Auto Ads by Adsense

Thursday, April 21, 2022

Review: A Thousand Brains - A New Theory of Intelligence

 A Thousand Brains is Jeff Hawkins' book about the way the neocortex of the brain works, and the implication it has for building machine intelligences. The premise of his book is as follows:

  • The neocortex is uniform in structure and function, and under a microscope, no parts of the neocortex looks any different from other parts
  • We can abstract away the non-neocortex portion of the brain as being unimportant for the development of intelligence
  • The brain is a prediction machine (he previously covered this in an earlier book, On Intelligence), and is constantly predicting what will come next and comparing its predictions with the actual sensory input
  • The unit of cognitive function in the neocortex is the cortical column
  • All cortical columns behave the same way by learning and building a model of its inputs, but differ in function mainly in what inputs it is wired to. For instance, a cortical column wired to the eyes will be trying to recognize and build models based on vision, while a cortical column wired to more abstract thinking as input will be operating on abstract concepts
  • Learning is the process of building a model. The only way to build a model of a physical object is to move around and explore it from multiple perspective --- consider how we build a model of a building by walking through its rooms, and how when you see a novel object you'll turn it and look at it from different angles
  • Along with the model, there are reference frames, which tell you about the relationship between pieces of the model. You have a model of your own body, and the reference frames tell you about the relationship between your various body parts like fingers, which is how you can recognize a mug in the dark by touch or just by holding it, even without other inputs
  • Recognition of an object or where you are is done by a voting process, a multi-sensory associative schema where all the salient evidence from your senses is brought together and the models that most closely match that input triggers the recognition and the prediction.
  • The model of the world or object is what knowledge is, not words, not data structures or labels. When you're asked a question about an object, your model of that object is what you use in order to answer those questions
  • Consciousness is a portion of your neocortex wired up to examine its internal state, with the ability to playback and remember what has happened in the past.
Hawkins (perhaps arrogantly) claims that this is the overall framework theory of how intelligence works, and that while the details might have to be refined, this framework will prove to be true. (There's substantial controversy about this claim, so take it with a grain of salt)

There are a few implications of this:

The neocortex never stops learning models. Every shift of attention—whether you are looking at the dishes on the dining table, walking down the street, or noticing a logo on a coffee cup—is adding another item to a model of something. It is the same learning process if the models are ephemeral or long-lasting. (kindle loc 1566)

 In particular, Hawkins claims that today's neural network models do not hold reference frames, which are key to knowledge, and therefore cannot learn and build models:

Robot designers are accustomed to using reference frames. They use them to keep track of where a robot is in the world and to plan how it should move from one location to another. Most roboticists are not concerned about AGI, whereas most AI researchers are unaware of the importance of reference frames. Today, AI and robotics are largely separate fields of research, although the line is starting to blur. Once AI researchers understand the essential role of movement and reference frames for creating AGI, the separation between artificial intelligence and robotics will disappear completely...Today’s neural networks rely on ideas that Hinton developed in the 1980s. Recently, he has become critical of the field because deep learning networks lack any sense of location and, therefore, he argues, they can’t learn the structure of the world. In essence, this is the same criticism I am making, that AI needs reference frames. Hinton has proposed a solution to this problem that he calls “capsules.” Capsules promise dramatic improvements in neural networks, but so far they have not caught on in mainstream applications of AI. (kindle loc 1904-1909)

 The implications of this for human learning is also significant. For instance, a lot of child development specialists criticize schools for not being good for learning, mostly because childhood has been transformed from being largely spontaneous and exploratory into something where kids are effectively jailed in a building and supervised continuously:

 from when she was five years old, Lenore would walk out of her house and walk to school on her own. It was about 15 minutes away. When school ended, Lenore would leave and just wander around the neighborhood freely on her own. She’d play games with the other kids that the kids would spontaneously organize, they’d run around, and she would go home when she was hungry.

That was how all childhood was, essentially, in the world at that point with very few exceptions. Children played freely with other children without adult supervision for most of the time. This was crucial for them. By the time Lenore was the parent in the 1990s, that had ended. She was expected to walk her kids to school, wait and watch them go through the door — even when they got pretty old — and to be there waiting at the gate to collect them at the end of the day. By 2003, only 10 percent of any American children ever played outdoors. So it essentially ended.

Childhood became something that happened either behind closed doors under tight adult supervision. And it turns out there are loads of things in this enormous and unprecedented transformation in childhood that are important for attention. Let’s give you a real no shit, Sherlock one: exercise.

Kids who run around can pay attention much better. The evidence for this is overwhelming. One of the single best things you can do for kids who can’t pay attention is let them go and run around. We have stopped that, right? Even before Covid, we stopped that.

We imprisoned our children. In fact, the only place where our kids get to feel they’re roaming around at the moment is on Fortnite and on World of Warcraft. We can hardly be surprised that they’ve become so obsessed with them. There are lots of other changes. Children learn when they play freely what’s called intrinsic motivation. (Ezra Klein interview 2022 02 11, New York Times)

So Hawkin's advocacy of learning through movement for AI can be compared to free range parents' advocacy of freedom for children to explore.  I found that fascinating to think about.

The last part of the book describes the way religion, right-wing theory, and other institutions have been constructed to hack the neocortex and use that to spread false believes. He notes that that false believe memes have to have the following characteristics:

1. Cannot directly experience: False beliefs are almost always about things that we can’t directly experience. If we cannot observe something directly—if we can’t hear, touch, or see it ourselves—then we have to rely on what other people tell us. Who we listen to determines what we believe. 2. Ignore contrary evidence: To maintain a false belief, you have to dismiss evidence that contradicts it. Most false beliefs dictate behaviors and rationales for ignoring contrary evidence. 3. Viral spread: Viral false beliefs prescribe behaviors that encourage spreading the belief to other people. (kindle loc 2758)

He applies this to vaccine denial, climate change denial, and the flat earthers. The final part of the book is a plea to teach kids about false believes and innoculate our children about how such false beliefs are harmful. Looking at the state of the world, I definitely believe that he's on the right track.

In any case, the entire book is well worth reading, and very much worth your time. I devoured it in a few evenings and didn't regret any time spent reading it.


No comments: