Something has occurred to me since my last post about where to implement the behaviour common to traditional neurons in the Clique Space(TM) Neuron.
Currently, no weights, activation, bias, or other well known mechanism commonly seen in traditional neurons (with the possible exception of structure that resembles a Clique Space synapse) is currently implemented in the Client Device, or the Neuron as an extension of the Client Device. It currently seems too premature to implement any traditional neurologic within these two projects; it instead seems better to delay implementation to an extension of the Neuron that can handle specific device types and can also implement any variety of neurological mechanism chosen by the organisation that implements such an extension, or no mechanism at all.
The Client Device and Neuron projects define the purpose and structure of qualia and Elements: Connections, Participants, Identities, the Sovereign - and all of their surrogates. The base protocol for the transmission of deliberations and their enclosed signals, as well as synapse and thinker structure are also defined primarily in the Client Device, but are extended in the Neuron to define similar structures that permit the transmission of messages. The Client Device uses the structure defined within it to facilitate the transmission of challenges only.
The rationale of my conclusion is the following: a relay Neuron is a specific type of Neuron; it acts in a mere signal forwarding capacity. These Neurons don't contribute in any substantial way to cognition, and therefore don't have a need to be equipped with the mechanisms - the Connection weights, activation, bias etc. - that equip a neuron with a greater ability to participate in cognition. A relay Neuron can extend the Neuron without addition of functional structure. Relay Neurons do not inherit anything more than those mechanisms that make a relay Neuron do what it needs to do if traditional neurologic is not implemented in the Neuron.
Hence, I think details relating to traditionally defined neurological mechanism can be reasonably deferred to the more specific types of Neuron that need these mechanisms.
Saturday, October 19, 2019
Friday, October 18, 2019
The exploration of a new hypothesis on or about the neurocognitive connectome.
It may all be bullshit, but thoughts over the past two days have led me down what I observe is an interesting garden path concerning the components of a Clique Space(TM) neural network and how a participating Neuron might conceive of itself and its neighbourhood.
Traditionally, a neuron shares physical connections with a subset of its neighbours. Each of these physical connections is assigned a weight and any given connection is assigned a value for its weight based on a process of "learning". That has always appeared as a reasonable if not abridged overview of the neural way things are generally known in this world.
I feel that I can give a slightly different slant on the above view. Clique Space is no different from this view in that Neurons indeed share similar physical connections (synapses) with a subset of their brethren; each connection is indeed realised through an algorithmic structure known as a synapse. However, what is slightly different is the following: weights are not assigned to synapses.
A Clique Space Client Device (a Neuron is a type of Client Device) is identified individually by a Connection. Connections are Elements which contain one or more features that are expressed by Participants (also Elements) which are members of Cliques. Cliques denote some cooperative endeavour is being undertaken. Synapses are a type of bipartite Clique which models the engagement of two Client Devices through the "this Element" feature of their individual Connections.
While a Client Device may not share a synapse with every one of its cohorts, it can still be aware of every Connection that makes up the entire population. Weights are assigned to the Connections known to a Neuron to represent other Neurons - even those Neurons with which the host Neuron doesn't share a synapse directly. Now, I think at this moment, there is an opportunity to do something rather useful with this knowledge; something that uses another structure developed over the past 11 years I have been working on my proof-of-concept. This something appears to me to advance the traditional connectionist model of neurons. Instead of assigning a weight to a synapse, weights are assigned to Connections.
A deliberator is a thinker (a type of Java thread) which is created to deliberate the arrival of a deliberation. Deliberators are associated with the Connection assigned as the first Neuron (actually Client Device but we will not split hairs here) to have entertained the topic being deliberated. Now, getting hypothetical, the wonderful thing about creating deliberators in this way is that a group of deliberators may be created to represent receipt of deliberations sent from neighbouring Neurons all of which are received at around the same time. Apart from the difference in originator, each deliberator is cogitating the same topic and hence each deliberator is constructed to have a identical parameters such as activation and bias because all deliberators are instructed by the same feature.
Accumulation of Connection weights for deliberators on this same topic may cross a threshold which when reached will cause the deliberator on which the threshold was crossed to send deliberations through synapses to its neighbours. Some neighbours may share synapses and may receive the deliberation immediately and others may only receive the deliberation through a neighbour who has faithfully passed the deliberation on.
This is all done by assigning weights to the Connections instead of to the synapses. I think that is a big deal, and an advancement in neuroscience that was published in this blog entry before anyone else in this world knew that assigning weights to representations of nodes in a Neuron's knowledge of its neural neighbourhood would be better than assigning them merely to the physical connections it would share with its neighbours.
This mechanism, and more as would answer most other questions about my concept is all there in my code. Really, it is... almost. I'm sitting on the greatest advancement in neuroscience since the perceptron!
Traditionally, a neuron shares physical connections with a subset of its neighbours. Each of these physical connections is assigned a weight and any given connection is assigned a value for its weight based on a process of "learning". That has always appeared as a reasonable if not abridged overview of the neural way things are generally known in this world.
I feel that I can give a slightly different slant on the above view. Clique Space is no different from this view in that Neurons indeed share similar physical connections (synapses) with a subset of their brethren; each connection is indeed realised through an algorithmic structure known as a synapse. However, what is slightly different is the following: weights are not assigned to synapses.
A Clique Space Client Device (a Neuron is a type of Client Device) is identified individually by a Connection. Connections are Elements which contain one or more features that are expressed by Participants (also Elements) which are members of Cliques. Cliques denote some cooperative endeavour is being undertaken. Synapses are a type of bipartite Clique which models the engagement of two Client Devices through the "this Element" feature of their individual Connections.
While a Client Device may not share a synapse with every one of its cohorts, it can still be aware of every Connection that makes up the entire population. Weights are assigned to the Connections known to a Neuron to represent other Neurons - even those Neurons with which the host Neuron doesn't share a synapse directly. Now, I think at this moment, there is an opportunity to do something rather useful with this knowledge; something that uses another structure developed over the past 11 years I have been working on my proof-of-concept. This something appears to me to advance the traditional connectionist model of neurons. Instead of assigning a weight to a synapse, weights are assigned to Connections.
A deliberator is a thinker (a type of Java thread) which is created to deliberate the arrival of a deliberation. Deliberators are associated with the Connection assigned as the first Neuron (actually Client Device but we will not split hairs here) to have entertained the topic being deliberated. Now, getting hypothetical, the wonderful thing about creating deliberators in this way is that a group of deliberators may be created to represent receipt of deliberations sent from neighbouring Neurons all of which are received at around the same time. Apart from the difference in originator, each deliberator is cogitating the same topic and hence each deliberator is constructed to have a identical parameters such as activation and bias because all deliberators are instructed by the same feature.
Accumulation of Connection weights for deliberators on this same topic may cross a threshold which when reached will cause the deliberator on which the threshold was crossed to send deliberations through synapses to its neighbours. Some neighbours may share synapses and may receive the deliberation immediately and others may only receive the deliberation through a neighbour who has faithfully passed the deliberation on.
This is all done by assigning weights to the Connections instead of to the synapses. I think that is a big deal, and an advancement in neuroscience that was published in this blog entry before anyone else in this world knew that assigning weights to representations of nodes in a Neuron's knowledge of its neural neighbourhood would be better than assigning them merely to the physical connections it would share with its neighbours.
This mechanism, and more as would answer most other questions about my concept is all there in my code. Really, it is... almost. I'm sitting on the greatest advancement in neuroscience since the perceptron!
Subscribe to:
Posts (Atom)