Another look at LIDA

One of the interesting things about the Madl, Baars, Franklin LIDA model is the number of memory stores that it envisages. I have thought of consciousness as the ‘leading edge of memory’, at least of episodic memory. Hence my interest in the model’s use of memory.

Let us walk through their cognitive cycle to see what forms of memory are mentioned.

In the perception part of the cycle, the percept is held in the preconscious buffers of LIDAs working memory (workspace) and temporary structures are built. I assume the temporary structures are a preliminary world model. This seems definitely wider than the local sensory memories like visual memory.

The residual contents of the working memory is associated with the incoming percept and with associations from episodic and declarative memory in order to create an updated working memory. In summary, a number of memory stores are associated to give the current workspace: the new percept, the previous workspace, episodic memory, declarative memory. But they do not include and input from a motor system memory.

This resulting workspace is examined by attention processes to bring some parts of the workspace into consciousness. These parts (the novel, relevant, urgent and insistent) are transferred to another memory, the global workspace. This is equivalent to conscious broadcast as the contents of the global workspace are available to many process in the brain.

A procedural memory is postulated, hold various relevant behavior schemes which can retrieve the information they require from the global workspace. Goals are adjusted, actions selected and taken. Results of the action are fed back into the cycle through effects on the environment being included in sensory input.

In summary and in different words we have: individual stores in each sensory mode, a store with the preliminary precept, a store for finished world model, stores for episodic/declarative long-term memory, a store for ongoing behavior plans/goals, a store of action implementation procedures. This model gives a look at the sort of memory stores that are likely to be needed for cognition.

I have a big reservation about this cycle (although in general it is attractive), it does not deal well with prediction. For improvement, if we start with an action procedure rather than incoming sensory information and draw the cycle from that starting point. The action procedure, as well as, and before, producing the action, also is integrated into the existing workspace so as to create a model of the world as it is expected in a fraction of a second. What fraction of a second? Long enough for the conscious broadcast of the global workspace to coincide with the prediction - so that ‘now’ is ‘now’. The incoming sensory information can be compared with the workspace to register any errors and allow the action procedure to be amended if required. In other words, sensory and motor systems have two links: there is an external link through the environment and an internal one through prediction and they are compared in the workspace.

Madl, T., Baars, B., & Franklin, S. (2011). The Timing of the Cognitive Cycle PLoS ONE, 6 (4) DOI: 10.1371/journal.pone.0014803

One thought on “Another look at LIDA

Leave a Reply

Your email address will not be published. Required fields are marked *