A Neural Mass Model of Selective Attention over Semantic Memory

Semantic memory and selective attention are usually studied apart, even though they share the same oscillatory machinery in the brain. This thesis brings them together in one neural mass model, extending a trained semantic-memory network with an alpha-gated attention mechanism, so stored concepts can be selected and retrieved one at a time. The model rests on two components: a semantic-memory network trained with an asymmetric Hebbian rule on 102 features across fourteen animals, and an alpha-rhythm gating mechanism that suppresses information through phase opposition between regions. These sit inside a three-region architecture: a sensory area for feature representation, an alpha oscillator, and a semantic recall area holding the auto-associative synapses and the fatigue mechanism. Three mechanisms are added: an inhibition mechanism, applying alpha suppression to the sensory area; a spotlight mechanism, releasing that suppression selectively to move attention between concepts; and a fatigue mechanism, modelled as activity-dependent depletion of excitation, letting an active concept fade so the network can move on. We simulated three animals presented in sequence, each cued by a single feature. The spotlight correctly segments them in most cases, retrieving each within its own 400 ms window with all features recovered in the recall area. Only in a few complex cases, where an animal shared many features with others, did retrieval need a small excitation increase. In one such case this pushed a third, less overlapping animal into saturation, only partly corrected by more inhibition. A further case showed retrieval can also depend on an animal's position in the sequence, not its own features, pointing to a timing effect alongside shared content. In conclusion, although alpha gating was built for suppressing distractors, our simulations show it can also select and retrieve stored knowledge, with fatigue alone enough to keep retrieval moving without external control.