[Hierarchal Temporal Memory (HTM) and NKS] - A New Kind of Science: The NKS ForumA New Kind of Science: The NKS Forum
Pages:1
Hierarchal Temporal Memory (HTM) and NKS
(Click here to view the original thread with full colors/images)
Posted by: g48150
http://www.numenta.com/Numenta_HTM_Concepts.pdf
We are all living, breathing examples of universal Turing Machines.
Our Neocortex is the foundation of every facet of our lives. There is one function that the brain provides, and uses seemingly endless "feedback" loops that make up the rest of our brains.
That is the very definition of a universal machine.
Posted by: Garrett Neske
I find the notion of emulating the dynamics of neural systems with simple machinery to be a fascinating route of research. Even more interesting is the question, do neural systems perform universal computations and how simple does such a system need to be to be universal? I happen to think that neural complexity is not in direct correlation with, for instance, the number of neurons in an organism or the number of synaptic connectivities, though this is the traditional, biophysically based definition of neural complexity. Consider the phenomenon of language, for example. Human language is sophisticated enough to refer to ideas and objects out of context, which is indicative of the cognitive capacities of the human. Yet, the honeybee has vastly less neurons than a human, but has a continuously infinite language (in contrast to the discretely infinite human language) that can refer to to the location of food outside of the immediate reference frame of the hive, as was discovered by von Frisch.
You mention the neocortex, which is strictly an aspect of the mammalian brain. If Wolfram is right about thresholds of complexity, does it not seem likely that a much simpler brain architecture could perform universal computation? While the notion of creating a machine at the computational capacity of the neocortex is exciting, I think it is more exciting to consider what might be the simplest neural system in the animal kingdom that is universal. I believe that such a system will almost certainly be found in an animal less evolutionarily derived than a mammal, perhaps even less so than a honeybee.
As always in computational neuroscience, one must be very cautious not to ignore the biophysical basis of one's models. The Hodgkin-Huxley equations, which form the basis of mathematical and computational neuroscience, are based on experiments on the ion channels of the squid giant-axon. One cannot simply formulate a neural model and then go search for it in the animal kingdom; neural systems are too nonlinear, and sometimes even chaotic, for that hypothetico-deductive approach. It seems to me that the HTM is based more on psychological experiments than on experiments dealing with the biophysical (i.e. molecular and cellular) bases for neurocomputation. I think this is also a serious issue for so-called "cognitive scientists," such as Marvin Minsky. We must never forget about the wet-lab data! But I also think that computational neuroscience needs a new direction. Even the Hodgkin-Huxley equations are highly nonlinear partial differential equations and these explain single neurons! We have not even gotten to neurons in a network! What we now know from Wolfram is that models based on simple rules can also model complex, nonlinear systems. In every computational neuroscience book or article I have read, the complexity of the mathematical model increases vastly with the complexity of neural system under investigation. These means that, often, only numerical approximations are available to the neuroscientist, and these approximations are not necessarily reproducible from computer system to computer system. And at this rate, it will never be possible to find the simplest neural system at the threshold of complexity because no one would agree on it! Those familiar with NKS, however, know that this need not be the case if one concentrates on simple computational systems. For the purposes of finding the simplest neural system that is universal, it is now necessary to frame the entire field of computational neuroscience in terms of simple computational systems. The study of neural complexity can no longer be carried out on the basis of numerical approximations, for these say very little about universality. We must reconcile the wet-lab biophysical data from various animals with Wolfram's computational systems.
Forum Sponsored by Wolfram Research
© 2004-2008 Wolfram Research, Inc. | Powered by vBulletin 2.3.0 © 2000-2002 Jelsoft Enterprises, Ltd. |
Disclaimer
vB Easy Archive Final - Created by Xenon and modified/released by SkuZZy from the Job Openings