smiron / cnnnet Goto Github PK

Instead on adding guidance force values, we should construct a force vector for each guidance force and average them together.

There are cases when a neuron cannot guide it's axon because all the non 0 (default) values are negative. In this case the max value is 0 and the axon cannot be guided.

This makes sense .. but a solution must be found.
Essentially a blank position is more favorable than any other position.

Refactor Soma guidance force to use the same infrastructure as Axons guidance.

Remove the NeuronCompute and NeuronInput classes

When all other guidance forces fail .. this guidance force should unblock the axon guidance.

These methods must be merged so that NeuronCompute has the same axon extending functionality as NeuronInput

Implement some kind of event that will trigger whenever the network is processed and attach to it all the available network state.

In the guidance process we take into account all guidance forces.
But in the method Neuron.ProcessGuideAxon the final IF statement only checks the Network.NeuronUndesirabilityMap to see if the new location is better

Implement output neurons.
This type of neuron should only have a soma visible in the network.

Two possible implementations:

1. If should have a desired state (active or not active). When the desired activity is not respected it should behave like a normal neuron (synapse plasticity etc). This would make the network learn an output for a given input.
2. A more experimental solution. In this implementation the output neuron doesn't have the desired state documented above. The output neurons would serve as a feedback to the input neurons.

I will implement the second solution.

We must have synapse formation and maintainability cost.

Create a synapse class similar to a neuron:

• Should make connection to assigned pre synaptic neuron only if it helps the post synaptic neuron
• only the synapses that help the post synaptic neuron should survive (should also help with computational optimization)

If the neuron doesn't fire often enough we need to do something.
So the neuron will increase the strength of every synapse that was just active (in the previous iteration).
TODO: this needs more work so we have a parameter or something similar that will determine if a neuron fires often enough or not.

Mechanism so that if the sensor neuron is triggered continuously .. we should decrease the firing rate until we reach a base line.
Also .. if a sensor isn't triggered at all we should have a base line of firing rate.

Score map is initialized to 0. We should initialize with a minimum value. This way the default value isn't bigger than some other negative value.