Cognitive science has developed a wide variety of theories about human cognition. While some of these theories are merely descriptive (i.e. they describe human cognitive behaviour), many are mechanistic, in that they postulate a set of internal components which interact over time to produce the observed behaviour. These components can be referred to as a cognitive architecture.
The most successful and widely used cognitive architecture is ACT-R. This includes components for memory, vision, motor behaviour, time estimation, and central executive control. These same components have been used for models of mental arithmetic, problem solving, task switching, car driving, phone dialing, sequence memory, GUI usage, and so on (see here for more examples and related publications). These models come from many different people and many different labs, making it the most widely used and tested cognitive archicture.
The focus of ACT-R research has been to develop this common setof components for modelling cognitive activities. By re-using these same components, they can minimize the amount of parameter fitting and model-tweaking needed to match human behaviour on all these tasks. However, they have been focussed on what the brain does, not how it does it. That is, they want to identify the algorithms behind cognition, not how neurons implement that algorithm.
Recently, John Anderson (the core person behind ACT-R) has been doing a lot of work with fMRI to try to identify the neurological correlates to the various components of ACT-R. The intent here is to bring in neural evidence to help constrain ACT-R theory, which right now is mostly constrained by psychology data (i.e. the model's behaviour has to match human reaction times, error rates, and so on). Adding in neural constraints can help to suggest modifications to ACT-R, or provide support for some of the ACT-R assumptions. This has led to a work comparing the activity of particular brain areas to various ACT-R components, in tasks such as the Tower of Hanoi, and mental symbol manipulation (for more, see here). There's also a short overview of this approach here. and for a more complete overview, see John Anderson's book "How Can the Human Mind Occur in the Physical Universe?".
It is definitely promising to see a degree of correlation between ACT-R predictions of BOLD signals and real data. However, without an explicit implementational story for the various components, many assumptions have to be made to create these BOLD predictions. Furthermore, if we had a fully neural implementation of these components, we could also bring in a lot of other neural evidence involving spiking patterns, connectivity, neurotransmitter use, drug effects, and even DBS.
Our goal is to create a fully neural implementation of each of the components of ACT-R, using realistic spiking neurons constrained by known neural anatomy.
We have chosen ACT-R as an interesting minimal target for building large-scale neural systems. If we can build a neural system with components that can perform the few basic functions needed by ACT-R, then we can relate the resulting neural model to a wide range of behavioural data, while still being able to look at low-level neural effects.
Currently, we are focusing on the basal ganglia, which for the ACT-R people correlates to a production system, along with a reinforcement learning system to use reward feedback to control action selection. This certainly doesn't completely capture everything about the basal ganglia, but there seems to be enough evidence that it's at least in the right general ballpark. It fits in nicely with the dimensionality reduction ideas, and the dopamine/reinforcement learning connection.
This basic model of this system is currently at a stage where it can do basic action selection based on the current context, choose an action that will modify the current context, modify that context, and choose a new action. Interestingly, constraining the neural model to use GABA results in a model that requires about 50ms to make a choice, which is the experimental best-fit to a wide range of behavioural data.