Abstract

When we listen to a lecture or read a book, each new word is interpreted in the temporal context of previous words. This is accomplished in a large language model with a long input vector, which keeps a running record of all previous input and output words. Temporal context is also needed for long-term working memory, which draws from long-term memory to interpret the input stream, form associations, and make decisions. These cognitive functions will be explored in cortical models of recurrent neural networks that support persistent travelling waves.