Kevin D. Shabahang

Supervisors: Prof Simon Dennis

"I am interested in operations and representations that underpin thought.  We need to know the data-structures that encode traces of events, and the operations that act on those data-structures in order to drive behaviour.  The two problems must be solved simultaneously, because the choice of one will necessarily constrain the choice of the other.

For my PhD, I will narrow my focus to the realm of natural language processing, working on Dr. Simon Dennis's Syntagmatic-Paradigmatic (SP) model.  The SP model exploits two forms of structure in its input.  Syntagmatic associations encode what words co-occur (e.g., coffee and cup) and paradigmatic associations encode what words are interchangeable (e.g., cup and mug).

Although the original SP model captured a multitude of behavioural benchmarks, both in memory and language processing, it was limited in its capacity to scale to real-world input (e.g., mass amounts of text).  Together, Dr. Dennis and I are working on a new, more scalable, version of the SP model.  Aided with insights from recent neurophysiological findings related to Short-Term Synaptic Plasticity (STDP), and a general understanding of the highly recurrent architecture of neural modules (e.g., CA3 of the hippocampus), we adopt a connectionist framework for our modelling.

Although connectionism has seen a re-birth in recent years, our approach differs from current neural-network architectures in its parsimony and dedication to psychological plausibility.  The neural-network architectures in use today are excellent pieces of technology, but are opaque and lack parsimony.  They are opaque in that they do not offer a clear explanation of the way they solve a given problem.  They lack parsimony in that they require hundreds, maybe thousands, of free-parameters for training.  A scientific theory must be both explanatory and parsimonious.

Aside from research, I have a passion for the outdoors, playing guitar, and practicing Brazilian Jiu-Jitsu."