| Humans typically employ multiple learning strategies when acquiring a skill. In particular, human learners are capable of learning both from feedback on their performance at a task and also from direct verbal instruction, provided by a knowledgeable teacher. These two sources of knowledge, induction over experiences and the reception of instructions, can interact to solve complex learning problems.;This research presents a general computational model of such instructed learning--of learning involving the integration of induction and instruction. This model is wholly connectionist in nature, being composed of a modular artificial neural network. Thus, the model extends standard connectionist learning beyond induction to incorporate explicit advice taking. A central feature of this model is the representation of explicit instructed knowledge in the dynamic activation space of a recurrent attractor network, while inductively acquired knowledge is encoded in the usual way, in the weight space of network connections. This allows for the rapid modification of behavior in response to instructions while still maintaining the process of slow learning from experience. Also, in the context of examining this model, several insights into the limitations of connectionist attractor networks arise and are investigated in detail.;The model is validated against human experimental data involving an established interference effect in instructed category learning. When provided with an explicit rule for categorizing visual stimulus items, followed by practice at applying the rule on a particular collection of labeled examples, human learners will sometimes come to violate the given rule as a result of the practice session, even if the practice items are labeled in a manner perfectly consistent with the given rule. For some stimuli, the similarity of the item to previously viewed practice items comes to dominate over instruction following, causing a decrease in rule following accuracy. The proposed model is shown to exhibit this interference effect, and results from new instructed category learning experiments with human learners are reported. Lastly, a Bayesian analysis is conducted of this interference effect, examining the degree to which this effect may be considered a natural result of optimal learning behavior. |
