Not the biological ones, though I’m sure there is a correlation.
In: Engineering
If you think of what you want your model to do as a desired function F that takes an input x and gives the correct output y, you can think of a neural network as a function R that tries to be as similar to F as it can by trying to learn to “mimic” it with its many layers of neurons. The idea here being that instead of explicitly defining this function R, you define its architecture, i.e., number of neurons and layers and the network learns the weights and biases through basically trial and error (a bit more complex than this).
A “neuron” takes all of the outputs of the previous layer (which could be your input), weighs them with the learned weights, sums them and adds a learned bias, and applies an “activation function”, you can think of a neuron as learning whats called a “non linear function”. As you could maybe see though, the weights and biases are used to make a linear function w1x1 + w2x2 + … wnxn + b (see the similarities to mx + b?), and this is where the activation function comes in handy because when it is applied to this linear function it makes it non-linear (by for example making its value 0 if it is negative). This is central to the idea of Machine Learning with Neural Networks as theyre based on the theory that we can learn any function F by stacking a lot of nonlinear functions on top of each other, and this is exactly what a neural network does.
Sorry for horrible formatting, im on a phone. Feel free to ask any clarifying questions, if ik the answer ill def try to answer
I wouldn’t attempt an explanation without a diagram. It’s a graph and it is all about the way you connect the nodes.
Simply said, the AI ones are an attempt to replicate the biological ones in data structures. At the beginning, there was the idea that if computers replicate brains as good as possible, they would also become “conscious” at some point. Hence the name “artificial intelligence”.
Of course, modern neural networks/AIs are extremely successful in replicating specific tasks that our brains can perform, but are still as far removed from consciousness as the automata of the 19th century (we don’t even understand how consciousness works in human brains yet).
There’s some detailed explanations so I’ll give something simple:
It is just a network of individual nodes, communicating from one level to the next. Each node gets a set of inputs, each with a weight, then gives an output according to those weights. Each node is basically just a small math function, feeding the next small math function(s).
Structured correctly, they can transform something like an image, into a 1 or 0 for “is this a dog?”.
Imagine you know the input and the output, but don’t know what kind of mathematical function is between them. So you try to guess what it is by adjusting the coefficients until the input and output line up for the ones you know, and then give it input you don’t know the right result for.
This is what neural networks are – complicated functions with thousands(or maybe millions) of input and output parameters, and coefficients guessed in the training process.
Starting with a biological neuron to explain why “neural networks” are named that: A biological neuron is a cell that receives an input, namely by having molecules known as neurotransmitters bind to their receptors. When they receive a strong enough input (i.e. having enough neurotransmitters bind to their receptors), they fire off an electrical signal down the length of their axon (a very long part of the neuron’s body), eventually releasing more neurotransmitters at the very end.
In computing, a neural network shares some of these features: it has nodes (which act like the network’s neuron) which receive an input (as a number, instead of as neurotransmitters binding to their receptors) and then pass a signal (again, as a number) onto other nodes in the network. Instead of axons carrying a signal from the central body of the neuron over to another neuron, you simply have connections between nodes.
Both kinds of neural network work in the same, very general way: you have tiny, discrete units of processing which listen for signals, and, based on the signal they receive, send a signal further down the network, either to activate or inhibit further neurons/nodes. Each node/neuron can be thought of as “looking” for something from the input to help the network as a whole determine what the output should be.
There is a correlation as they are bioinspired. The neurons in your brain are connected to other neurons through synapses, and neurons in your brain tend to fire when they receive sufficient excitatory input from other nearby neurons. That is, if you have a neuron connected to a few other neurons via synapses, and those few neurons fire, the connected neuron may also fire depending on the relative strength of the synaptic connections and whether the combined input exceeds the neuron’s firing threshold. The act of that neuron firing will then contribute to the input of some other neurons, which influences other neurons, and so on. This complex web of synaptic connections forms a network of neurons, called a neural network.
Contemporary Artificial Neural Networks are modeled on biology but are dramatically simpler in structure and connectivity than biological networks. There is no one kind of ANN. They have different structures depending on what kind of network they are, with the simplest in my opinion being static layers of fully connected neurons which are usually visualized as chains of neurons lined up next to each other, with all neurons in a chain connected to every neuron in the next chain. By analogy, neurons in the early layers contribute to the activation of the neurons in the next layers based on the strength of their synaptic connections. This is done through simple multiply-accumulate-add operations that are extremely well-studied in computer hardware. So we can create an artificial and emulated version of a biological neural network (be it one found in a human brain, mouse brain, or insect brain), but with a much simpler functional mechanism and connectivity suitable for computer hardware, that can still achieve some strong artificial intelligence capabilities.
Again it doesn’t have to have one particular structure. There are also neural networks that are more closely informed by biology, like Spiking Neural Networks. There are neural gases, that are not static and dynamically change the number of neurons. But nearly anything that is based on the concept of neurons connected to each other through some kind of web falls under the umbrella of neural networks, and usually, they can “learn” by re-organizing or evolving their state in response to stimulus.
Latest Answers