How is memory accounted for in the NEF?

The Neurological Engineering Framework can be used to create systems that use memory in interesting ways. One system (Spaun) is able to memorize (and forget) lists much in the same way as humans do. Another system can learn tasks and apply the knowledge to subtasks using hierarchical reinforcement learning. Both of these tasks involve memory, but how is memory contained within the Neurological Engineering Framework?

1 Answer

The Neurological Engineering Framework does not explicitly state a mechanism for memory. There is no "hard-drive" in the brain for easy retrieval and access. Rather, memory is captured in the connection weights between neural populations and the dynamics of the network.

In the Hierarchical Reinforcement Learning example, linked to in the previous question, the "memory" of the map for the navigation task is saved by modifying the connections weights of the neurons representing the environment.

In the serial list memorisation task, the working memory for saving the semantic pointer list items is preserved in an integrator, which is essentially a neuron population that feeds back on itself.

These are examples of working memory, in other words, the memory that allows you to drive a car without really remembering how exactly you drove it. Long term memory modeling requires quite a few more mechanisms (memory reconsolidation) and structures (hippocampus, amygdala) that haven't been captured in any system built with the NEF yet.

• Are there any developments on this issue? – gota Jan 24 '18 at 16:35
• @NunoCalaim quite a bit! There's a task-specific memory module. Jan Gosmann is finishing his thesis on a detailed model of the hippocampus. This would be a lot easier to link to if I had finished updating the website, but for now you're going to have to type memory into the search box of the publications tab. – Seanny123 Jan 24 '18 at 16:43