User:Graeme E. Smith/Skill Memory: Architecture Scope and Limitations

Skill Memory Architecture, Scope and Limitations Graeme E. Smith, GreySmith Institute of Advanced Studies  http://en.wikiversity.org/wiki/Portal:GreySmith_Institute   http://en.wikiversity.org/wiki/User:GreySmith_Institute  grysmith@telus.net When Dr. Edelman wanted to win a robot soccer match against MIT, he pulled out a Brain based device designed to work like the cerebellum. Perhaps their robots are more efficient now... In the IDA cognitive Architecture which was the basis of LIDA, the first so called functionally conscious robot, Stan Franklin included in the design an "action oriented system" which learned behavior rules in both a top-down and a bottom-up fashion. It is generally accepted that a cognitive architecture should have some sort of skill memory. In this article I introduce the idea of a two phase skill memory based on the cerebellum. I carefully explain how this relates to what we know of the attention system, and postulate the existence of ASOs, a specific kind of signal that will allow us to detect automation sequences arising from the cerebellum. I also suggest that this same mechanism might work to create a macro-language from complicit attention functions. To understand skill memory, we must first understand demand addressing and the triangular circuit of attention. Demand addressing is a concept first thought up in DesCartez time when it was thought that the brain was a reservoir for a memory fluid. The concept was that somehow when we demanded a memory, that particular memory valve would open and the memory fluid would be released. By the time of computers, we were thinking more about telephone switch boards, where the operator plugged in the switches to make a particular memory available. Today we talk about memory addresses, and every memory location in a computer has its own address by which it can be accessed. Brains don't work that way, as Dr. Miller found out in the 1950's. Instead if we can take David Marrs word for it, the basic memory circuit is a content addressable memory, and if we can take Jerry Fodors word for it, the memory is implicit, and phenomenal. in The Dual Mode Cortex I describe how a demand type memory based on place-code addressing is implemented using the 5th layer apical dendrites. This memory access relies heavily on CLUMP addressing via the Thalamus to pre-activate neural groups and thus output qualar memories. It is exactly this type of addressing that is needed to trigger the cerebellar cortex, and so our story begins with the Thalamus outputs triggered by the clump addressing mechanism. Some of these activations trigger action centers such as the primary motor cortex, and some of these activations trigger a special kind of attention I call complicit attention, which combines a data quale with a function quale, to direct processing to a specific function. According to David Marr, activating these action/function areas of the brain, triggers a reaction in an area called the inferior olive, where the action/function areas are mapped out before being sent to the brain stem to trigger actual movement. It is at this point that we bring in the cerebellar cortex. The cerebellar cortex has a connection to the cerebral cortex, which is fed into its mossy dendrite layer at layer 1, Then, it has a connection to the inferior olive, which maps out to a one-to-one relationship with the purkinje cells in the cerebellum, and is probably connected via the rising fibers, which in turn trigger the purkinje cells. What this creates is a sensitivity to the patterns that form in the mossydendrites which are thought to be context related patterns, and the activation of specific action/function areas in the cortex. The purkinje cells are capable of holding multiple patterns and defining pseudo-sequences by linking context to preactivation of another purkinje cell, in such a way as to create an implicit learned reflex that echos popular activities and processing sequences. These pseudo sequences are linked in turn to contexts, so that a pseudo-sequence will begin eventually to respond when the context exists without requiring prior inferior olive activity. Not only that, but the sequence is sped up by the fact that the context is all that is required to trigger the next stage in the sequence, so time can be compressed especially when adrenaline is involved. This is why slow dancing such as in Tai Chi, or Kung Fu, can be smoothly translated into faster action in an emergency or at a tournament. Just as in the other memory systems that have implicit states, the Cerebellum does not produce one sequence at a time, it produces a number of alternate sequences from which to pick. To tag each one with a different frequency, in the Alpha frequency range to form an ASO or Alpha (frequency) Synchronous Oscillation, is required so that the top-down attention probably in the ACC can select from among the options. This is where we think the Ventro-Lateral PFC comes into play. The Ventro-Lateral PFC is connected to the SMA where sequences from the cerebellum are first recognized, and then compared to come up with a filter that will allow the ACC to select one. If this fails, the ACC then selects another option, it looks for intentional rules that can be used to disambiguate the selection. If that fails the ACC then selects the last option which is the triggering of a conscious state, so that it can use volition to select from among the options. Volition has the option of not only not selecting one of the previously rejected sequences, but of creating its own new sequence. This is the late Volition model, and is considered by some to be controversial, because the motivation for intentional states that are not volitional is not well understood in classical or folk psychology. Another problem with this model is that it supposes that people are not actually operating at full consciousness most of the time. This vies with the viewpoint experienced by the individual that they are conscious all the time they are awake. Perhaps this dichotomy exists because we are talking about two different things, the experience by introspection of a "feeling of consciousness" versus a functional state of consciousness during which the brain does volitional activities. I think that what I should stress here is the difference between awareness and consciousness. Awareness is the state of receiving a near constant input about our relationship to the environment, but need not involve consciousness, while consciousness is a volitional state, during which we are aware of our awareness, and actively choose among the options, which interactions we should have with our environment. I maintain that this is a much rarer state than awareness alone, because most of the time we can "go through the actions" without having to think. You can tell when someone is actively conscious because they "take time to think" instead of just reacting. Libett has indicated that Volition doesn't start until 500 milliseconds after the ACC is first lit up. Other researchers have suggested that 80 milliseconds is a better estimate. The question seems to be related to specific signals indicating that a choice has been made by the ACC sending a signal that releases one of the selections. Part of the reason there is a discussion is that we don't have a consensus agreement on which signal indicates volition. This model might indicate that intention involves one signal while volition involves another. Since Volition must wait for intention to complete, the later signal (Possibly Libetts signal) is likely to be the one that is associated with volition. Part of the problem with this discussion is that the sensation of thinking does not seem to take that long. One of the problems with this theory is that our sensation of making a choice, does not happen immediately it is delayed. One of the reasons that Libett's work is still quoted even though it is controversial, is because he can clearly show that the ACC lights up long before subjects acknowledge that they are making a decision. This means that the subjective experience of making a choice is not concurrent with the mental task of making a choice. I will get into that later in my consciousness theory. What is important at this point in time is the idea of the ACC as being the area that decides which sequence we execute, and the fact that it can draw on successively higher order processing to disambiguate or replace the sequences that have been suggested by the skill memory.