While ML may operate on a variety of architectures, the cutting edge of AI in the 2010s has largely run on artificial neural networks (ANNs) – computer programs that are inspired by, albeit quite different from, the human brain. ANNs roughly mimic the electrical operations of the brain’s neuronal connections rather than emulate high-level logic like GOFAI does. ANNs are ‘based on the assumption that cognition emerges through the interactions of a large number of simple processing elements or units (i.e., ‘neurons’)’ (Sun 2014: 109). Artificial neurons are organized into a series of layers and each layer is connected to the layers above and below. The lowest level receives inputs – e.g. images, text or speech – in the form of data that has been vectorised (converted into long strings of numbers). Higher levels – called hidden layers – process data that is sent up from the layers below them. Early networks had only one hidden layer, but today’s networks have many more. In general, the more layers the underlying ANN architecture of the ML system has, the more complex patterns it can find. The most advanced ML in the 2010s is called ‘deep learning’ because these networks are many layers deep (LeCun, Bengio and Hinton 2015: 436–44), with some networks possessing as many as 1,000 (He et al. 2016). While previous ML networks were constructed by hand, a ‘key aspect of deep learning is that these layers of features are not designed by human engineers: they are learned from data using a general-purpose learning procedure’ (LeCun, Bengio and Hinton 2015: 436).
The artificial synapses which connect the layers of artificial neurons are ‘weighted’ with numeric values representing the strength of the connection. The network ‘learns’ through adjusting the weights of these connections. We can thus think of an ML system as a fixed template with changeable parameters; ‘by assigning different values to these parameters, the program can do different things’ (Alpaydin 2016: 24). ANNs are exposed to a data set, which might be images of faces or audio clips of people saying hello. In a process called ‘training’, the network is exposed to many instances of the chosen object(s) and the weights of the synapses are adjusted by a learning algorithm until the network learns to output the correct response, recognizing faces or the word hello, as the case may be (Kaplan 2016: 30).
introduction: ai-capital
13