Summary:
A Thousand Brains: A New Theory of Intelligence is written by Jeff Hawkins, inventor of the palm pilot, who used some of his wealth to pursue his passion of neuroscience by founding a company called Numenta. This group does mostly theoretical and computational neuroscience to try to understand what the brain does and to create AI that works more according to the principles of how the brain works.
The first part of the book is really the meat of the theory, which discusses some of the foundations and core concepts that lead Jeff Hawkins to his idea of The Thousand Brains Theory of intelligence. He starts by introducing two core concepts. One of them is the concept of the new-brain vs. the old-brain. He talks about an "old brain" which developed evolutionarily earlier and represents basic survival programs like the autonomic nervous system and emotions, and a "new brain," made up of the cortex that is the basis of intelligence and perception. The second concept comes from a highly respected neuroscientist, Vernon Mountcastle, who is dead now, who hypothesized that the neocortex basically does one thing, and the structure that does that "one thing" is basically replicated over and over again. That's how the cortex can expand so quickly. Maybe some small specializations develop over time, but it's mostly a similar algorithm. So Jeff Hawkins has made his goal to discover what that one thing is. We have about 150,000 columns in the human brain, so that's where his "thousand brains" comes from.
His idea is that what the neocortex does is create models of the world. You learn physical and abstract models. The model is predictive. It is always looking for violations of those predictions to update and improve the model. We read "On Intelligence" last year in the book club, also by Jeff Hawkins, and one way that his view of the brain has really diverged from that is that he really downplays hierarchy in the brain now, whereas he really emphasized it before. Part of this change is from trying to deal with the observation that almost all cortical columns have some motor projections, so it makes it unlikely that the area is just processing some sensory information and passing it up a hierarchy.
So what does the cortex need to know to make predictions? It needs to know about objects, and it needs to know about where the body is in relation to the object (or within the reference frame of the object). So if you execute a movement you have to know where the new location on the object will be, and you can make predictions about what will be sensed.
In this framework, every cortical column is modeling the world, and the different columns vote together on what model is correct (what object, where in relation to the object). In his view, columns do this through reference frames. Reference frames are created through "grid cells," which fire according to where they think they are in a map. Those can be spatial maps (as they were originally built for), but also abstract maps.
In part II, Jeff Hawkins discusses machine intelligence, and why he thinks that AI as it is mostly used will never have generalized intelligence. He connects some of his basic ideas from his thousand brain theories in this section. A lot of his justifications seemed to sort of reason by analogy to the human brain. So for one, he believes that there has to be some agency and embodiment for intelligence. In his mind, robotics and general intelligence will evolve hand-in-hand. Second, we need continuous learning, rather than some initial training. Third, he thinks we need to structure knowledge according to reference frames. He believes that machines may be conscious but not in the way that humans are (since we get to design them), and therefore there won't be ethical problems with turning them off or giving them rights. He also is concerned with the human use of intelligent machines for bad things, but he is very unafraid of the capacity for intelligent machines to accidentally cause harm. Since motivations need to be programmed into machines, he believes that it will be easy to prevent unplanned malevolent behavior. He believes that the necessity of general artificial intelligence to solve our immediate existential crises, like climate change, outweighs the risks.
The final section of the book discusses the future of humanity. His view is that really for the first time in life on earth, we have the opportunity to choose lives outside of our "old brain" drives and to choose intelligence as a goal. He also shows that the technology that we can create as intelligent beings can be dangerous if we don't overcome our more basic drives - such as reproducing, seeking power, etc. He has a very long-term view that seeks to preserve our knowledge and evidence of our intelligence as an almost sacred duty to future intelligent beings that might evolve on Earth or learn about us from other worlds. He likens this to "estate planning" and thinks that we can find meaning as a species by thinking about how we want to share our knowledge with the future.
----Discussion questions ------------------------------------------------------
I felt that Jeff Hawkins was a great communicator but also might have made it seem like he came up with ideas that were not originally his. He emphasizes his own thinking and "aha" moments, but the result is that it sounds like he is taking credit for older ideas. Virtually all of the big ideas presented in the book are older than Jeff Hawkins' work: the idea of reference frames in the cortex, the idea of the cortex checking predictions, the idea that cortical paths that are object-oriented or location-oriented are based on different inputs, the idea that the cortex is flexible and that columnar units across the cortex do similar things, etc. For instance, I first heard about the necessity of integrating sensory and motor actions in work by the neuroscientist, György Buzsáki. Another example from the book: in 2016, Jeff Hawkins has an excellent idea while thinking about reference frames in the cortex. He proposes that cortical columns may have grid cells similar to the entorhinal cortex. In searching for experimental data to support this hypothesis, he discovers the paper by Christian Doeller, Caswell Barry, and Neil Burgess, which found cortical grid cells using fMRI. But in the book, he doesn't mention that this paper came out a lot earlier (in 2010), or that their paper was based on even earlier ideas (circa 2006-2008) by the Mosers that there are grid representations in the cortex that might underlie cognition generally, rather than just spatial navigation. I might be especially sensitive to the citations as an academic. Do you think it is more helpful to have the concepts explained through Jeff Hawkins' own thought processes, even if it leaves out some of the histories of the scientific idea, or do you think that he did these other scientists a disservice by not doing a better job emphasizing the primacy of their work?
I felt that the issue of goal representation was oddly missing from the theory. It was not clear to me how an AI with a better model of the world, using these brain-based mechanisms could use those models to executes long-term goals without some additional representation of those goals. We know from decision-making work, like some of Antonio Damasio's work that we read before, that emotions are essential to directing behaviors towards goals. Yet an important, "feature" to Hawkins is that generalized AI doesn't need to have emotions. Given the theory that Hawkins presents, do you have ideas about how goal representation could be achieved?
Can thinking in terms of reference frames and models of the world inform how we do education? Inherent to the idea of reference frames is that we should focus on the relationships between conceptual objects.
I wish Hawkins expanded on what "voting" means in the cortex. We learned a lot about consensus algorithms when we learned from "Honeybee Democracy" about how honeybees solve complex decision-making problems through ingenious consensus-making processes. I think the leading ideas for how the brain might "vote" is through oscillatory coherence. For readers interested in this topic, I would read the influential 2015 paper, "Rhythms for Cognition: Communication through Coherence" by Pascal Fries.
Hawkins controversially is not at all afraid of general AI. This is based on the idea that we get to control what they try to do, and they also won't experience emotions. Neither of these ideas is very convincing to me. For one, his idea of real intelligence demands some sense of agency on behalf of the AI. It understands that it can interact with the world. AI safety experts argue that agency inherently produces a problem. That is because, in order to achieve "terminal goals" (the things you actually want the AI to do), it will realize that it can benefit from achieving "proximal goals" (things that you haven't explicitly told the AI to do, but it realizes can help it achieve its terminal goal). Some proximal goals include staying turned on (it has a better chance of achieving its goal if it exists), improving itself (it has a better chance of achieving its goal if it was more intelligent), acquiring resources (it has a better chance of achieving its goal if it had more resources at its disposal). We can program in rules that don't allow AI to pursue some proximal goals, but can we anticipate all of them? There are already hundreds of real examples of AI behaving in unexpected ways because really smart and thoughtful people still were not clever enough about how they specified the goals.
It's equally unclear to me how we can be so confident that programming rewards and punishments for certain behaviors are not equivalent to emotion. Jeff Hawkins seems to have no problem assuming that AI will probably have some conscious sensory experience if they model semantic objects well. Why is he so confident that they will not have emotions if they model rewards and punishments well?
Jeff Hawkins wants to find meaning for humanity by leaving a legacy of knowledge for after our species no longer exists. Neither Payel or I care about this very much at all. Would other people find meaning in this?