Neural networks change the future of computing
Jacob Herbert
An artificial neural network is a type of machine-learning algorithmwhich vaguely copies the structure of the human brain. It uses layers of nodes which are interconnected by weighted neurons. Numbers are input which could represent images, statistics or any type of data. The number are then processed through the neural network, which outputs another series of numbers which might correspond to the direction a vehicle should travel, or the type of object seen in an image. To set the weights on a neural network or refine the specific structure of a neural network, we must train it using a large training data set and then test it on another data set to determine its accuracy. There are many ways to train a neural network depending on how it will be used.
Neural networks were originally theorized in the mid-1940s by mathematicians as a way machines could be artificially intelligent. The idea was largely forgotten about until the 1980s when breakthroughs in computing power and a new interest in machine-learning prompted the idea to be revisited. Today, the neural network underpins almost all
modern machine-learning systems. Although the neural network is an old idea, with the exponential growth in the computing power available and new types of neural networks being invented, the neural network will become more and more integral to our computer systems.
There exist many types of neural networks, all based on the underlying architecture of nodes and neurons. For instance, in recent years we have seen the increasing development of convolutional neural networks, which are used to recognize images. There exists an annual image recognition competition known as ImageNet, where competitors must design the best programme to classify and detect objects and scenes in images from the ImageNet database. This is a database of over 14 million images of objects from 20,000 different categories such as animals, flowers, etc. The current reigning champion is Microsoft’s convolutional neural network – Microsoft ResNet - which has 152 layers of nodes. It can correctly identify 96.43% (Microsoft Research, 2015) of the objects in given images from the database. That level of accuracy is both impressive and concerning, given the implications of a computer which can understand its surroundings almost as well as a human. The future of image recognition software now lies with CNN-based image recognition algorithms instead of more traditional algorithms that don’t use neural networks.
192
Made with FlippingBook Digital Publishing Software