Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="EzekielRaiden" data-source="post: 9086555" data-attributes="member: 6790260">Yeah, I would expect a neural network involved unless explicitly told otherwise. Otherwise, there's no "training" involved.Edit: Oh, if anyone thought this meant literal neurons, as in something actually functioning the way human brain cells function, then no, there is no such thing.[SPOILER="Explanatory digression"]A "neural network" in AI parlance is, as a general rule, a set of "layers" of "nodes." E.g. you could have 10 layers, each with 6 nodes in them. These nodes are the analogic "neurons' of the neural network. Generally, layers only take data from the layer before them (except the first layer, which naturally must take data from human input.) They get assigned assigned random instructions (weights) for how to pick up data from the layer which came before. So, for example, with a very small neural network of 3 layers with 4 nodes, you'd have...Layer 1: A1, B1, C1, D1Layer 2: A2, B2, C2, D2Layer 3: A3, B3, C3, D3The first layer receives input data, so its "training" is about what information it should factor into its calculation to determine what value it outputs. E.g., node A1 might pick out certain pixels of an image, while B1 would pick out a different set of pixels. A given node takes in all of its weighted input data, and then returns a number, which can be used by other nodes down the line. The sum total of the neural network is thus a HUGE list of parameters which all feed one layer into another until you reach the final layer, at which point its output is parsed for human use.Relatively simple applications of neural networks, such as optical character recognition, it is often possible to get at least a loose idea of what the network is doing because you don't need very many layers or nodes to do that task. Often, nodes will develop weights which (for example) identify the presence of curved structures in specific areas, approximating certain kinds of convolutions (which are how regular software performs edge-detection) in a much more application-specific way.[/SPOILER]I don't think it's possible, at least at this time, to create software that can "enhance" artistic work in this way without using this layers-of-nodes structure. And, as said above, the common parlance for this "layers-of-nodes" structure, in computer science, is "a neural network."</blockquote>

Verification