Microsoft Word Kurzweil, Ray The Singularity Is Near doc

The brain's circuits are very slow.
Synaptic-reset and neuron-stabilization times (the amount of time required for 
a neuron and its synapses to reset themselves after the neuron fires) are so slow that there are very few neuron-
firing cycles available to make pattern-recognition decisions. Functional magnetic-resonance imaging (fMRI) 
and magnetoencephalography (MEG) scans show that judgments that do not require resolving ambiguities 
appear to be made in a single neuron-firing cycle (less than twenty milliseconds), involving essentially no 
iterative (repeated) processes. Recognition of objects occurs in about 150 milliseconds, so that even if we "think 
something over," the number of cycles of operation is measured in hundreds or thousands at most, not billions, 
as with a typical computer. 
•
But it's massively parallel.
The brain has on the order of one hundred trillion interneuronal connections, each 
potentially processing information simultaneously. These two factors (slow cycle time and massive parallelism) 
result in a certain level of computational capacity for the brain, as we discussed earlier. 
Today our largest supercomputers are approaching this range. The leading supercomputers (including those 
used by the most popular search engines) measure over 10
14
cps, which matches the lower range of the estimates 
I discussed in chapter 3 for functional simulation. It is not necessary, however, to use the same granularity of 
parallel processing as the brain itself so long as we match the overall computational speed and memory capacity 
needed and otherwise simulate the brain's massively parallel architecture. 
•
The brain combines analog and digital phenomena.
The topology of connections in the brain is essentially 
digital—a connection exists, or it doesn't. An axon firing is not entirely digital but closely approximates a digital 
process. Most every function in the brain is analog and is filled with nonlinearities (sudden shifts in output, 
rather than levels changing smoothly) that are substantially more complex than the classical model that we have 
been using for neurons. However, the detailed, nonlinear dynamics of a neuron and all of its constituents 

---

(dendrites, spines, channels, and axons) can be modeled through the mathematics of nonlinear systems. These 
mathematical models can then be simulated on a digital computer to any desired degree of accuracy. As I 
mentioned, if we simulate the neural regions using transistors in their native analog mode rather than through 
digital computation, this approach can provide improved capacity by three or four orders of magnitude, as 
Carver Mead has demonstrated.
13
•
The brain rewires itself.
Dendrites are continually exploring new spines and synapses. The topology and 
conductance of dendrites and synapses are also continually adapting. The nervous system is self-organizing at all 
levels of its organization. While the mathematical techniques used in computerized pattern-recognition systems 
such as neural nets and Markov models are much simpler than those used in the brain, we do have substantial 
engineering experience with self-organizing models.
14
Contemporary computers don't literally rewire themselves 
(although emerging "self-healing systems" are starting to do this), but we can effectively simulate this process in 
software.
15
In the future, we can implement this in hardware, as well, although there may be advantages to 
implementing most self-organization in software, which provides more flexibility for programmers. 
•
Most of the details in the brain are random.
While there is a great deal of stochastic (random within carefully 
controlled constraints) process in every aspect of the brain, it is not necessary to model every "dimple" on the 
surface of every dendrite, any more than it is necessary to model every tiny variation in the surface of every 
transistor in understanding the principles of operation of a computer. But certain details are critical in decoding 
the principles of operation of the brain, which compels us to distinguish between them and those that comprise 
stochastic "noise" or chaos. The chaotic (random and unpredictable) aspects of neural function can be modeled 
using the mathematical techniques of complexity theory and chaos theory.
16
•