Posted: February 18, 2013 at 5:02 pm
The importance of simulation (prediction) in dreaming and mind-wandering literature should be integrated into the current conception of DM3. There are two aspects of continuing development: (1) The current propagation of activation for free-association is inherited from previous projects (MAM, DM1 and DM2) and not well situated in theory. (2) There is no feedback from the world that can be used as a reward that could be used to drive intrinsic motivation.
Priming occurs when a previous stimulus improves current perceptual performance. For example, if you show a person a string of arbitrary images the time it takes a subject to report their name is quite long. If all the images are in one conceptual context, like kitchen appliances, then recognition and reporting is significantly faster. It’s easier for us to recognize a stimulus if it appears in the context of conceptually related stimulus (or even located in the same place, or similar in shape or colour). Priming seems to predispose perception to the recognition of stable patterns of concepts and features. Priming could be explained by the activation of all related concepts/categories when any recognition takes place. This existing activation of related categories is a shortcut for future perception of those same categories. The recognition of a toaster primes all other related categories, allowing quick judgements for toast and jam. An alternative explanation is that related concepts are activated because they are expected to be related to the current stimulus. A toaster primes bread because bread is expected to appear in close succession to the toaster.
Thinking of priming as prediction leads to the conception of perception as a predictive process itself, this fits well with the theory behind this project that considers perception as a highly constrained form of dreaming / prediction, and dreaming as a highly unconstrained form of perception / prediction. One idea of how to manifest this in the system is to introduce the notion of a context. A context is an area of space and time where percepts co-occur. In the absence of supervision, goals or action, these contexts can be considered semantically, or to some degree causally, related. All percepts within a certain context will be linked to a degree.
During the first perceptual stimulus frame, activated percepts are linked to a degree proportional the distance in time and space between them, creating an initial context. For the consecutive frames, each percept that is activated is checked to see if it has been primed (connected to the previous context). For all those percepts activated and primed, their connection to the percepts in the previous frame (previous context) is increased. For those percepts that have not been primed, but are are activated, their connection to the previous context is increased. For percepts that were primed but were not activated, their connection to the previous context decreases.
Percepts appear to the viewer when that are activated to a sufficient degree, but the degree of priming is not directly visible. Priming is manifest in the propagation of activation between percepts in the same context. As many percepts will be associated with multiple contexts, activation may spread between contexts. Precept A may cause the activation of percept B, to a degree proportional to the weight that connects them. The weaker the connection, the less likely the signal would be propagated. If percept B is activated, then it passes that activation onto neighbours in its context. The system exhibits habituation, where percepts that are repeatedly activated require more stimulation to be activated in the future. At this point it is unclear whether context links should continue to be learned for habituated percepts.
During mind-wandering and dreaming, the system behaves similarly as previously proposed. During mind-wandering, context links are still being learned, but the presence of priming coupled with habituation causes a mix of external and internally activated percepts to be visible. During dreaming, activations are due to latent activation perpetuated by the amplification of signals. While activations are decaying, so could the associated contextual links. While activations are amplifying, associated contextual links would increase. Dreaming would then prune weak links, which would be reduced to 0, and leave strong links intact at levels similar to those before dreaming.