For years, the best AI systems have been a result of a technique called "deep learning."
What is deep learning? Well, it's actually a relatively new term to describe an artificial intelligence approach called neural networks. A basic definition would be a computer system modeled on the human brain and nervous system. Which is exactly what it is.
Neural networks are a means for a machine to learn. The networks loosely mimic the human brain and enable a computer to learn to perform a task by analyzing examples, that have already been labeled by a human. An object recognition system could be given thousands of labeled images and the computer would learn to find patterns in the images that are consistent with specific labels.
A neural network is comprised of thousands of processing nodes that are meant to be artificial neurons. These "neurons" are usually grouped into layers and are connected to nodes in the layer below them and the layer on top. These connections are simplified versions of synapses in the brain. An individual node may be linked to several nodes below it, from which it receives data, and several nodes above it, to which it sends data.
When receiving data, the node will assign it a number called a “weight.” The weight increases or decreases the strength of the signal at a connection. The weight of an input is a number which when multiplied with the input gives the weighted input. These weighted inputs are then added together and if they exceed a pre-set threshold value, the node sends information to the next layer. In any other case, the neuron does not.
When a neural net is being trained, all of its weights and thresholds are initially set to random values. Data is fed to the input layer and passes through the next layers, getting multiplied and added together in complex ways, until it arrives, completely transformed, at the output layer. The weights and thresholds are continually adjusted until training data with the same labels consistently yield similar outputs.
The current deep-learning revolution is the result of the need for hardware that can keep up with the complex imagery and rapid pace of today’s video games. The solution? A graphics processing unit (GPU), which packs thousands of relatively simple processing cores on a single chip, something that was actually very similar to a neural net. Modern GPUs enabled the one or two layer networks to become the 10-, 15-, even 50-layer networks of today, putting the "deep" in "deep-learning.
As deep-learning develops, so will our understanding of neural networks, aiding advancement in the best-performing artificial intelligence systems in the years to come.
References:
Hardesty , Larry. “Explained: Neural Networks.” MIT News | Massachusetts Institute of Technology, 14 Apr. 2017, news.mit.edu/2017/explained-neural-networks-deep-learning-0414.
Written by: Mahathi Somula
留言