1.32 Neural Computation Theories of Learning

This postulate was experimentally confirmed in the hippocampus, where high-frequency stimulation (HFS) of a presynaptic neuron causes long-term potentiation (LTP) in the synapses connecting it to the postsynaptic neurons (Bliss and Lomo, 1973). LTP takes place only if the postsynaptic cell is also active and sufficiently depolarized (Kelso et al., 1986). In many brain areas, this is due to the N-methyl-D-aspartate (NMDA) type of glutamate receptor, which opens when glutamate is bound to the receptor and the postsynaptic cell is sufficiently depolarized at the same time. Hebb’s rule has served as the starting point for studying the learning capabilities of artificial neural networks (ANN) and for the theoretical analysis and computational modeling of biological neural systems (Hertz et al., 1991). The architecture of an ANN determines its behavior and learning capabilities. The architecture of a network is defined by the connections among the artificial neural units and the function that each unit performs on its inputs. Two general classes of network models have feedforward and recurrent architectures. The simplest feedforward network has one layer of input units and one layer of output units (Fig. 1, left). All connections are unidirectional and project from the input units to the output units. Perceptron is an example of such a simple feedforward network (Rosenblatt, 1958). It can learn to classify patterns from examples. It turned out that perceptron can only classify patterns that are linearly separable – that is, if the positive patterns can be separated from all negative patterns by a plane in the space of input patterns. More powerful multilayer feedforward networks can discriminate patterns that are not linearly separable. In a multilayer feedforward network, the “hidden” layers of units between the input and output layers allow more flexibility in learning features (Rumelhart et al., 1986). Multilayer feedforward networks can solve some difficult problems (Rumelhart and McClelland, 1986) and underlie current rapid development of the field of deep learning in machine learning (LeCun et al., 2015). In contrast to strictly feedforward network models, recurrent networks also have feedback connections among units in the network (Fig. 1, right). A simple recurrent network can have a uniform architecture such as all-to-all connectivity combined with symmetrical weights between units as in Hopfield network (Hopfield, 1982), or it can be a network with specific connections designed to model a particular biological system (Sporns, 2010).