Running head : HEBBIAN LEARNING How Far Can You Go with Hebbian Learning , and When Does it Lead you Astray ?

This paper considers the use of Hebbian learning rules to model aspects of development and learning, including the emergence of structure in the visual system in early life. There is considerable physiological evidence that a Hebb-like learning rule applies to the strengthening of synaptic efficacy seen in neurophysiological investigations of Long-Term Potentiation, and similar learning rules are often used to show how various properties of visual neurons and their organization into ocular dominance stripes and orientation columns could arise without being otherwise pre-programmed. Some of the plusses and minuses of Hebbian learning are considered. Hebbian learning can strengthen the neural response that is elicited by an input; this can be good if the response made is appropriate to the situation but it can also be counterproductive if a different response would be more appropriate. Cases in which this outcome-independent Hebbian type of strengthening might account at least in part for cases in which humans fail to learn are considered, and computational models embodying the Hebbian approach are described that can account for the findings. At a systems level, Hebbian learning cannot be the whole story. From a computational point of view, Hebbian learning can certainly lead one in the wrong direction, and some form of control over this is necessary. Also, experimental findings clearly show that human learning can be affected by feedback. Several ways in which sensitivity to feedback might be incorporated to guide learning within a fundamentally Hebbian framework for learning are considered.