|
Artificial neurons and neural networks share a number of similarities with their biological counterparts, though there are a number of abstractions as well both at the level of a single neuron and the network as a whole.
One of the similarities is that both the artificial and biological neurons work with inputs and outputs that connect them to their peers, either exciting or inhibiting them. This allows the sending and receiving of signals between them. Another similarity is that both work with thresholds, which can be exceeded depending on the inputs a neuron receives from it peers. Once the threshold is exceeded the neuron, both biological and artificial, fire an output.
Considering the differences, one is how the actual inputs and outputs work. According to Franklin, in real nervous systems, “the significance of the signal most frequently is carried by the rate of firing, not the magnitude of a single firing”, which describes that difference clearly. Another difference is that a single artificial neuron can excite or inhibit another through a positively weighted synapse or a negatively weighted one respectively; however, Franklin explains that a single biological neuron is “either inhibitory or excitatory to all subsequent neurons to which they connect”.
One of the similarities between artificial and biological networks is some of the ways of learning or training styles. Some of training styles shown by Franklin on chapter 7 can also be considered when analysing a biological neural network. For example, hard-wired or programmed networks in the brain are defined by the genetic code, these may control basic functions; another example is reinforcement, which can be paralleled to training a dog. Another similarity is that if a part of some biological neural networks (not the stem brain) is damaged (some neurons die) the performance degrades gracefully, which is what happens in an artificial neural network when some units fail.
One important difference is that biological networks work with different kinds of chemicals (neurotransmitters) and this changes how the network works. This is not currently modelled in artificial neural networks. Another difference is that biological do not implement backpropagation learning as some artificial networks do.
Source:
Franklin, S., Artificial Minds. 2001.
|
