|
|
|
|
|
|
 |
Neural Networks ComponentsLearning Algorithms in Neural NetworksOnce an artificial neural network has been structured for a particular application, it is ready to be trained. This training allows the network to learn appropriate behaviour for the task at hand. As with our discussion of artificial neurones and neural networks, it is useful to consider machine learning in the context of neurobiology. Unfortunately, our scientific understanding of learning and memory within the brain remain limited. However, several lines of research on learning and memory in animals suggest that it is the modification of the connections between neurones (at least in part) that allow animals to modify their behaviour with experience. Many neural networks also learn through the modification of their synapses with experience. Broadly speaking, there are two approaches to training neural networks depending on how synapses are modified with experience: supervised and unsupervised. Supervised learning involves a "teacher" providing the network with the desired output with the input. For example, when learning to recognise different musical instruments a teacher would reveal the instrument playing so that the neural network can compare its response to the correct one. Although often used in artificial neural networks, the learning rules used in supervised learning are generally considered to be unrealistic and unlikely to occur in biological brains. Unsupervised training is where the network has to process data without outside help. Instead the networks task is to re-represent the inputs in a more efficient way. Unsupervised learning plays a role in perception such as the visual system and also in motor structures like the cerebellum. Although unsupervised self-learning networks are closer in function to animal brains they researchers have had difficulty implementing them to solve real-world problems. A possible third class of learning mechanism termed reinforcement learning lies close to supervised learning but differs in that it is biologically feasible. For correct responses, reinforcement learning resembles unsupervised learning: in both cases the network is told that its actions were correct. However, the different types of learning differ when mistakes are made. In these situations, supervised learning lets the network know exactly what it should have done (the "desired" output) whereas reinforcement learning only says that the behaviour was inappropriate and, perhaps, how inappropriate it was. So, from the above we can distinguish between at least two, and possibly three, major classes of learning within networks. Index | Components | Learning Algorithms © 2008 Marcus bros
|
|
|
|
