Take a photo of a barcode or cover
3.98 AVERAGE
3.98 AVERAGE
A truly amazing read. Provides a broad and novel framework for 'how the brain might work'. What's great is it states it as a hypothesis with a listing of testable predictions to chase. And here's the thing- it just might be correct!
Truly remarkable book. A simple concept that behavior is memory laid out relatively comprehensibly.
Unfortunately, the meat of his argument gets very dense in the middle, and it was very hard to stay focused. That said, it is worth it. A great read.
Unfortunately, the meat of his argument gets very dense in the middle, and it was very hard to stay focused. That said, it is worth it. A great read.
Jeff Hawkins' theory of cognition -- of the way the brain works is novel and compelling. Jeff provides a very easy-to-understand description of how he thinks the cortex functions and why today's computers will probably never be intelligent.
I spent some time studying models of cognition, and Jeff's work aligns with biases I already had. It is fairly easy and very interesting reading.
I spent some time studying models of cognition, and Jeff's work aligns with biases I already had. It is fairly easy and very interesting reading.
informative
medium-paced
inspiring
slow-paced
An easy read up to Chapter 6. Describes the theory of how the brain processes information using a hierarchy of a repetitive pattern.
Jeff Hawkins has done a remarkable thing. He's essentially synthesised all of the information we have on how the brain works into a simple, elegant and utterly comprehensible theory of intelligence that will pave the way to the creation of truly intelligent machines. That's a massive claim I know but I honestly don't think I have ever read a simpler, more straightforward account of what intelligence is.
Hawkins' theory, in a nutshell, is that intelligence is a manifestation of the brains ability to predict the future and test its perceptions against its predictions. Like a fractal there is a mass of self-similarity at work here. At the very fine-grained level the predictions the brains making are very mundane but as sensory information is handled, and exceptions passed up the hierarchy, and predictions passed back down the hierarchy, our brains learn from their experiences and, over time a genuine, common understanding of the world emerges.
Anyone working on machine intelligence should read this short, simple book.
Hawkins' theory, in a nutshell, is that intelligence is a manifestation of the brains ability to predict the future and test its perceptions against its predictions. Like a fractal there is a mass of self-similarity at work here. At the very fine-grained level the predictions the brains making are very mundane but as sensory information is handled, and exceptions passed up the hierarchy, and predictions passed back down the hierarchy, our brains learn from their experiences and, over time a genuine, common understanding of the world emerges.
Anyone working on machine intelligence should read this short, simple book.
A nice read which gives some insights into the mechanisms of cognition and intelligence. The author proposes an overarching theory of human intelligence which is quite interesting. Only problem is it's not completely rigorous and the author hasn't given a good list of references.
A book exploring intelligence in neurons and its parallels in silicon. Some decent bits on neuroanatomy and cell networks in the brain and sheds a little light on this topic through a treatment of the visual cortex. Good stuff but is a little out of date the book came out in 2005 and machine learning is the hot thing when I last checked however there is some fun stuff from this book.
Okay. This book and I didn't get along terribly well, but the experience was nevertheless a valuable one. So, 3 stars, even though I disagree fundamentally with some of the theory and the style of presentation. This will be a long one; bear with me.
To put it simply.... Jeff Hawkins is a very intelligent computer engineer who thinks he understands brains in ways that no neuroscientist ever has before, mostly because he is willing to stand by a grand picture where most neuroscientists want to investigate every small chunk before declaring they've solved brains. He has read a lot of books about neuroscience and has spoken to a lot of neuroscientists, and has trudged up a (not patently incorrect) theory from the 1970's and used it as a foundation for what he considers the first general theory of cortical brain functioning (it isn't). He then equates "cortex" with "intelligence" and takes off on a grand tour of his theory: that we can build intelligent* machines to perform complex pattern recognition tasks in much the same ways that he proposes an organic cortex does.
*NOT human-like, though you wouldn't know it from the wording in any blurb you read about the book, including the book's own jacket summary.
There are a lot of theoretic assumptions in this book, and unpacking them is quite unfortunately left up to the reader, who may not have the requisite background knowledge to separate out responsible assumptions, backed by data that Hawkins rarely mentions in order to keep it digestible to a lay audience, from irresponsible ones. There are many of the latter, detectable only by those who know the field or the scientific method well enough to know a red flag when they see it. Hawkins plays loose on owning up to these assumptions, even when they are cornerstones without which his theory loses a lot of its appeal. I was relieved when he admitted to the oversimplification of his view at the end of the chapter on neurology, but it felt like an afterthought that could (should?) have been used to temper his conviction and factual flippancy up to that point. The tone of the book is occasionally that of a conspiracy theorist who has figured it all out, against all conventional belief, and pulls you along a fast sequence of premises and conclusions while waving his hands and telling you the details of the premises are too complicated to get into.
One of the bafflers that stood out to me (this is a bit technical, sorry) was the notion that the basal ganglia and cerebellum are old structures whose functions have been largely subsumed by the neocortex, and thus were unnecessary for a theory of intelligent motor behaviour. The balls it takes to make that claim, when such vastly debilitating diseases as Parkinson's, Huntingtons, or Ataxia exist, blew my brains out a bit.
At that point, it became clear that Hawkins was so fixated on the neocortex that he was willing to push aside contradictory evidence from subcortical structures to make his theory fit. I've seen this before, from neuroscientists who fall in love with a given brain region and begin seeing it as the root of all behaviour, increasingly neglecting the quite patent reality of an immensely distributed system. It's pretty natural, and honestly not limited to neuroscientists: when you stare too closely at one piece of a puzzle, you begin to forget that there are other pieces. For most scientists, however, this view need not be a detriment, because they generally aim their research programs at very specific questions-- questions that this atomized focus are fit to answer. In the case of Jeff Hawkins' general theory of brain function, however, it's entirely disingenuous.
Putting aside my qualms with his approach to the theory, there is actually some overlap between his views my own, and some points of valuable and probably instructive disagreement. Hawkins views intelligence as a result of hierarchical and recursive neural organization: basically, there are higher and lower levels, with communication tracing both upwards (from sensory input) and downwards (from higher levels of analysis) via patterns of activation. What we experience is a complex interaction between external input and internal input from memory, resulting in a continuous stream of online prediction. Up to this point, my theories of what we might call consciousness match his theory of intelligent pretty well. Where we differ is in the details (and in our respective convictions that we are correct!).
In neuroscience, there is a theoretical construct called the 'grandmother cell,' to illustrate the ludicrous idea that there is a single neuron in your brain that represents your grandmother, another for your cat, and so on. The grandmother cell has been disproven time and time again: the brain is a HIGHLY distributed system, and a given representation is the result of patterns of activations across many cells, not one cell. Jeff Hawkins acknowledges this.... before proposing instead that representations in the cortex are handled by (my term, not his) grandmother cellular columns. Briefly, the visual cortex has been shown to have a columnar organization that traverses six parallel layers of anatomically and biologically different cell types, such that cells at Point X of layer 1 respond to the same sorts of basic visual information (e.g., line angles) as cells at Point X of layers 2 through 6. Because the rest of the cortex also seems to be organized in six distinguishable layers, Hawkins suggests that the entire cortex operates in columns, such that the composite idea of your grandmother should be represented by a given cellular column in a high-level area of cortex. He never states these logical conclusions outright, but they follow from the way his theory proposes hierarchical organization to work. He briefly admits the oversimplification at play, and then nevertheless uses the oversimplification as the foundation for the rest of the theory. This is not a novel theory so much as it is an outdated theory with the goalpost pushed back one step. And while oversimplifications are a necessary evil in scientific progress*, they need to be acknowledged and admitted so that they can be refined, again, especially where a general theory for a lay audience is the goal.
*as my advisor says it, the goal of a scientist is to maintain a productive level of ambiguity.
The rest of the book was (to me) less controversial. There was the requisite chapter to answer such questions as "does this mean animals are intelligent!?" for readers who've never thought about the implications of a physical and evolutionarily-derived brain before. Yawn. This was followed by a chapter on what I assume is the whole reason Hawkins wrote the book: the prospect of intelligent machines.
Having defined 'intelligence' as 'cortex,' he rather plainly announces that an intelligent machine will be one organized with recursive and flexible hierarchies, a reveal that will shock or excite no cognitive scientist. He very clearly explains why current artificial intelligence built on existing computer memory structures are not up to the task, an argument that AI researchers have been ignoring for decades. Much to my pleasant surprise, again given the blurbs on this book, he then laments the cold hard reality that we will never have viable machines that are intelligent in the way humans are: humans are intelligent the way humans are because of all the sensory and proprioceptive input coming in from their human bodies to shape their brains. Unless we build almost impossibly costly and cumbersome human-like bodies to go with our fancy intelligent machine brains, it's a moot point trying to make machines like humans-- and why would we want to anyways? Hawkins outlines some realistic goals that are achievable (e.g., self-driving cars, diagnostic machines, weather prediction machines...), none of which I particularly disagree with except for the optimistic time-frames forecasted.
However, I can't help feeling that most readers are set up to be vastly disappointed by the propositions. With the majority of the book devoted to neurological theory, it's hard not to anticipate that the machine intelligence he will eventually propose will mimic neurology. The book jacket itself claims that Hawkins' theories will "make it possible for us to build intelligent machines, in silicon, that will not simply imitate but exceed our human ability in surprising ways." But nothing about the analogical applications of his neurologically-based theory are intended to imitate humans. He expressly states, in fact, that the aim to build artificial humans is wrong-footed and fated to fail. The message is a bit confusing, and while I would have personally been offended to hear him say what most readers likely wanted to hear-- that we can build human-like machines by analogy to human cortex-- I again get a sticky sense of disingenuity, this time to sell book copies.
Overall, this was an interesting but infuriating book that takes some great ideas from existing cognitive science, laudably exposes them to a lay audience in ways that most cognitive and neuroscientists won't bother to, shoves them into a flawed neurological framework, and then announces brains to be solved. The ego involved is staggering, the conclusions less so, and the applications underwhelming. I am admittedly very interested to see, hopefully in my lifetime, just how intelligent Hawkins' intelligent machines can get with only an analogical neocortex. Since he never discusses this fact, spoilers: a neocortex is not enough for either humans or nonhuman animals to function, let alone intelligently. The (rather expensive...) exercise of trying to evoke intelligence from cortex alone could provide us with a better appreciation of subcortical structures, much needed in this species-self-congratulatory era of cortical fixation.
To put it simply.... Jeff Hawkins is a very intelligent computer engineer who thinks he understands brains in ways that no neuroscientist ever has before, mostly because he is willing to stand by a grand picture where most neuroscientists want to investigate every small chunk before declaring they've solved brains. He has read a lot of books about neuroscience and has spoken to a lot of neuroscientists, and has trudged up a (not patently incorrect) theory from the 1970's and used it as a foundation for what he considers the first general theory of cortical brain functioning (it isn't). He then equates "cortex" with "intelligence" and takes off on a grand tour of his theory: that we can build intelligent* machines to perform complex pattern recognition tasks in much the same ways that he proposes an organic cortex does.
*NOT human-like, though you wouldn't know it from the wording in any blurb you read about the book, including the book's own jacket summary.
There are a lot of theoretic assumptions in this book, and unpacking them is quite unfortunately left up to the reader, who may not have the requisite background knowledge to separate out responsible assumptions, backed by data that Hawkins rarely mentions in order to keep it digestible to a lay audience, from irresponsible ones. There are many of the latter, detectable only by those who know the field or the scientific method well enough to know a red flag when they see it. Hawkins plays loose on owning up to these assumptions, even when they are cornerstones without which his theory loses a lot of its appeal. I was relieved when he admitted to the oversimplification of his view at the end of the chapter on neurology, but it felt like an afterthought that could (should?) have been used to temper his conviction and factual flippancy up to that point. The tone of the book is occasionally that of a conspiracy theorist who has figured it all out, against all conventional belief, and pulls you along a fast sequence of premises and conclusions while waving his hands and telling you the details of the premises are too complicated to get into.
One of the bafflers that stood out to me (this is a bit technical, sorry) was the notion that the basal ganglia and cerebellum are old structures whose functions have been largely subsumed by the neocortex, and thus were unnecessary for a theory of intelligent motor behaviour. The balls it takes to make that claim, when such vastly debilitating diseases as Parkinson's, Huntingtons, or Ataxia exist, blew my brains out a bit.
At that point, it became clear that Hawkins was so fixated on the neocortex that he was willing to push aside contradictory evidence from subcortical structures to make his theory fit. I've seen this before, from neuroscientists who fall in love with a given brain region and begin seeing it as the root of all behaviour, increasingly neglecting the quite patent reality of an immensely distributed system. It's pretty natural, and honestly not limited to neuroscientists: when you stare too closely at one piece of a puzzle, you begin to forget that there are other pieces. For most scientists, however, this view need not be a detriment, because they generally aim their research programs at very specific questions-- questions that this atomized focus are fit to answer. In the case of Jeff Hawkins' general theory of brain function, however, it's entirely disingenuous.
Putting aside my qualms with his approach to the theory, there is actually some overlap between his views my own, and some points of valuable and probably instructive disagreement. Hawkins views intelligence as a result of hierarchical and recursive neural organization: basically, there are higher and lower levels, with communication tracing both upwards (from sensory input) and downwards (from higher levels of analysis) via patterns of activation. What we experience is a complex interaction between external input and internal input from memory, resulting in a continuous stream of online prediction. Up to this point, my theories of what we might call consciousness match his theory of intelligent pretty well. Where we differ is in the details (and in our respective convictions that we are correct!).
In neuroscience, there is a theoretical construct called the 'grandmother cell,' to illustrate the ludicrous idea that there is a single neuron in your brain that represents your grandmother, another for your cat, and so on. The grandmother cell has been disproven time and time again: the brain is a HIGHLY distributed system, and a given representation is the result of patterns of activations across many cells, not one cell. Jeff Hawkins acknowledges this.... before proposing instead that representations in the cortex are handled by (my term, not his) grandmother cellular columns. Briefly, the visual cortex has been shown to have a columnar organization that traverses six parallel layers of anatomically and biologically different cell types, such that cells at Point X of layer 1 respond to the same sorts of basic visual information (e.g., line angles) as cells at Point X of layers 2 through 6. Because the rest of the cortex also seems to be organized in six distinguishable layers, Hawkins suggests that the entire cortex operates in columns, such that the composite idea of your grandmother should be represented by a given cellular column in a high-level area of cortex. He never states these logical conclusions outright, but they follow from the way his theory proposes hierarchical organization to work. He briefly admits the oversimplification at play, and then nevertheless uses the oversimplification as the foundation for the rest of the theory. This is not a novel theory so much as it is an outdated theory with the goalpost pushed back one step. And while oversimplifications are a necessary evil in scientific progress*, they need to be acknowledged and admitted so that they can be refined, again, especially where a general theory for a lay audience is the goal.
*as my advisor says it, the goal of a scientist is to maintain a productive level of ambiguity.
The rest of the book was (to me) less controversial. There was the requisite chapter to answer such questions as "does this mean animals are intelligent!?" for readers who've never thought about the implications of a physical and evolutionarily-derived brain before. Yawn. This was followed by a chapter on what I assume is the whole reason Hawkins wrote the book: the prospect of intelligent machines.
Having defined 'intelligence' as 'cortex,' he rather plainly announces that an intelligent machine will be one organized with recursive and flexible hierarchies, a reveal that will shock or excite no cognitive scientist. He very clearly explains why current artificial intelligence built on existing computer memory structures are not up to the task, an argument that AI researchers have been ignoring for decades. Much to my pleasant surprise, again given the blurbs on this book, he then laments the cold hard reality that we will never have viable machines that are intelligent in the way humans are: humans are intelligent the way humans are because of all the sensory and proprioceptive input coming in from their human bodies to shape their brains. Unless we build almost impossibly costly and cumbersome human-like bodies to go with our fancy intelligent machine brains, it's a moot point trying to make machines like humans-- and why would we want to anyways? Hawkins outlines some realistic goals that are achievable (e.g., self-driving cars, diagnostic machines, weather prediction machines...), none of which I particularly disagree with except for the optimistic time-frames forecasted.
However, I can't help feeling that most readers are set up to be vastly disappointed by the propositions. With the majority of the book devoted to neurological theory, it's hard not to anticipate that the machine intelligence he will eventually propose will mimic neurology. The book jacket itself claims that Hawkins' theories will "make it possible for us to build intelligent machines, in silicon, that will not simply imitate but exceed our human ability in surprising ways." But nothing about the analogical applications of his neurologically-based theory are intended to imitate humans. He expressly states, in fact, that the aim to build artificial humans is wrong-footed and fated to fail. The message is a bit confusing, and while I would have personally been offended to hear him say what most readers likely wanted to hear-- that we can build human-like machines by analogy to human cortex-- I again get a sticky sense of disingenuity, this time to sell book copies.
Overall, this was an interesting but infuriating book that takes some great ideas from existing cognitive science, laudably exposes them to a lay audience in ways that most cognitive and neuroscientists won't bother to, shoves them into a flawed neurological framework, and then announces brains to be solved. The ego involved is staggering, the conclusions less so, and the applications underwhelming. I am admittedly very interested to see, hopefully in my lifetime, just how intelligent Hawkins' intelligent machines can get with only an analogical neocortex. Since he never discusses this fact, spoilers: a neocortex is not enough for either humans or nonhuman animals to function, let alone intelligently. The (rather expensive...) exercise of trying to evoke intelligence from cortex alone could provide us with a better appreciation of subcortical structures, much needed in this species-self-congratulatory era of cortical fixation.