[mikehahn]

Here is the scenario: you have a fast CPU, with between 4 and 16 GB of RAM to work with, along with a typical 250 GB hard drive. Can a simplified neocortex be modeled under this scenario? Our goal is to replicate intelligent behavior, such as learning how to play simple computer games, or even language. Each neuron's size is a power of 2, between 256 and 32,768 bytes. The axon side of each synapse is a 32-bit pointer, and the dendrite side, also 32 bits, contains (among other things), the strength of the synapse, along with a 4-bit value indicating the size of the parent neuron. The beginning address of each neuron is always a multiple of the size of that neuron. So given the axon pointer and the dendrite it points to, it's easy to compute the beginning address of the destination neuron. Since each synapse is 8 bytes, we can fit in RAM between 2^29 and 2^31 synapses (500 million to 2 billion).

I realize that the human neocortex contains trillions of synapses, but what if the environment of our potentially intelligent entity is made drastically simpler than our own sometimes chaotic, fast-paced world? I'm talking 2 dimensions, and when it comes to language: one ASCII character in or out at a time.

Under this scenario, using current knowledge of how neurons work, and how cortical columns are wired together, along with the framework developed by Jeff Hawkins in "On Intelligence," is it possible to model intelligent behavior? What do you think?

I am a newcomer to the field of computer modeling of neural networks. For more information about my own project, which I call a Toody Machine, and which elaborates on the above scenario, I invite you to visit:

http://www.treenimation.net/toody.html

Mike Hahn

[deespona]

According to my propossal at "what if we have no memory at all" topic, I´m convinced its possible to implement human mind purely in software, and this have few to do with hardware advances, but in how the implementation is done. Also I think it has no or few connection in how the brain works at a biochemical level, but in how the imput patterns -images, noises, tactile info- are processed to obtain output patterns -muscular orders, new image or sound patterns, etc- in a different brain area.
If you check my old theory about a "no memory data stored in the brain" model, the method is based in a flexible neural network that is constructed or changed in the course of our lives, connecting input pattern areas of the brain to other output pattern areas in the brain. In this way, an input "sound" pattern can generate an "output" image pattern in another brain area, then memory is not stored in any part but the rendered result of an analogical neural connection path from one zone of the brain to other.
I think these tridimensional, constantly reconfigured paths created in our brain can be simulated by a tridimensional simulation of the whole net in a given moment, probably without programming, but with the 3d duplication of the brain´s net in a virtual frame buffer.
I think the brain transform the input patterns to other patterns in two ways: 1) by creating a phisically, long term connected network of synapses between neurons during the sleeping process to obtain instant output patterns from input patterns and 2) temporal, electromagnetical connections between different neural paths to obtain temporal pattern output responses during vigilia, and imprinting new neural areas for creating new more stable phisical synapses during the next sleeping process.
Then the steps to obtain a software based duplication of the brain must be done in the following steps:
1. Scanning tridimensionally the human brain at a under micron resolution
2. cleaning the obtained tridimensional voxel model of the synapses for creating a simplified wire connection 3d model of the human brain.
3. assigning two ways of electromagnetical connectivity to this virtual 3d wire representation of the human brain, one physically connected (only connected networks transmits the pattern) and other by proximity (close neural paths can capture patterns from one path to another to access different brain areas)
4. detecting input patterns areas and output pattern areas in this virtual 3d brain, without the difficulties generated by analizing the brain´s electrical activity in a living organism.

If all is ok in the propossal, a mind performance can be duplicated virtualy,-that is , constructed in software- simply by duplicating the neural network shape in a virtual frame buffer, without programing the mind mechanisms and without researching the mind activity at a biochemical level.

The shape is the way, also for the mind!

Jose Maria de Espona
http://www.3dturtles.com

[mikehahn]

[deespona]

yes you are right, no scanning at 1 micron available today, but some systems works in the direction. But maybe at a lower scanning resolution as those available today at 0.01 square mm can allows a rough 3d representation of the main neural networks, probably enough for checking the theory. As the voxel model is transformed to a single 3d wired simulation of the brain, the requeriments of RAM can be much more lower, and current memories can surelly allow a framebuffered 3D representation of the brain network.
But pls try to get the essential idea: duplicating brain activity in a virtual memory by constructing a 3d model from a 3d scanning of the brain, to check if the brain neural network works like an analogical filter that converts an input pattern into another output pattern that is read in a different brain area, that is, not checking an HTM hawkins´memory based system but an analogical signal filtering system based in neural networks that works processing an input pattern for rendering an output pattern, a model that don´t requires programing but simply a virtual network 3D shape duplication that manage patterns.
A newborn´s brain works practically without neural connections, but those previously constructed under DNA orders (innate knowledge and pattern input paths). The fact that external input pattern stimulation can activate neural connections that are constructed during the sleeping period is the base of the theory. If we want to construct an artificial intelligence system that resembles the human or animal brain we must work with the same utilities: physical and electromagnetical neural conections without previous programing, but those more basical to allow the initial input of the patterns into the neural network and those that allows the reading of the output patterns processed in the 3d net.
Im also interested in different approaches to AI based in much more empirical methods, as those propossed in my 3d turtles system (www.3dturtles.com) that is based also in a conceptual principle about intelligence based in 3d shape understanding, but this is not the true target of a forum like this one, that is , making AI based in the human brain performance, as supposedly Jeff is doing with his neocortex simulation based in memorized patterns. But my key question here is if memory data is really stored in the neocortex. or as contrary the neocortex is a simple analogical filter that converts input patterns into output patterns, then processing again these output patterns in sucessive cycles (that is, the thinking process).
Then I agree with your view in the software implementation of the brain, but not necessarely in the programming of this vitual implentation. The shape of the human brain´s network can be enough for emulating it without knowing what in the hell the brain really does in a so deterministic way.

Jose Maria de Espona
http://www.3dturtles.com

[mikehahn]

If anyone out there happens to be monitoring this thread, I would just like to announce that I have added 2 paragraphs to the bottom of my Toody Machine web page, under the heading: "Toody vs. HTMs". Also, the Toody Machine is now an open source project. The main Toody Machine web page is located at:

http://www.treenimation.net/toody.html