy than passes. He thinks a diligent person may already be there, and says he believes he is. He points out that life expectancy figures assume little scientific progress, which is wrong given weekly advances in AI and biotech, and you can still be hit by the proverbial bus, which self-driving cars are meant to address.

In Q&A, asked about combining neural and symbolic approaches, he says he does not want symbolic models as they have been used, citing the Cyc project, which became impossible to debug as fixing one thing broke three others. Instead a connectionist system can be reliable enough to look like a rule while still capturing soft edges and exceptions, so we can learn from accumulated human wisdom that looks symbolic. He wants those systems to develop hierarchy rather than be one massive net.

Asked about brain regions below the neocortex, he says the cerebellum has more neurons than the neocortex and stores simple scripts, like a signature, though much movement has migrated to the neocortex. People without a cerebellum function fairly normally, if a bit clumsily. For intelligence, he says, the neocortex is the region to study.

Asked what happens after the singularity, Kurzweil says computation could go on well past 2045, to trillions of trillions of times today's capacity under molecular computing limits. On jobs, he compares today's fears to the Luddites of the early 1800s and to 1900, when about 38% of workers were on farms and 25% in factories. Nobody then could have named the new jobs, and today's are similarly hard to foresee, as shown by mobile apps and data analytics.

He says employment has risen from about 30% to 45% of the population, new jobs pay about eleven times as much in constant dollars, and people increasingly want careers with purpose, moving up Maslow's hierarchy. We could not do these jobs without enhancing intelligence, which we did through education, from 38,000 college students in 1870 to 15 million today, and now through brain extenders such as smartphones that will eventually go inside our bodies and brains.

That, he says, is the answer to movies about a brave band of humans fighting AI for control: we already have billions of smartphones, are deeply integrated with AI and will keep merging with it.

Asked about examples that are not exponential, Kurzweil says price-performance and capacity of information technology reliably follow exponentials, while social effects can be linear, such as the steady growth of democracy from a handful of countries to dozens. Chess scores look linear only because they are logarithmic measures.

Responding to a question on Heidegger and the meaning of technology, Kurzweil says we first used technology to shore up weaknesses, from muscles to multiplying numbers, and the promise of AI is to master the unique strengths of humanity, such as loving expressions in poetry and music. He thinks of technology as an expression of humanity, so AI is human, and we use it to extend our mental reach just as we once invented tools to extend our physical reach.

On existential risk, he recalls the civil defense drills of his childhood and says nuclear weapons are still on a hair-trigger. He says he has written extensively about the downsides of genetics, nanotechnology and robotics while remaining optimistic. His model is the Asilomar conference on biotechnology, whose guidelines have since been refined and baked into law, and he says virtually no one has been harmed by biotechnology; a conference on AI ethics followed in January.

He also argues that the cost of caution is real: a gene therapy death led to halted research, and he claims many unnamed people died from the delay. For AI, he admits there is no subroutine that guarantees safety, since intelligence is inherently not controllable, so the best approach is to practice democratic ideals and ethics today, because the singularity will emerge from our present society. Measures of world progress, such as literacy and poverty, keep improving even though people think things are getting worse, and he allows that violence could disrupt that.

To a final question about why ideas, not just hardware, should grow exponentially, he cites a study from the Obama administration's scientific advisory board on 23 classical engineering problems, in which hardware gave about a thousand-to-one gain over a decade and software and algorithmic improvements gave an average of about 26,000 to one. He says each innovation can double or multiply performance, as seen in deep learning, where architectures, data, compute and memory all improve together.