CrX Architecture Part-1

 

Documentation of CrX architecture

 

            This Documentation talks about the CrX – Cognitive Revolution X architecture with a detailed step-by-step explanation, which was developed from BrainX theory. This architecture will initially function as a general problem solver/AGI.

Ø  N – network

Ø  P – partition gives specification

Ø  C – cluster gives direction / dynamics

Ø  R – returning of output

 

Ø BrainX theory into CrX architecture

The brain processes input and allows it to travel throughout its pathways, regardless of where it originated (it does not discriminate). Neurons serve as the paths for impulses to travel, with these paths branching in various directions, thereby altering the direction of the impulse at any given time. The impulse is not consciously choosing its direction; rather, it changes direction due to the laws of physics and biological assistance. Several factors determine the direction of the impulse, and a regularly attended pathway has a higher probability of impulse transmission (thus, a higher likelihood of the impulse traveling this route). When impulses from the sides converge into the main impulse, it alters the direction of the main impulse. Of course, there is considerable noise in the brain – a challenge addressed by the concept that a large sum of correctly functioning impulses is sufficient for output (60% of correct impulses offset 40% of noise). The neuron's diverse paths and their changing direction abilities are equivalent to clusters that alter directions. Unlike the brain, the direction-changing in this architecture depends on factors like the presence of clusters around the chained neurons. Even though the brain does not discern the type of impulse it receives, it processes impulses because they change the physiology of neurons. It is the input and output regions in the brain that give representation to the impulse. Following the same principle, denser regions in the input region prioritize receiving impulses as they receive more impulses than non-denser regions (more impulses sustain the propagation of impulses in the brain, aiding in noise reduction). This is akin to the encoder and decoder having a probability density distribution function that assists in the destabilizing mechanism – a mechanism aimed at reducing noise in the model. Finally, the wiggle connection, where outputs are merged to yield a combined final output, is equivalent to the brain's formation of shortcuts between input and output, thereby bypassing the computational part. The network replicates the property of neurons clumping together into layers. Partitioning and clustering replicate the plasticity of neuron properties.

Ø Complete picture of the architecture

The network

 

Ø What is network

Networks are the millions of layers that are combined together. Initially, we are going to start with only two networks that contains 10 layers each (as a prototype). If it works well, then we can increase the number of networks. Increasing the number of networks will help in storing more different information, which will bring more understanding of the incoming information to the model. However, this will reduce the information reduction capacity (information will be reduced to its smallest truth due to over-consumption – same like brain) of the model, which will increase the correct output rate (other effects are unknown).

[Smallest truth – it is reduced bits of information; each model can have its own smallest truth when given the same input. It largely depends on how the model is trained.]

In the layers, there are random neurons that get attracted to the chained neuron if it contains any values (similar to how the impulse of active neurons in the brain attracts other resting neurons via chemical cues). The random neurons are allowed to move randomly, but if they are within the clusters around the chained neuron, then they are fixed onto that position. This fixation of random neurons around the chained neuron helps us to activate the right set of chained neurons for every input. They do not reduce the information (not yet), but they just use the already-in-place information (cluster).

The brain exactly does the same thing by activating the right set of neuron clusters and outputting the reduced form of output. The type of reduced format is dependent on the type of information that the brain regularly stores. This has the capacity to induce curiosity and a problem-solving attitude in the human brain. Since it cannot be fully followed by this prototype, curiosity and giving up on the problem are not possible at the maximum level.

The first network processes inputs, and the second network receives combined clusters. The output is again fed as input to the first network to update the clusters in the first network and to put required clusters within the same layer, which will be in different layers and it will not be able to process input. Because of the wiggle connection, processing input is made possible, and the rate and accuracy are increased. This model should be continuously updated, and with the destabilizing mechanism in action, it can solve problems.

These networks, which consist of chained neurons in their layers, will act as input takers or input placeholders (in other words, the impulse generated by input in the brain is equivalent to chained neurons containing the input values in this model).

The layers contain vectors in their spaces, which are used to form clusters around the chained neurons. The clusters around the chained neuron will act as plasticity in the network. Impulses in the brain are transferred to other neurons only by plasticity (also known as connections). The same principles are applied here. The input values of chained neurons from one layer are passed to another layer only if there are clusters around the chained neuron. In this way, direction for input can be created, which is analogous to what the brain is doing with information, just changing the direction of impulses, and leading to activating a certain set of outputs.

Math behind the network

The network layers consist of empty spaces where chained neurons are fixed to each other at fixed intervals. These chained neurons take the scaled pixel values, with each layer containing 100 chained neurons receiving scaled pixel values. The random neurons serve as vectors, and they move randomly within the spaces. In each layer, the number of random neurons can be up to 500.

The input enters the model by chained neurons and the block diving function divides the block and partition of the clusters eliminates noise and clustering fixes the random neurons. These clustered active chained neurons value are transferred to second network respective layers via wiggle connection.

 

Ø Block dividing function

The layers spaces that consist of random neurons and chained neurons were first divided into equal parts to specify random neurons to specific locations in the layer. This idea of dividing into blocks was to reduce complexity and increase the activation of the right set of neuron clusters (to activate most of the chained neurons, which helps to create pattern), which leads to the right output. Dividing the space into blocks is a way of maintaining the equal dispersion of random neurons. Because of over-concentration, it takes away the ability to find the smallest truth. This gives specialization to random neurons' activity in their respective blocks.

The brain provides spaces for each cluster to grow, and they won't clump all things together. Neurons are connected to the spaces more than to other neurons. So, to give their own spaces for the random neurons, we have to insert specifications for them to stay and connect to the chained neuron.

Math behind the block dividing function

This function divides the spaces of the layer to confine or limit the movement of random neurons within the blocks of this function. The spaces are first divided into equal blocks then the scaled pixel values are inserted into each chained neuron. And the chained neurons that contain value are active so they will attract the random neurons towards the chained neuron. Chained neurons that contain value of 0 are unactive and they will not attract the random neurons in the space. and only the random neurons within the blocks space are attracted to that particular blocks chained neuron active.

Ø Loop

This model follows a continuous closed loop to continuously update the clusters in the layers. The first network's layer output goes to the respective layers of the second network to store the required clusters nearby with the help of the wiggle connection. Each output from the second network is tested for similarities within each output. If the outputs are the same more than two times, then the output is returned back as input to the first network.

Ø Increase of Network effects

The network was the way to store and mix clusters by increasing their numbers. Both the mixing-up of clusters and storing can be increased. However, complexity will rise and optimization has to be adjusted. For now, going with simple two networks would be good.

 Ø Biologically - inspired explanation

The network layers of this architecture mimic the property of clusters of neurons being condensed into thin layers. Not all properties were mimicked, but only the essential ones necessary for processing information in the brain.

The partition

Ø What is Smallest truth?

Random neurons are vectors in the layer, so after many iterations of pixel input values into the layer and with the work of the destabilizing mechanism, there may be only a few small clusters all over the layer. They do not get reduced but can increase with small increments, but stop after a certain limit, as the input region has an inbuilt predefined intensity limit. So, all these functions work together to extract the smallest truth from the input and help in problem-solving.

Math behind the smallest truth

When the model continuously receives information from the source, the clusters undergo significant changes, but smaller clusters that persist or remain unchanged even after numerous information inflows into the model are referred to as the smallest truth of reality. The attachment of these vectors (random neurons) to these chained neurons always indicates that these chained neurons are consistently active in response to most of the input information from the source.

 

Ø What is destabilizing mechanism?

The destabilizing mechanism is equivalent to the activity of neurons not being connected to one neuron always; neurons are connected to locations, not to other neurons. This was the reason for introducing the concept of the destabilizing mechanism to this model.

The destabilizing mechanism makes the model continuous. That is, the input will continuously form and remove clusters, with the exception of true facts or small truths remaining always as clusters, as they are always getting activated.

There should be a preference for some chained neurons, and it should be possible by partitions, as where blocks get narrower, the random neurons should increase in those blocks. Think of each random neuron as an impulse in that chained neuron. As more impulses in that chained neuron, the higher the probability of that chained neuron's values being the output.

 

Ø Partition role in destabilizing mechanism

Partition is the first step towards the destabilizing mechanism. It reduces over-concentration by eliminating noise in the layer. This helps to find the smallest truth and eases the process of the destabilizing mechanism.

Math behind the partition

First, the spaces are divided by a block dividing function, and then random neurons are attracted to the active chained neuron. Next, a partition algorithm, like a quadtree, is applied to partition the layer space. The partition algorithm continues partitioning the space until only 4-5 random neurons remain inside each partitioned space. Then, this partitioned space containing 4-5 random neurons is subjected to a clustering algorithm.

 

Ø Partition role in finding the smallest truth

By filtering the noise and preventing the clustering of unwanted random neurons in the layer, sometimes there will be a loss of information by this partition mechanism. However, if that lost information is one of the smallest truths, then it will be stored eventually by other types of inputs.

 

Ø Biologically - inspired explanation

The assumption is that neurons are not limited to their connections in the brain but rather to spaces (this also contributes to the problem-solving abilities of the brain, by mixing similar information – connecting randomly with closer neurons). The block dividing function and partition algorithm were introduced to replicate this property of neurons in the brain.