Comments about abstract and figures

Hi -

I’ve spent a great deal of time doing very basic simulations in SPA to understand the effects of feedback on the speed of processing. I am attaching a draft of an abstract based on that work. I would like to make the figures, code, etc available, but I am unclear where (?github) to place it. Thanks

Howard

Working Memory, Feedback Connections, and the Speed of Human Cognition:

Computational models suggest a mechanism for cognitive slowing with aging and disease

Howard Crystal, Terrence Stewart, and William Lytton

The term working memory describes the limited capacity, limited duration, often conscious, store of information used by humans in a multitude of everyday cognitive tasks. One possible substrate for working memory is repetitive firing of neurons in the prefrontal cortex (Fuster 1971, Levy, Goldman-Rakic, 2000). This firing persists even when the original stimulus is no longer in view. We used computational modeling to assess effects of one possible mechanism of recurrent firing – feedback connections - on the speed of cognition. Without feedback, memories decay in less than 40 milliseconds. With perfect feedback, memories will persist indefinitely but at the cost of cognitive slowing. Without feedback, a new memory can be represented in less than 20 ms, but with perfect feedback, it takes around 100 ms to lose the old memory and represent the new memory. A linear regression model with feedback , threshold, and noise as predictors accounted for 76 % of the variance in speed of response (changes in F : 132, 33, and 16 respectively, all with p < 0.001). Excess noise interferes with maintenance of working memories. Regenerative feedback (defined as feedback > 1) can compensate for excess noise or neuronal loss, but at the cost of even more slowing. These results suggest a mechanism for cognitive slowing that occurs with aging and with disease - to compensate for increasing noise and/or neuronal loss, feedback of neuronal ensembles is increased at the cost of cognitive slowing.
Serial processing of units with perfect feedback compounds slowing. We show that If information is serially passed among 3 regions, each with perfect feedback, no more than 2 items can easily be processed per second. This timescale is consistent with the speed humans complete many cognitive tasks including the symbol-digit test, the Stroop task, and even reading. Although information can pass from neuron to neuron in a few ms, we hypothesize that recurrent connections with high levels of feedback slow some forms of cognition down to rates closer to about two items a second.

Hi Howard,

We don’t host models, so you’ll have to make your own repository (e.g. on your Github) to share this.

When you’ve made the repository, you can post in https://forum.nengo.ai/c/creations to let the Nengo community know about your abstract/code.

thanks

Howard