[bkumnick]

Hi everyone. I just joined the board and thought I'd introduce myself.

I just got Jeff's book yesterday and skimmed through half of it last night. Very interesting...

I've been a software engineer, software architect, and software project manager since 1987. I've also spent the last 20 years or so studying AI, machine learning, neural science, computational neuroscience, philosophy of mind, ontology, and knowledge representation. Like Jeff, I find myself driven to build machines that think based on the principles used by the brain.

I have come to many of the same conclusions I read about in Jeff's book, but from the opposite direction. Instead of starting with the architecture of the cerebral cortex and working my way down, I started with a detailed study of single neuron computation, and dendritic integration and worked my way up. My current research interest lies in neural knowledge representation. I've made more progress in that area than anybody in the world as far as I know. Jeff's focus seems to be centered on memory. Mine is centered on the representation of thought.

I've cracked the neural code. I've discovered how single neurons represent abstractions and concepts. I've discoved how neurons represent meaning, and how self awareness works at a neural level. I've also discovered what allows us to think in context and in context free fashion, think associatively, think in terms of similarities and differences, think in terms of analogies, perform conceptualization, and generalize. I just recently developed the core mathematical algorithm responsible for all neural processing. It is smoking fast and incredibly efficient. It basically provides a mathematical transform that allows neurons to represent how anything relates to anything else in any number of dimensions in any combination of dimensions in any context at any level of abstraction and transform all of those equations into a single 4 dimensional equation for processing via spatio-temporal correlation. Think about the ability to perform the equivalent of vector addition or dot products, but in any combination of any number of dimensions simultaneously. This gives a neuron the innate ability to represent and solve n-dimensional problems where each term in the equation can have a different number of dimensions, yet be able to solve the entire system of equations, by converting it all to 4 dimensions and representing all of it in terms of how things relate to other things in space-time...

Regards,

Barry

[danshawen]

Sounds like a great start to a mathematical representation of a reptilian brain.

Now you'll be needing to integrate that with sensors or sense organs, and somehow I don't think that is going to work "symbolically" as the rest of your approach strongly suggests, you are married to.

Process using only lengths (what sense organs can input to the neurons) if you wish to impress us.

[bkumnick]

Actually, the neural processing algorithm does process everything as "lengths" -but they aren't 1 dimensional scalar lengths. They are 4 dimensional vector "lengths". Biological neural systems transform and reduce all input into 4d space-time "lengths" (vectors) for internal processing. They relate sensory signals and abstract representations of objects and relations to each other at multiple levels of abstraction in context, in spacetime.

No symbolic information processing is involved. There are no symbols. It's pure vector processing. It's all number crunching. That is why its so fast.

The neural processing algorithm is very unusual. It is a fixed algorithm, but it operates within the confines of the 3-dimensional spatial topology and geometry of the biological neural network. The 3 dimensional geometry of the neural network forms the basis for the representation of 4 dimensional spatiotemporal relationships between the flow of electrotonic potentials as they flow through the dendritic trees and are integrated. The 3 dimensional geometry of the dendritic trees and the relative timing and location of synaptic firing patterns represent the input to the neural processing algorithm. There is only one processing algorithm, but the functions it computes vary as a function of the spatial geometry of the neural networks' dendritic trees and the relative timing and location of synaptic activation potentials in its input.

In computer terms, the 3 dimensional shape of the dendritic trees and the pattern of synaptic connections in the neural network represent the neural systems' "memory" and stored program. The flow of electrotonic potential represents execution of the program. The neural processing algorithm is analogous to microcode that defines the "neural instruction set". The fine grained shape of the dendritic trees is dynamic. The detailed 3 dimensional structure of the network and its synaptic connections are refined and "grown" based on what the network experiences. In essence, the network optimizes its memory storage, knowledge representation, and program based on what it experiences and what it learns. Consequently, processing and storage both become more efficient as the network learns better ways to organize and represent its knowledge.

The mathematical and logical functions computed by the neural network can (sometimes) be expressed symbolically, but doing so is not practical. The functions and logical expressions computed by the network are usually extremely lengthy and complex in symbolic form. They often involve equations and logical expressions that contain tens of thousands to millions of related terms. Evaluating these expressions symbolically would be prohibitively expensive, both in terms of computation time and storage. None of these logical expressions are programmed or stored in memory by a programmer. The network can be designed to generate and output the symbolic information based equivalent of neural processing as a byproduct of neural "execution", but the equations must be generated dynamically, on the fly, as a by product of neural execution. The network does not use symbolic information for internal processing. It can use it for external communication, but doing so is relatively innefficient (and error prone). Not all neural processing can be expressed symbolically, and even when it can be, the equivalent symbolic expressions are too voluminous and too complex for most practical uses.

Best regards,

Barry Kumnick