Artificial neural network (ANN) is a simulation of biological neural network with some mathematical functions. However, it might be better if ANN be simulated closer to the real biological structure and mechanism.
Current ANN designs come with input nodes, middle layers and output layers. Each layer is connected to previous layer and next layer. This is multi-layer ANN. A good suggestion should be using the real meaning of the word network and create a graph-based ANN instead of layer-based, eg. recurrent neural network is kinda graph than layered. Graph-based ANN allows circulation in the network during training, not only feed-forwarding.
Biological neuron soma is supposed to be much more complex than the mathematical artificial neuron with only regression function and activation function. Should an artificial neuron soma come with a full complex programme instead of just only 2 functions: regression and activation?
Biological neuron comes with multiple axon terminals which carry similar outputs instead of exactly the same output as current ANN designs.
Some types of neural networks:
- Feedforward dense multi-layer neural network (basic and simple neural network with feedforwarding and backpropagation)
- Convolutional neural network (CNN, simulate visual cortex) for image recognition
- Recurrent neural network (RNN), time delay neural network (TDNN) for speed recognition
- Radial basis function network (2 layer neural network)
- Evolutionary neural network (the network which changes neuron connections by itself)
- Modular neural network (just like the brain, each different part of the brain is a neural network by itself)
- Biological simulated neural network (the neural network with features simulating closely to biological neurons and parts of brain)
More details about types of neural networks: