Nemulator: A Neme Emulator

Jared Schiffman, Max Van Kleek, Tom White

Overview

Central to the Society of Mind are the concepts of 'nemes', which are agencies responsible for representing fragments of states of minds. Together with K-lines, they form critical bridges interconnecting agents in different ways. One variety, polynemes, bridge together the major properties of an idea or concept, such as color, shape, size, or substance. Other varieties, such as ascending and descending nemes, interface with K-lines, recognizers, and memorizers to create feedback loops in the mind. These memory-neme feedback loops enable the mind to, among other things, retrieve and organize ideas by their similar properties, and identify or describe ideas in words.

Since the concept of nemes seems to help explain numerous processes of the machinery of the mind, it would prove useful to evaluate the theory by determining how feasible it might be to build nemes as physical systems. One way to do this is to build a simulator which could experiment with various topologies of nemes, and demonstrate how each perform their basic duties outlined in Society of Mind.

Our project will be to create a Neme simulator that illustrates the three axes of topology, function, and environment. Each of these axes will be illustrated by a system that allows active manipulation and graphical visualization of nemetic structure, its interactions with agents, or exposure to a variety of stimuli. Additionally, since the human mind would have to contain millions of complex interconnected nemes in order to enable us to recognize, memorize, and associate concepts the way we do, we will in this project, discuss the scalability of nemetic structures.

The objectives of our simulator are to create a system of nemelike structures to accomplish the following:

to learn sensoral input; that is, to associate sensoral inputs to form a concept
to recognize sensoral input; that is, to identity incoming stimuli and retrieve related concepts
to combine recognition and learning, so that concepts activated by sensoral input can be re-shaped by new input
do all of the above using simuli as input as opposed to stimuli, that is, to use ideas as input to allow us to retrieve associated properties and other ideas.
to all of the above in a network that can be generated from a simple neural growth model.

Goals and Precedents

The goal of our nemetic structure in general is to construct a reliable auto-associative memory that exhibits properties of human memory and neurophysiology. For example, this memory should store concepts in a network, such that it does not require dramatic rewirings of neurons to learn concepts. If such rewiring was in fact necessary, it would not be consistent with our current theories of our neurophysiology, which contest that only minor local reconnections can be done in the adult human brain.

Several earlier neural-network based auto-associative memories exist. Perhaps the closest network is the Hopfield Network (1, 2), which is an auto-associative network that uses fully-connected, bidirectional connections and performs annealing to memorize and recognize patterns in arrays of digital inputs. This network is frequently used in the domains of Optical Character Recognition and Image Restoration. However, there are several weaknesses inherent in Hopfield Networks that led us to seek an alternate representation. The foremost weakness is that Hopfield networks exhibit the odd characteristic of being unable to distinguish a stimulus from its exact inverse. Secondly, since they require digital inputs, it would be difficult to adopt it to an analog input space, where an object could have any varying intensity of a property.

Topological Structure

Organization

One of primary intents of this project is to determine what topological structures we may use to build structures that can function as nemes. The structures in this system will be represented as weighted Directed Acyclic Graphs (DAGs) that connect nodes representing sources of stimuli, input nodes , to nodes representing concepts, concept nodes. The responsibility of these topological structures, thus, to learn associations between particular input nodes and concept nodes.

Learning in the system is accomplished by activating the input nodes, which induces edges connected to them to increase their conductivities. Edge conductivity is also influenced by the concept nodes to which the edges are incident; specifically, an inactivated concept node will discourage an edge from becoming more conductive. If the stimulus is strong enough and the conductivity is pulled up adequately, the signals reaching the concept node may begin to excite it. This, in turn, will encourage an increase in conductivity in a positive feedback manner. The result is that the edges from relevant stimuli are pulled up to their maximum conductance, and the concept is learned (see Figure 2).

All concept nodes in the system start in an unallocated state (see Figure 1). This means that the weights of the paths to the nodes from inputs start out unbiased. As the system learns concepts, paths from sets of stimuli nodes to particular concept nodes become more heavily weighted (see Figure 3). When those stimulus nodes are re-activated later, the same concept nodes as those during the original learning process become activated, indicating that the concept was successfully recalled.

An extra structure, called the supressor bus links all concept nodes. This structure aids in "locking-in" to a concept, by permitting nodes excited by a particular sensory input to supress all other nodes on the bus. This way, only the first concept to be excited by a stimulus, or the most excited concept may possess the network's attention. Without this mechanism, two or more concepts may be excited simultaneously.

Growing Topologies

In addition to trying a variety of simple topologies for the nemetic stcutures in our simulation, we allow the topologies to grow in an organic fashion. This is an attempt to use an automaton-based computational model to represent the period of neuron growth of the cortical development phase in early infancy. Neurons which will be used as nemes in the manner described above are constructed by the automata to create paths from input nodes to concept nodes.

Recent research led by the department head of MIT Brain and Cognitive Sciences, Dr. Mriganka Sur, has revealed the high plasticity of neurons during development, particularly in the visual cortex. Namely, he has shown that by slightly modifying electrical signals, neurons originally intended for connecting the visual cortex may be rerouted to an entirely different region, such as the auditory thalamus. Further, he has shown that external sensory stimuli may have the same rerouting effect, as he has demonstrated with his experiments with the visual cortices of ferretts (Sur 1). This discovery has led us to attempt to model this in our organic growth of neuron structures, by allowing external stimuli to influence and govern the final structure of the neurons.

To model this phenomenon, the neurons being formed are immersed in a space filled with scents, which are fields that induce positive and negative attraction towards particular nodes. The external stimuli induce these scents to encourage the automata to connect the neuron to relevant agencies, and to ensure a uniform distribution of connections (See Figure 4).

When the nemetic structure is grown in this fashion, input nodes and concept nodes must be chosen carefully to ensure invariants are preserved, such as preventing input nodes from feeding into other input nodes. This process of structural interpretation also untangles the grown web, and displays a schematic of an equivalent, untangled structure.

Sensory Inputs

The sensory values at Input Nodes represent measurements of a particular property, such as temperature, softness, or heaviness. Rather than being raw sensoral inputs, such as a representation of light intensities across the retina or a audio waveform as detected by our ears, we will assume that some lower-level subsystems will have already performed appropriate preliminary processing and feature detection. We believe this assumption is justified, for, as described in his famous work "What the Frog's Eye Tells the Frog's Brain", Prof. Jerome Lettvin in the 1950's found that in creatures such as the frog, significant processing is done directly in the sensoral organ prior to even entering the brain. Lettvin describes one such processing as "a bug detector", that sends a signal to the brain as soon as the retina detects a moving convex shape of a particular diameter in its field of vision. Other types of signal processing are likely to be in human sensoral organs as well, and feature detection may be handled by other agencies in the brain.

Results

Unfortunately, due to time constraints, we were unable to develop our simulator to the extent we had hoped. In this section, we will outline what we were able to accomplish, and what further research is to be done.

Learning

Our initial experimentation with the neme network structure discussed in the first section of "Topology" yielded encouraging results. The system successfully demonstrated the ability to learn an association between both single-stimulus and multi-stimulus situations, provided that these stimuli were influential enough and disjoint (please see "Limitations" below). Additionally, two desirable properties emerged from the network; first, the network does not distinguish between learning and recall modes, and thus could be said to be constantly learning. Secondly, it exhibits a slow learning strength decay when the system is exposed to more and more stimuli. We decided that this behavior was a desired phenomenon, since we would expect that the greater the length of time elapsed since a human has revisited a stimulus he or she remembers, in most cases the less strong the associations should be.

Additionally, we were able to successfully separately implement and were successful in getting a favorable and interesting neuron growing system.

Limitations

While individual sections of the nemulator were successfully implemented, they were only tested on simple cases in isolation from other parts. More work would need to be done to integrate the sections to a system as originally envisioned.

The topological growth simulator is very effective at growing neural looking structures. Features from these structures can be extracted into various types of nodes and connection patterns. But to test different structures on different environments, a language describing the salient features of the neural structures would need to be developed to succently describe different classes of structures that could be grown.

The functional network showed that the operation of our network structure, including the suppressor bus, was strong enough to support strong orthogonal inputs and to associate these patterns into unique concept nodes. But the math did not scale to more complex inputs with multiple activations. We do not see this as a overall failure in our model, but believe that more investigation into a mathematical model of weight adjustments needs to be done for the general cases. We feel that our functional model was successful in that it was not opaque - the learned concepts were not distributed throughout the network but instead localized in polynemic nodes - and we feel that with care this result could be scaled up to more complicated input patterns.

Next steps

The environmental simulator we created was a simple fruit classifier that included associating locations for different types of fruit. The most natural immediate step would be to improve the mathematical model in the functional network to successfully classify the fruits into unique trained locations.

Once a viable functional netork model was established, it would be instructive to integrate the functional and topological network data structures. This would insure that the functional network is feasible from a simple neural growth model, which would give insight into the morphegenic requirements for our functional model.

Also, assuming the functional network could be mended, it would be instructive to try various other mappings between the environment and the functional network. The bidirectional input/output nodes of the functional network could be applied to various high and low level sensory inputs, attributes, or other concepts. The output could also be wired directly to motor control, for example in a robotic simulator.

With a topological network that generates a viable functional network that can successfully complete tasks in various environments, it would be interesting to generate many topological networks for a specific enironment and discover which topological paramaters correlate with different environmental processing. That is, given a concise language describing classes of topological structures, different topological possibilities could be tested directly against environmental simulations to see how the envionmental task at hand determines the structures necessary to support the necessary functionality.

Finally, it would be interesting to begin to scale multiple simple local networks into larger networks. Polynemic nodes at one level could become attributes at a higher level. Different topological structures specialized for specific tasks could be networked together to form more complex and less specialized tasks.

Appendix

The following diagrams figuratively illustrate the nemulator system.

These diagrams concentrate on the inputs to the nemulator. There would also be outputs from the systems with both input and output based on the current environment.

References

Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. In Proceedings of the National Academy of Sciences, pp. 2554-2558. National Academy of Sciences.

Lettvin, Jerome Y., H. R. Maturana, W. S. McCulloch and W. H. Pitts, ``What the Frog's Eye Tells the Frog's Brain,'' Proc. IRE 47 (1959) 1940--1951, reprinted in Warren S. McCulloch, Embodiments of Mind (MIT Press, 1965).

Minsky, Marvin. Society of Mind. New York: Simon and Schuster, 1986.