Experiment: Neural network modeling of semantic and associative networks

Awhile back, I posted a neural network application I developed that uses back propagation to support machine learning -- basically a bare bones network capable of learning to map any given input pattern to a specific output pattern. One of the things that interests me about neural network modelling, is that they provide a neurologically plausible means of modelling cognitive processes such as language production. For example, a few years back, I developed a different neural network to simulate the types of errors commonly observed in aphasics or patients undergoing cortical stimulation mapping (CSM) - a procedure designed to identify "eloquent" language cortex so that these critical regions can be preserved during neurosurgery for epilepsy or tumor resection. During CSM, an awake patient is given a confrontation naming task (i.e. shown pictures of common objects and asked to name them), and during certain trials, the neurosurgeon will apply current directly to the cortex. Sometimes, if the stimulation will result in a naming error such as a semantic paraphasia (CAT-> "dog") or phonological error (DOG -> "log"). 


The application attempted to simulate these error types in order test one plausible model of underlying semantic/conceptual organization. Specifically, I defined concept nodes (e.g. DOG) that mapped to a layer of feature nodes (e.g., furry, four-legged, has tail, etc.) that were based on a set of feature norms developed by McRae and colleagues. The feature layer of nodes was then linked to an output layer of phonological form, in order to model the process of seeing an object, activating its semantic/conceptual features, and then producing the object's name. Because objects can share common features (e.g., DOG & CAT share "furry"), this introduces the possibility for naming to go awry if the activation state of a particular node breaches its threshold level. In order to induce these types of errors, I would then "lesion" the model (analogous to an aphasic lesion or the temporary lesions in CSM) by introducing some amount of noise in the model. The basic findings were that this model of semantic/conceptual organization did quite well in capturing the sorts of errors as observed in real CSM data. I'll have to dig through some of my old hard drives, but hopefully I can post the source code and application up here at some point.


Anyhow,yesterday I received a request to help an MA student acquire some data for his master's thesis entitled “Neural Network modeling of semantic and associative networks.” I'm not sure what exactly his study is about, but he is looking for volunteers to spend about 5 minutes doing an online chained word association task. If you are interested, you can find the experiment here:

http://itp.uni-frankfurt.de/~mehran/WordAssociations/index.php?mode=ChainStart