Bitwise

by David Auerbach

  Not yet, anyway. McCulloch was speaking of the calculating machines of the mid-twentieth century. But now that large systems like Google and Facebook are persisting and growing for years and decades, we can contemplate the possibility of an evolving, maturing network whose intelligence is not intrinsic to its algorithms but lies in its evolved complexity, developed over great periods of time and through repeated, varied, and error-prone interactions with the world—just like a child. We don’t debug these networks; we educate them.

  Machine and Child Learning

  We don’t need something more in order to get something more.

  —MURRAY GELL-MANN

  I did not understand how my child was changing. She grew in size, and she used more complex sentences, but I was far less certain of what was going on inside her head. In the first months of her life, I kept a spreadsheet of her milestones, and declared a new “version” whenever my wife and I deemed her sufficiently different to appear as though a software upgrade had been installed. Hardware upgrades to her height and weight were ongoing. I had:

  v1 (0 weeks): Initial conditions.

  v2 (2 weeks): She smiles. Her first facial expression.

  v2.1 (4 weeks): Turns head and moves arm.

  v3 (9 weeks): Tracks objects with her eyes.

  v3.1 (10 weeks): Sleeps through the night.

  v4 (15 weeks): Grips and picks up a rattle with both hands.

  v5 (22 weeks): Rolls over from back to stomach.

  v5.1 (23 weeks): Wraps her arms around my neck and hugs me.

  It was tempting to see these changes as upgrades because I wasn’t doing anything to trigger them. My daughter was just figuring it out on her own! Having spent two decades of our lives in front of computers, my wife and I weren’t used to seeing our “projects” alter their behavior without long and hard intervention. “Maintenance” was required (nutrition came in, waste went out), but there was no clear connection between these efforts and the changes taking place in our daughter.

  “Upgrades” became more difficult to track as my daughter’s skills expanded and her comprehension of the world around her developed. I settled into the rhythm of her changes, so it wasn’t until eleven months in, when I had to be away from her for two weeks, that I was presented with drastic advancements in her walking and object manipulation.

  Yet all these behavioral changes paled in comparison to the most mystifying and transcendent leap of all—her acquisition of language. For anyone who has wrestled with the ambiguities and frustrations of how language works (and doesn’t work), seeing a child learn to maneuver its verbal gears is both revealing and confounding. My daughter learned the conventions of English through a fair amount of trial and error. As she learned more sounds and began to experiment with using words to mean more than just “I want that!” I let go of the fantasy that any sort of “upgrades” were taking place at all and I came to see her as a mysterious, ever-evolving network. There are algorithms that guide the development of the network we call the child, chief among them the workings of DNA, but those algorithms are the builders, not the building itself, and they are hidden from us.

  There is a period from roughly ages two to five when children have yet to learn the full semantic logic behind words and phrases, and so their utterances have often beautiful and hilarious malapropisms. But they can also be revelatory. When, at two and a half, my daughter said, “Worms and noodles are related by long skinny things,” she lumped together two entities based on superficial appearance, but she hadn’t yet learned what a relation was. She used the word “related” in a similar way to how she’d heard it used by her parents.

  These confusions can be more abstract. Children create fantasies when playing. The boundary between play and real life sometimes doesn’t make sense to an adult, because children treat the play like reality while they know it isn’t. Once, seeing our three-year-old playing with a stuffed turtle named Squirt, my wife asked, “Does Squirt like going to school? What did he eat for breakfast?” My daughter looked up at her quizzically and said, “Mama, do you know that Squirt is not real?”

  At age three, she made up this story:

  THE ADVENTURES OF LION THE LION

  There are people in the house. They are cooking food and then a camel comes in the house! The people are very confused. The people asked the camel, “What are you doing here?” But the camel didn’t answer because he can’t talk.

  She had forgotten that the story was about a lion (named “Lion”) when the time came for the main character’s entrance, so she substituted a camel. She wanted a camel in the story, but she had no prepared scripts for how camels and people interact, since people talk and camels don’t. So the story ended there, koan-like.

  My daughter was keen to use logic to argue her position when she needed to. Sometimes it took the form of threats, particularly at bedtime: “If you don’t give me any milk, I’ll stay awake all night. Then you’ll never get any sleep and you’ll die sooner.” Or her plaintive objection when she was upset and rejected our efforts to comfort her: “I want nothing!”

  Gradually, rationality asserts itself and shoves the nonsense out of the way. By three and a half, Eleanor was modeling our motives. She didn’t always do so flatteringly, as when she said to her blanket, “Now I will raspberry you. You will not like it but I enjoy it and that is why we will do it.” At this point, she was able to determine that everyone around her had goals and that sometimes those goals conflicted with hers. She couldn’t necessarily determine others’ motivations, but she knew they were there.

  Similarly, at that time, my daughter began to understand that things could be made out of other things, and sometimes there were things that she could not see. She knew human bodies were made of bones and blood, which led to this song, to the tune of “Frère Jacques”:

  I hear David, I hear David

  Here he comes, here he comes

  Now he’s walking through us, now he’s walking through us

  Now he’s wet, now he’s wet

  ME: Why did I get wet?

  HER: You’re wet with blood.

  ME: Where did the blood come from?

  HER: Because you walked through us.

  She created the explanation out of the raw materials of her observations of the world around her and her attention to how people talked. The leap from observational data to thought is one of the most amazing and incomprehensible processes in nature. Any parent will know how baffling it is to see this happening in stages. There are limits past which a child cannot go in understanding, until one day those limits mysteriously vanish, replaced by new and deeper ones. Children come by and large to the same shared understanding that we all possess. But it can’t be rushed. Even as she knew that bodies were made of blood and bones, she could not yet conceive of a hierarchy of substances, judging by the frustration she exhibited after we saw a picture of an atom in a book:

  ME: Everything is made of atoms. Even you.

  HER: No, I am made of bones.

  ME: Your bones are made of atoms!

  HER: No, my bones are made of…muscles.

  ME: Your muscles are made of atoms.

  HER: [very skeptical] My muscles are made of…muscle.

  ME: Muscle is made of atoms.

  HER: [Utterly fed up, turns page]

  What remains a puzzle to me, and to researchers in general, is how children leap from superficial imitation and free association to reasoning. The brain grows and develops, with billions of neurons added year after year—but no matter how much memory or processing power I add to my desktop server, it never gains any new reasoning capabilities.

  There are many different types of networks coming into existence besides giant informational system
s like Google and Facebook. There are neural networks, deep learning networks, and belief networks, among others. All these fall under the broad rubric of machine learning. Today’s most powerful machine learning techniques, such as those employed by Google’s DeepMind, excel at recognizing similarities between explicit patterns, whether those patterns are made of words or pixels or sound waves. They can judge whether two passages of text have similar lexical structures and word choices, but can say little about the texts’ meaning. They can determine whether a creature in a photograph looks more like a dog or a cat, but they know nothing of what a cat or a dog is. They can beat humans at Go, but they cannot discern whether a particular Go board is beautiful or not—unless we train an algorithm on a set of “beautiful” and “non-beautiful” boards and have it try to learn that classification. While these machine learning networks can perform feats that leave humans in the dust, they inherit contexts, standards, and judgments from humans, and they are unable to generalize from a given task to similar yet distinct tasks without human guidance. They cannot reason about the application of labels, as my daughter did at age four:

  HER: These ballet shoes are so soft. I bet they are made out of polyester.

  ME: Maybe they are made out of marshmallows and you could eat them.

  HER: You can’t eat ballet shoes.

  ME: Then they aren’t made out of marshmallows.

  HER: Nothing’s made out of marshmallows.

  ME AND HER (simultaneously): Except marshmallows.

  HER: Only marshmallows are made out of marshmallows. That’s why they are called marshmallows. All the other names are used up by people and other stuff.

  In contrast to the image classification performed by machine learning networks, children quickly learn to categorize by far more than visual similarity, and in fact learn to reject visual similarity in favor of other categorizations. As Susan Gelman found, once told a pterodactyl is a dinosaur and not a bird, even a young child will tend to infer based on that category membership rather than any visual similarity, in guessing, for example, that a pterodactyl does not live in a nest or that a dolphin cannot breathe underwater.

  A machine learning network could not switch from visual to nonvisual categories of its own accord. It needs explicitly coded directives given by humans. It can label pictures of marshmallows as “marshmallows,” but cannot infer that “marshmallows are made of marshmallows.” That’s meta-reasoning. But if there’s any such directive encoded in our genes or in our neurons, we have yet to discover it.*5 I wondered how my daughter merged her cognitive skills to create a whole greater than the sum of its parts. There are competing systems in play: grammar, fantasy, rationality, narrative, taxonomy, visual imagery, and more. Somehow, they coalesce into what we think of as a rational adult. There is no clear explanation for how this occurs. Many models of development have been suggested, but the only certainty is that, somehow, rational human adults are, as far as we know, the beings most capable of coping with and exploiting the world. We are also the beings most capable of complicating the world and creating problems for ourselves, and posing existential threats to the whole planet—another by-product of our evolutionary fortunes.

  The psychologist Lev Vygotsky postulated that the development of child speech and child thought were fundamentally separate, and integrated only once both had reached a certain level of maturity in toddlerhood. Until that point, speech serves a communicative but nonrational and nonrepresentational function. Then children discover that words name things. They aren’t just noises that get the child attention or food or hugs. At that point, “thought becomes verbal, and speech rational.”

  Vygotsky’s simple schema isn’t necessarily right, but it is evocative and useful. It reminds us that a mechanistic theory of rote learning and knowledge acquisition isn’t sufficient to explain how children get from point A to point grown-up. There is no single executive function running the show and directing the development of language, but multiple constituent parts that somehow become integrated, at which point children’s reasoning power soars by leaps and bounds, and continues to do so for at least a decade. As thought becomes more abstract around the age of five, even identifying the precise changes taking place is an impossible task. My daughter, who has lost her memory of her early years as all children do, has not been of much help. At age five, I asked her about the meaning of her song, which she’d forgotten she had made up:

  I hear David, I hear David

  Here he comes, here he comes

  Now he’s walking through us, now he’s walking through us

  Now he’s wet, now he’s wet

  As to why I was wet, she said, “Is it because I like to splash you with water at bathtime?” An excellent, logical explanation, which brings back many soggy memories. She was perplexed to hear the real explanation she’d given: “Did I say that?” The entire nature of the word “wet” had changed for her. It was no longer something that linked disconnected things like blood and water. It was now a real concept. Now she knew that blood had nothing to do with how I could get wet. In moments like this, in which the whole of child development seems packed into the changing use of a single word, I think of Vygotsky’s maxim “The meaningful word is a microcosm of human consciousness.”

  She even came up with explanations for why she hadn’t known things before. She was perplexed that there was a time when she hadn’t known that the natural numbers went on forever. So she said:

  Once babies get to 100, they don’t realize there are still other numbers that are hiding behind it. 100 is the king of the numbers, but it hides secrets from babies. The secrets are 101 and all of the other numbers above it.

  The transition from free association to rational explanation that children unknowingly make is a mystery that artificial intelligence has yet to conquer. The problem facing AI at this time is how to move from the specific to the general in a humanly rational way: how to take the knowledge from one clearly defined task, like labeling images or playing Go, and put it to new and different use in a general-purpose thinking network. I suspect that accomplishing this will require the creation of networks that engage with the physical world in a variety of different ways, processing visual and verbal information in a variety of contexts and learning—slowly—what approaches do and don’t work in various situations. If this were possible, the network would also need to be sufficiently powerful to apply the same broad set of techniques to varied and novel problems.*6 We still have a long way to go.

  If we do indeed create such a general network, it’s not clear that any great secret of the nature of intelligence will be revealed. We’ll have created something as complicated and irreducible as a human infant itself. We will be able to watch these networks grow, learn, and mature, but we will not be able to debug them any more than we can debug a child. Nor will we understand how or why they function in the way that we understand how an algorithm functions. To say, “Oh, well, it said ‘Goo’ instead of ‘Ga’ because this set of network weights was not triggered and this one was” is not an explanation. Rather, we will see, as I did with my own daughter, that a complex set of predispositions and behaviors, when encoded into a single creature, results in even more complexity when that creature starts to engage with the world in myriad ways.

  When Alan Turing coined the Turing Test (which, contrary to frequent news reports, has not been beaten by any computer), he implied that behavior and speech are sufficient evidence to determine whether a creature can “think.” As my daughters grow up, I witness them increasingly thinking in ways they have never done before, just as AIs are starting to impress us with their “thinking.” If something acts like it’s thinking, that will be good enough for most people. It is no wonder we are desperate to program AIs to love us. We had better be prepared to love them as well. As my older daughter once asked me, “If we break through a screen, are we in the computer’s
life, and do we get to feel what it feels like?”

  *1 I was, in fact, able to mount a counterattack after finding a fan-made My Little Pony album by “Horselover Fat” that distorted, rearranged, sped up, and otherwise mutilated two of Pinkie Pie’s songs. I thought it was an improvement, but my daughter disagreed.

  *2 It is trivial to write a program that does not terminate, but algorithms must be finite. Fans of Gödel may be interested to know that an incompleteness proof shows that it is not possible to write a program that determines whether a given program (any program) does or does not run forever (the so-called halting problem).

  *3 Roederer uses the term “pragmatic information” to describe what constitutes the purpose-driven changes in a system, but since this is a distinct concept from Claude Shannon’s purely statistical definition of information, which discusses only the degree of uncertain possibility in a message rather than any meaning it might have, I am mentioning “information” only in this note.

  *4 To put it another way, for any operation, the algorithm takes the current state of the system as one of its inputs. New inputs do not occur in a vacuum, but are themselves dependent on past outputs produced by the system and the environment’s response to those outputs.

  *5 I personally am fond of Jaak Panksepp’s layered models of emotions, as described in The Archaeology of Mind: Neuroevolutionary Origins of Human Emotion.

  *6 By “technique” I mean some low-level mechanism like a brain neuron or a hormone.