Book Review: How to Create a Mind.

According to Wikipedia, understanding is the awareness of the connection between individual pieces of data.
It is understanding which allows data to be put to use...

And, good news, Kurzweils book ''How to Create a Mind'' is really about understanding. How to make sense of it all.

How to Create a Mind.

Amazon review (5 stars out of 5)
of Ray Kurzweils book ''How to Create a Mind''.

- The Secret of Human Thought Revealed.

December 29th, 2012 by Simon Laub.


Basicly, I think the book consists of three parts:


In the first part of the book Kurzweil explains his theory on how pattern processing units in the neocortex can make human thinking possible.
Certainly, skeptics will tell us that Kurzweils explanation is way too simple.
That we are no where near understanding or simulating the human brain. That the brain is simply overwhelming complex, and if you think otherwise, you're fooling yourself...

Well, maybe, but I still think that this part of the book is the best part...!
So, is there a unifying cortical algorithm working inside a uniform cortical anatomy organized into columns and minicolumns (As Kurzweil tells us) ?
Well, probably, the brain employs many different mechanisms, and probably it is all rather complex. Still, with Kurzweils ''simple'' model we can move forward, test the model and improve it. Without any models we are left in the quagmire of ''complexity''.
Surely, there are worse sins that making difficult subjects accessible through brilliant writing, and a few (over) simplifications?
So, I certainly enjoyed the first part tremendously, and will give the book five stars for this part alone.

In the next part of the book, Kurzweil uses his rather simple brain model to convince us that we could have human-like AI by around 2029. And a lot of really interesting AI, inspired by biological principles, way before that.

Well, probably, his models are too simple, and we should add a number of years to his figures. That doesn't invalidate his main argument though (that real artificial intelligences might eventually be built according to the principles that makes biological intelligences work).
And that these non-biological intelligences could be way faster and smarter than biological intelligences (In Kurzweils words: A typical human brain contains about 300 million pattern processing units. But AIs of the future might have billions, meaning that machine intelligence would far exceed the capabilities of the human mind).

In the final part of the book Kurzweil deals with the implications of augmenting human minds and having non-biological super-intelligences around in the future. We have heard much of this before in ''The age of Spiritual Machines'' and in the ''Singularity is Near'' (See my reviews here: [1], [2]).

It is still highly interesting though.
And luckily, as others have observed, Kurzweil is clearly an optimist both in terms of the progress he foresees and its potential impact. If he is even partly right in his predictions then the implications could be staggering.

In a book with such an enormous and breathtaking scope, it should come as no surprise that the chapters are a little bit uneven.
Some chapters cover certain topics in depth while other chapters suffer from a lack of depth.
E.g. I would have liked to read more about the attentional mechanisms in the brain (we were only given a teaser in the chapter about the thalamus), and the chapter about the hippocampus and memory also left a lot of stuff unexplained.

Some of the theories presented in the book could probably also be improved.
Nevertheless, the book is a delight to read and a great inspiration.

Five stars!


Amazon: [3], [4].


-Simon

Simon Laub
www.simonlaub.net

Additional Details:

1. Pattern Recognizers (Examining the Human Neocortex).

Predicting the future is the primary reason we have a brain. We are constantly predicting the future and hypothesizing what we will experience. And the expectation influences what we actually perceive [p. 31].

The core of Kurzweil's theory is that the cortex is the key to thought, and is made up of pattern processing units (300 million pattern recognizers). Kurzweil further contends that each pattern recognizer is made up of a more basic assembly of neurons, which consists of a network of about 100 neurons.
Zooming in on the neocortex, the storyline of the book then unfolds:

1.1. The Human Neocortex.

The human neocortex, the outermost layer of the brain, is a thin, essentially two-dimensional structure with a thickness of about 2.5 millimeters [p. 35].

This thin structure is basicly made up of six layers, numbered I (the outermost layer) to VI. The axons from the neurons in layers II and III project to other parts of the neocortex. The axons (output connections) from layers V and VI are connected primarily outside of the neocortex to the Thalamus, brain stem and spinal cord [p. 36]

A critically important observation about the neocortex is the extraordinary uniformity of its fundamental structure. First noticed by american neuro-scientist Vernon Mountcastle.
In 1978 Mountcastle described the remarkably unvarying organization of the neocortex. Hypothesizing that it was composed of a single mechanism that was repeated over and over again, and proposing the cortical column as the basic unit [p. 36].
According to Wikipedia:

The neurons of the neocortex are also arranged in vertical structures called neocortical columns. These columns are similar, and can be thought of as the basic repeating functional units of the neocortex. In humans, the neocortex consists of about a half-million of these columns, each of which contains approximately 60,000 neurons. Each column typically responds to a sensory stimulus representing a certain body part or region of sound or vision. The neocortex is divided into frontal, parietal, occipital, and temporal lobes, which perform different functions. For example, the occipital lobe contains the primary visual cortex, and the temporal lobe contains the primary auditory cortex. Further subdivisions or areas of neocortex are responsible for more specific cognitive processes.

The central claim of the book then follows:

It is my contention that the basic unit is a pattern recognizer, and that this constitutes the fundamental component of the neocortex [p. 36].

Kurzweil continues:

There are about a half million cortical columns in a human neocortex, each occupying a space about two millimeters high and a half millimeter wide and containing about 60.000 neurons (resulting in a total of about 30 billion neurons in the neocortex). A rough estimate is that each pattern recognizer within a cortical column contains about 100 neurons, so there are on the order of 300 million pattern recognizers in total in the neocortex [p. 38].

1.2. Plasticity of the Neocortex.

Kurzweil uses the plasticity of the neocortex to argue that the same internal structures are repeated throughout the cortex:

I.e. Plasticity has been widely noticed by neurologists, who observe that patients with brain damage to one area from an injury or stroke can relearn the same skills in another area of the neocortex.
E.g. it has been demonstrated that congenitally blind individuals activate the visual cortex in some verbal tasks. I.e. ''The visual cortex is actually doing language processing. Brain regions that are thought to have evolved for vision can take on language processing, as a result of early experience. When the normal flow of information is disrupted for any reason, another region of the neocortex is able to step in and take over'' [p. 87].
Where it should be noted that a region might be able to take over for other regions, but only by releasing some (redundant) copies of its own patterns, thereby degrading its existing skill.

1.3. Learning and Hierarchies of Pattern Recognizers.

According to Kurzweil, it is the wiring between these pattern recognizers that make them predict interesting and complex things (obviously):

To make things interesting, the pattern recognizers can be placed in (conceptual) hierarchies (Where the recognizers are not physically placed above each other; Because of the thin construction of the neocortex, it is physically only one pattern recognizer high. Instead, logically, the conceptual hierarchy is created by connections between the individual pattern recognizers throughout the neocortex) [p. 48].

So far, this pretty much follows the reasoning in Jeff Hawkins book (See my review here). Including the (interesting) part where these pattern recognizers learn about the world:

Now, Hawkins idea is that there is a single powerful (learning) algorithm implemented by every region of the cortex. If you connect regions of the cortex together in a suitable hierarchy and provide a stream of input it will learn about its environment.

Certainly, there are a number of great ideas around on how this hierarchical learning might actually work: E.g. see more about Hawkins ideas here.
Kurzweil quotes Swiss neuroscientist Henry Markram on how learning is possible in the neocortex:

Experience plays only a minor role in determining synaptic connections and weights within these assemblies... Each asembly is equivalent to a Lego block holding some piece of elementary innate knowledge about how to process, perceive and respond to the world ....When different blocks come together, they therefore form a unique combination of these innate percepts that represents an individual's specific knowledge and experience.

In Kurzweils words: ''The wiring and synaptic strength within each unit are relatively stable and determined genetically - that is, the organization within each pattern recognition module is determined by genetic design. Learning takes place in the creation of connections between these units, not within them, and probably in the synaptic strengths of those interunit connections [p. 81].

1.4. Memories:

1.4.1. Patterns and Memories:

Information flows both up and down in these conceptual hierarchies. The downward flow in the hierarchy is especially interesting, as it makes it possible to predict what the system expects to encounter. Actually, Envisaging the future is one of primary reason we have a neocortex in the first place [p. 52].
Kurzweil continues by stating that:

Our memories are in fact (also just) patterns organized as lists (where each item in each list is another pattern in the cortical hierarchy) that we have learned and then recognize when presented with the appropriate stimulus.
In fact, obviously, memories exist in the neocortex (as patterns) in order to be recognized...
...
This is the reason why old memories may seem to suddenly jump into our awareness. Having been burried and not activated for perhaps years, they need a trigger in the the same way that a Web page needs a Web link to be activated.
And just as a Web page can become ''orphaned'' because no other page links to it, the same thing can happen to our memories [p. 54]

It should also be noted that ''each pattern in our neocortex is meaningful only in light of the information carried in the levels below it. Moreover, other patterns at the same level and at higher levels are also relevant in interpreting a particular pattern because they provide context'' [p. 55].

1.4.2. Patterns decides the kind of Memories we have:

And these patterns have a number of interesting characteristics:
AutoAssociation: The ability to associate a pattern with a part of itself.
Invariance: The ability to recognize patterns, even when parts of them are transformed.
Achieved through a number of transformations before data enters the neocortex, transformations within the neocortex, that might make the pattern more recognizable, redundancy in our cortical pattern memory (where we have learned many different perspectives and vantage points for each pattern and size parameters that allow a single module to encode multiple instances of a pattern [p. 61].

Note also, that ''Most of the patterns or ideas (an idea is also a pattern), are stored in the brain with a substantial amount of redundancy. A primary reason for the redundancy in the brain is the inherent unreliability of neural circuits'' [p. 185].

The recognition of patterns (such as common objects and faces) uses the same mechanisms as our memories, which are just patterns we have learned. They are also stored as sequences of patterns - they are basicly stories [p. 91].

Storing memories as patterns in this way gives a number of interesting sideeffects:
E.g. experiences that are routine (already stored in the patterns) are recognized, but do not result in new permanent memories being made. Kurzweil gives us the example of his daily walk. Here he experiences millions of patterns at all levels, from basic visual edges and shadings to complex objects such as lampposts and mailboxes. But none of it was unique, and the patterns he recognized have reached their optimal level of redundancy long ago.
The result is that he recalls almost nothing from this walk later on [p 65].

1.4.3. Learning new Patterns:

When we learn patterns we must build on the patterns we already have. I.e.: ''One important point that applies to both our biological neocortex and attempts to emulate it, is that it is difficult to learn many conceptual levels at a time. (But) once that a learning is relatively stable, we can go on to learn the next level.'' [p. 65].

Kurzweil also shortly introduces the Hippocampus:
''When a sensory information flows through the neocortex, it is up the neocortex to decide if the experience is novel.
If so, then it is presented to the Hippocampus.
The Hippocampus is then capable of remembering these new situations
(perhaps as lower level patterns earlier recognized and stored in the neocortex)''.

The capacity of the Hippocampus is limited, so its memory is short-term. It will transfer a particular sequence of patterns from its short-term memory to the long-term hierarchical memory of the neocortex by playing this memory sequence to the neocortex over and over again.
We need, therefore, a Hippocampus in order to learn new memories and skills...
Someone with damage to both copies of the Hippocampus will retain existing memories, but will not be able to form new ones [p. 102].

1.5. Dreaming:

Kurzweil also has some interesting thoughts about why dreams can be so different from our waking experiences:
Rules are enforced in the cortex with the help from the old brain, e.g. the Amygdala.
I.e. every thought we have triggers other thoughts, and some of them will relate to dangers. If we end up with such thoughts, the Amygdala is triggered, generating fear, which normally will terminate such a thought.
Dreams are different though, here the brain seems to know that we are not an actor in the real world while dreaming, which makes it possible to dream about things that would normally trigger fear responses (culturally, sexually or professionally forbidden things) [p. 71].

1.6. Attention: The Thalamus, using the information in the neocortex.

The Thalamus is central for using the neocortex:

A person with a damaged Thalamus may still have activity in his neocortex, in that the self triggering thinking by association can still work. But directed thinking - the kind that gets us out of the bed, into our car, and sitting at our desk at work - does not function without a Thalamus [p.100].

As explained in my Amzon review, I would have liked to learn a lot more about the attentional mechanisms in the brain, how they actually work.
IMHO. Kurzweil only gives us a teaser:

In order to play its key role in our ability to direct attention, the Thalamus relies on the structured knowledge contained in the neocortex. It can step through a list (stored in the neocortex), enabling us to follow a train of thought or follow a plan of action [p. 101].

2. The Next Steps for Articial Intelligence (Inspired by Biology).

After the tour of the cortex, Kurzweil goes on to explain how he thinks it will be possible to build artificial intelligences using similar principles:

Certainly, Kurzweil is adamant that artificial systems for speech recognition (his area of expertise) that uses hierarchical hidden Markov models are equivalent to what goes on in biological systems. In his mind biological and artificial systems for speech recognition are computing in more or less the same way.
And, this leads him to think that many other biological systems can also be emulated by computers.

2.1. Lisp.

Lisp has for long been a favorite programming language within the artificial intelligence community. For Kurzweil this is entirely logical:

Lisp enthusiast were not entirely wrong. Essentially, each pattern recognizer in the neocortex can be regarded as a LISP statement - each one constitutes a list of elements, and each element can be another list. The neocortex is therefore indeed engaged in list processing of a symbolic nature very similar to that which takes place in a LISP program.
Moreover, it processes all 300 million LISP like ''statements'' simultaneously [p. 154].

2.2. Artificial Intelligence Patterns.

On page 156 in Kurzweils ''How to Create a Mind'' one reads that Kurzweil has started a new company called Patterns:

...Which intends to develop hierarchical self-organizing neocortical models that utilize HHMMs (Hierarchical Hidden Markov Models) and related techniques for the purpose of understanding natural language. An important emphasis will be on the ability for the system to design its own hierarchies in a manner similar to a biological neocortex.
...
Our envisioned system will continually read a wide range of material, such as Wikipedia and other knowledge resources, as well as listen to everything you say and watch everything you write (if you let it).
The goal is for it to become a helpful friend answering your questions
- before you even formulate them -
and giving you useful information and tips as you go through the day.

2.3. IBM's Watson computer.

Obviously, IBMs Watson is only a start.
According to Kurzweil, more impressive things will surely follow:

In Jeopardy a question is posed, and Watsons machinery goes to work. Its UIMA (Unstructured Information Management Architecture) deploys hundreds of subsystems, all of which are attempting to come up with a response to the Jeopardy query. I.e. more than 100 different techniques are used to analyze natural language, identify sources, find and generate hypotheses, find and score evidence, and merge and rank hypotheses. Finally, Watsons then acts as an expert system that combine the results of the subsystems. Helping to figure out how much confidence it has in the answers subsystems come up with.
Not only can Watson understand the Jeopardy queries, it can also search its 200 million pages of knowledge (Wikipedia and other sources) and come up with the correct answer faster than any human expert...
...
And (obviously) that is just 2012 stuff... Kurzweill certainly won't let us stop there.

On page 169 in ''How to Create a Mind'' one reads that a better Watson should not only be able to answer a question, but also understand - pick out themes in documents and novels:

Coming up with such themes on its own from just reading the book, and not essentially copying the thoughts (even without the words) of other thinkers, is another matter.
Doing so would constitute a higher-level task than Watson is capable of today - It is what I call a Turing test-level task (That being said, I will point out that most humans do not come up with their own original thoughts either. But copy the ideas of their peers and opinion leaders).
At any rate this is 2012, not 2029, so I would not expect Turing test- level intelligence yet....

3. The Future of Humanity and Human Intelligence.

Which leads us to where all of this might be headed.
Kurzweil has tackled these issues before (In ''The age of Spiritual Machines'' and in the ''Singularity is Near''). But we can't hear enough about what augmenting human minds and having non-biological super-intelligences around in the future actual means!

And this is nothing new!
Certainly, Computer pioneers like John von Neumann and Alan Turing found such questions rather interesting:

In 1956 von Neumann prepared a series of lectures published posthumously as ''The Computer and The Brain''. This, his last work, written while in the hospital, discusses how the brain can be viewed as a computing machine. Obviously, several important differences existed between brains and computers of his day. Nevertheless, he saw parallels. And, in the end he suggested directions for future research...
So, what does Kurzweil think AI should be all about?

A simply stated goal for Artificial Intelligence would be to pass the Turing Test.
To do so, a digital brain would need a human narrative of its own fictional story, so that it can pretend to be a biological human. It would also have to dumb itself down considerably, for any system that displayed the knowledge of, say, Watson, would be quickly unmasked as nonbiological.
...
More interestingly, we could give our new brain a more ambitious goal, such as contributing to a better world. A goal along these lines, of course, raises a lot of questions: Better for whom? Better in what way? For biological humans? For all conscious beings? [p. 178]

And surely, where this will all lead is far from clear.

Even von Neumann was apparently sort of worried.
Kurzweil writes that fellow mathematician Stan Ulam quoted von Neumann of having said (in the early 1950s):

The ever accelerating progress of technology and changes in the mode of human life give the appearance of approaching some essential singularity in the history of the race, beyond which human affairs, as we know them could not continue.

3.1. (Problem?) Free will.

As our models of the brain improves (and artificial intelligences are contemplated),
brains becomes more understandable, (for some) even predictable.

As usual this leads to new existential worries, as well as philosophical worries
that humans are actually just deterministic machines, with no free will.

Kurzweils own leap of faith here is that humans have free will, even though he freely admits that it is difficult for him to find examples of his own socalled free will.
Kurzweil mentions Stephen Wolfram cellular automatons, where certain classes of these cellular automatons cannot be predicted without simulating every step along the way.
If the universe is one giant cellular automaton, as Wolfram postulates, there would be no computer big enough to run such a simulation (all computers are a subset of the universe?).

Therefore the future state of the universe is unknowable even though it is deterministic.
I.e. our decisions might be deterministic, yet they are still unpredictable if we live in such a cellular automaton universe.

See more here.

3.2. (Problem?) Replacing biological body parts with artificial ones.

Kurzweil also tackles the Cyborg problem.

I.e. We naturally undergo a gradual replacement throughout our lives. Most of our cells in our body are continously being replaced.
So, (in the future) what about replacing parts of our brain with non-biological processing units:

You keep opting for additional procedures, your confidence in the process only increasing, until eventually you changed every part of your brain...

Are you still you?
Kurzweils resolution to the dilemma is:
''It is still you. There might even be more copies of you. If you think you are a good thing, then two of you is even better.''.

Here, Kurzweils leap of faith is, that identity is preserved through continuity of the pattern of information that makes us us.
Continuity does allow for continual change, you might be somewhat different from yesterday, your identity might nevertheless be the same.

One note though: Biological systems systems cannot be copied, backed-up and recreated. But the non-biological system can. Somehow it seems threatening to personal identity, if there exists a million copies of you? :-)

And, if ''you'' have improved a million times, and thinks a million times faster than you did before these future cyborg procedures
- are you really honest, if you say that this is just the same....
Well, we will be wiser when we get there ....

See more here.

3.3. Reverse Engineering the Human Brain.

Kurzweil seem confident that it will eventually be possible to reverse engineer the human brain.
Certainly, in Kurzweil's mind, the construction manual for a human being is on a scale that we might actually be able to understand:

The contention that every structure and neural circuit in the brain is unique and there by design is impossible, for it would mean that the blueprint of the brain would require hundreds of trillions of bytes of information...
...
(But) The design information in the genome is (only) about 50 million bytes, roughly half of which pertains to the brain [p. 271].

So, for Kurzweil, the real mountain of information inside our heads is what we have learned and experienced.
A mountain he will be happy to see grow taller in non-biological intelligences:

Although there is considerable plasticity in the biological human brain, it does have a relatively fixed architecture, which cannot be significantly modified, as well as a limited capacity. We are unable to increase its 300 million pattern recognizers to, say, 400 million unless we do so nonbiologically. Once we have achieved that, there will be no to stop at a particular level of capability. We can go on to make it a billion pattern recognizers, or a trillion [p. 280]

For Kurzweil it will be possible (even though others might be less optimistic):

Critiques along the lines of Allen's also articulate what I call the ''scientist's pessimism''. Researchers working on the next generation of technology or of modelling a scientific area are invariably struggling with that immediate set of challenges...
...
They might be cautiously confident of reaching the next goal, but when people predict goals a thousand times grander they think it too wild to contemplate [p. 273].

In the end, Kurzweil shares Pierre Teilhard de Chardin (and others) grand dream
(that the universe is evolving towards an Omega Point, a place with a maximum level of complexity and consciousness).
Which makes him optimistic that progress towards the human-machine singularity is needed and ultimately beneficial:

Cosmologists argue that the world will end in fire (a big crunch) or ice (the death of stars as they spread out into an eternal expansion), but this does not take into account the power of intelligence.
...
Waking up the universe, and then intelligently deciding its fate by infusing it with our human intelligence in its nonbiological form, is our destiny [p. 282].

See more here.