City Research Online

Abstract

In this thesis, the Double Error model, a general computational model of real-time learning is presented. It builds upon previous real-time error-correction models and assumes that associative connections form not only between stimuli and reinforcers, but between all types of stimuli in a connectionist network. The stimulus representation uses temporally-distributed elements with memory traces, and a process of expectation-based attentional modulation for both reinforcers and non-reinforcing stimuli is introduced. A modified error-correction learning rule is proposed, which incorporates both an error-term for the predicted and predicting stimulus. The static asymptote of learning familiar from other models of learning is replaced by a similarity measure between the activities of said stimuli, resulting in more temporally correlated stimulus representations forming stronger associative links. Associative retrieval based on previously formed associative links result in the model predicting mediated learning and pre-exposure effects. As a general model of learning, it accounts for phenomena predicted by extant learning models. For instance, its usage of error-correction learning produces a natural account of cue-competition effects such as blocking and overshadowing. Its elemental framework, which incorporates overlapping sets of elements to represent stimuli, leads to it predicting non-linear discriminations including biconditional discriminations and negative patterning. The observation that adding a cue to an excitatory compound stimulus leads to a lower generalization decrement as compared to removing a cue from said compound also follows from this representational assumption. The model further makes a number of unique predictions. The apparent contradiction of mediated learning in backward blocking and mediated conditioning proceeding in opposite directions is predicted through the model’s dynamic asymptote. Latent inhibition is accounted for as occurring through both learning and selective attention. The selective attention of the model likewise produces emergent effects when instantiated in the real-time dynamics of the model, predicting that the relatively best predictor of an outcome can sustain the largest amount of attention when compared to poorer predictors of said outcome. The model is evaluated theoretically, through simulations of learning experiments, and mathematically to demonstrate its generality and formal validity. Further, a simplified version of the model is contrasted against other models on a simple artificial classification task, showcasing the power of the fully-connected nature of the model, as well as its second error term in enabling the model’s performance as a classifier. Finally, numerous avenues of future work have been explored. I have completed a proof-of-concept deep recurrent network extension of the model, instantiated with reference to machine learning theory, and applied the second error term of the model to modulating backpropagation in time of a vanilla RNN. Both the former and latter were applied to a natural language processing task.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	School of Science & Technology Doctoral Theses School of Science & Technology > School of Science & Technology Doctoral Theses

Export

Downloads