wbverwey
commented
8 years ago Hi guys,
Below a short overview of the current state of cognitive models of serial motor skills. You may know a lot of it already, but I guess it is still good to provide a short overview here.
First, it is good to realise that serial motor learning occurs at various processing levels and involves at least two different control mechanisms (for an extensive overview, see Verwey et al., 2015;PsychBull & Rev (http://link.springer.com/article/10.3758/s13423-014-0773-4) ). Knowing this, allows us to focus our modelling efforts on one mechanism first, and gradually extend it to then extend it.
Importantly, we make a distinction between two types of learning, that have typically been investigated in two different, but seemingly similar, tasks: (1) After substantial practice (typically 500 repetitions of a single sequence), short (keying) sequences yield the so-called 'motor chunks' (see e.g., Ramkumar et al., 2016 in Nature Communications; Acuna et al., 2014 in J of Neurophysiol). These motor chunks allow execution based on the motor system (i.e., execution is 'automated' because it does not require central/cognitive resources once the sequence has been initiated). Individual stimuli are not required any more. This has been demonstrated especially with the Discrete Sequence Production (DSP) task (Abrahamse et al., 2013, Frontiers in Human Neuroscience, for an overview). This is called the 'chunking mode'.
(2) The more commonly used serial RT (SRT) task consists of a continuous cycling through, say, a series of 12 responses to stimuli (usually key presses). Here, motor chunks do not develop and participants continue to be dependent on the stimuli displayed. In cognitive models, improvement is attributed to associations that allow 'activation' to prime forthcoming responses. This priming occurs at all levels of processing (Verwey et al., 2015, PsychBull & Rev; Abrahamse et al., 2010, PsychBull & Rev). This type of sequence execution is said to involve the 'associative mode'.
Often results of the DSP and SRT tasks are mixed up, but as said these tasks use different mechanisms. (In the DSP task, associative learning develops too, but in tasks with sequences of limited length, motor chunks soon become dominant because they trigger individual moveents fastest).
As said, motor sequence learning occurs at various processing levels in various ways. The properties of the task determine what mechanisms (i.e., neural structure) are best suited and become dominant. Still, strategic control is possible too, like when people consciously execute a highly practiced, automated, movement sequence again. Also, the use of motor chunks does not mean that translating stimuli into responses is no longer contributing to execution at all.
For our modelling purposes, it seems good to start with 1) the chunking mechanism, i.e., with the development of representations that control short series of movements. These motor chunk representations, for some reason, are limited to about 4-5 elements. Motor chunks may well be stored in the SMA. SMA would control (individual movements in) M1. 2) Longer sequences involve a concatenation of different motor chunks. Concatenation, when occuring consistently over practice, automates too (i.e., requires no central/cognitive guidance). Concatenation is attributed to the basal ganglia (with little practice the basal ganglia would trigger the invividual movements; later, entire motor chunks are represented in SMA). Together, this yields the well-known hierarchical control structure.
We now could start off with the basic chunking mechanim, and then extend this neural model to also cover a) concatenation, b) parallel triggering of movement representations (chunking and translating stimuli), and c) cognitive control. Validating and tuning these neural models is possible with the data of the tens of DSP experiments I have carried out since Verwey (1996, JEP: HPP).
Willem
Two tasks that could be broadly described as "motor chunking":
I took a shot at this during the Nengo Summer School and got increasingly fast reaction times for learning an association, but got kind of overwhelmed about how to scale up to the problem. It's more than just learning an associative memory using Voja and PES. I think it should involve representing the actions as SPs, figuring out a compressed representation from them (circular convolution of something?) and then learning associations based off of that.