Current interests · Relationship between cognitive load, attention and learning · Brain imaging · Linking neurophysiology to behaviour · J Senior Researcher (1999-) National Institute of Advanced Industrial Science and Technology (AIST), Neuroscience Research Institute, Cognitive and Behavioral Sciences Group. Researcher (1997-1999) Electrotechnical laboratory, Information Science Division, Cognitive Development Group. STA Fellow (1995-1997) Electrotechnical laboratory, Information Science Division, Cognitive Science Section. relational theory of cognition and development; measuring relational complexity in cognitive tasks; systematicity and its implication for classical and connectionist cognitive models; Research Associate (1994-95) After submitting my thesis I worked with Professor Graeme Halford in the Department of Psychology at The University of Queensland. There, I worked on: STAR (Structure Tensor Model of Analogy and Relations); Relational versus associative modes of cognitive processing; and a Relational theory of cognition and development. Doctorate (1991-94) Title: Connectionism and the Problem of Systematicity Abstract Description: Systematicity is the property whereby human cognitive capacity is organized around structural similarity. For example, the capacity to understand the concept "John loves Mary" extends to the structurally related concept "Mary loves John". My thesis is concerned with how well connectionist models support systemacity. In particular, I consider systematicity as generalization to novel position. For example, can a network correctly compute complex objects of the form (Subject loves Object) where "Mary" appears in the Subject position having only every been trained on examples with "Mary" in the Object position. I found that the feedforward and simple recurrent networks could not demonstrate generalization across position assuming no similarity between atomic objects (e.g., John, Mary) representations. This is because of the independence between weights which encode and decode component representations across different positions. However, a third architecture called the Tensor-recurrent network (proposed in this thesis) does support generalization across position under the same assumptions. Weight dependence is ensured by exploiting role-filler (position-component) method of representing structured objects of the tensor network. This means that internal component representations are constructed independently of their position within a complex object. In addition, though, appropriate role (position) and filler (component) representations are learned by exploiting the learning ability of feedforward and recurrent networks. That is by backpropagation update information though the tensor network and weights which generate internal role and filler representations. However, the Tensor-recurrent network was only designed for flat structures. Further work is needed to handle recursively structured objects. Department: Computer Science, The University of Queensland Supervisor: Dr Janet Wiles Accepted: February, 1995 University studies* Doctor of Philosophy, PhD (Department of Computer Science) Bachelor of Arts with 1st class honours, BA[hon] (Department of Computer Science) Bachelor of Science, BSc (Departments of Mathematics and Physics) Postgraduate Diploma in Education, PGDip. Ed. (Department of Education)