Publications
- Exploring Symmetry and Representation in Vector Symbolic Architectures
- 2024
- Ozen, R.
- Master's Thesis
- Abstract: In this thesis, I explore and extend the methods of Vector Symbolic Architectures (VSAs), with a particular focus on Holographic Reduced Representations (HRRs) and group VSAs. Through a consideration of abstract groups and linear representation theory, I provide detailed analyses of these systems to clarify their underlying algebraic structures, and suggest a unifying framework that enables the construction of new HRR-like VSA systems. In addition I present novel VSA models, including a memory retrieval model based on Hopfield networks for use with unitary HRRs and other group VSA generalizations, and a resonator network model that solves for rotations in 3D space. I also introduce unitary group VSAs, and show that such VSAs satisfy the notion of a maximally expressive VSA system under certain formal assumptions, enabling the representation of a diverse set of structured data relationships.
- Read full publication >
- Minerva-Q: A Multiple-Trace Memory System for Reinforcement Learning
- 2022
- Ozen, R., West, R. L., & Kelly, M. A.
- Proceedings of the 20th International Conference on Cognitive Modelling
- Abstract: We propose Minerva-Q, a multiple-trace memory model capable of perceptual-motor reinforcement learning. This model combines Q-learning with the Minerva family of memory models. In our simulations we found our Minerva-Q agent learned increasingly optimal solutions to the Cart Pole task and reproduced human-like performance when presented with minimal expert training examples in a sparse reward task.
- Read full publication >
- Workshop Session
- 2022
- Maloney A., Schellinck J., Ozen R., Conway-Smith B., and West, R.L.
- 29th Annual ACT-R Workshop
- Abstract: Discussions and presentation titled "gactar: A tool for creating & running basic ACT-R models on multiple implementations using a single declarative file format."
- Read full publication >