03-28-2025, 08:49 AM
Hello. 
We all know that a neural network is not just the part that works after it has been trained.
The neural network also contains a learning "mechanism".
The advantage of SB-neurons is that they have a system of precise and unambiguous recognition of EVERY situation that arises at the inputs of the SB-neuron. The weighting system of classical neurons allows for the possibility of the SAME weighted sum occurring for several different combinations of values at the neuron's inputs.
This is bad.
The activation function of the SB-neuron contains training information. The activation function of a classical neuron is a "line on a graph" that has nothing to do with the training data.
This fact absolutely kills me.
If a neural network needs to learn to identify ANY cross (cat) and ANY zero (dog) anywhere in the pixel matrix, then a SB-neural network will do this job better than a classical neural network, because a SB-neural network is better than a classical one.
The training program of the SB-neural network will better determine the features of "cross" and "zero". And it will more accurately store information about these features in the activation function of SB-neurons.
I'm ready to prove this on a real task.
We all know that a neural network is not just the part that works after it has been trained.
The neural network also contains a learning "mechanism".
The advantage of SB-neurons is that they have a system of precise and unambiguous recognition of EVERY situation that arises at the inputs of the SB-neuron. The weighting system of classical neurons allows for the possibility of the SAME weighted sum occurring for several different combinations of values at the neuron's inputs.
This is bad.
The activation function of the SB-neuron contains training information. The activation function of a classical neuron is a "line on a graph" that has nothing to do with the training data.
This fact absolutely kills me.
If a neural network needs to learn to identify ANY cross (cat) and ANY zero (dog) anywhere in the pixel matrix, then a SB-neural network will do this job better than a classical neural network, because a SB-neural network is better than a classical one.
The training program of the SB-neural network will better determine the features of "cross" and "zero". And it will more accurately store information about these features in the activation function of SB-neurons.
I'm ready to prove this on a real task.
![[-]](https://litdev.uk/mybb/images/collapse.png)