In today’s post, I will introduce an example of one proposed mechanism of cognitive control that is discussed in detail in my book, On Task. This is the mechanism of working memory gating.
Consider an example of an everyday act of control. You are driving your car when your cell phone buzzes to notify you of an incoming message or call. For many people, that buzz is a strong cue that elicits the act of checking your phone. And if you were relaxing in your home, this might be an appropriate action to take. Indeed, you might even check your phone without having set a goal to do so. However, cell phone use while driving is dangerous, and in some places, it’s illegal. Thus, in this context, you need to override your tendency to check your phone and do something else, like redirect your attention to the road or click a mute button.
If you are successful in controlling yourself in the car, then you must have in some way used a representation of your context, i.e., driving your car, to choose an appropriate action over the more habitual, but inappropriate one. To do so, you likely held this context in your working memory. Working memory is a system for maintaining a limited amount of information in an active state so that it can be used for on-going processing. Thus, maintaining your car-driving-context in working memory allows it to provide internal state support for choosing the appropriate action.
This simple and widely held conception of how control operates places working memory at the center of successful control. However, in practice, this notion is complicated by the fact that the world is complex and dynamic. We have access to lots of potential contextual information we might maintain that changes from moment-to-moment. Meanwhile, our working memory capacity is limited. Thus, we need a means of controlling the input to working memory to select what information to maintain versus what information to ignore and when to do so.
We also need a way to control output from working memory. In general, for any on-going behavior, we are not constantly monitoring what we are doing with respect to our internal goals. Indeed, with a well-learned action, like a practiced piece of music or a tennis serve, attending to the details of our actions can disrupt rather than help. Conversely, if a context support is needed, but we fail to apply it, we are susceptible to errors and slips of action. In our example, failing to exert control at the right moment might result in us checking our phones before we realize we are doing so. So, there are moments when control is needed, and it is crucial to align our use of information in memory with those moments.
Working Memory Gating Policies
Mechanisms that control the input and output of working memory are sometimes conceptualized by computational cognitive neuroscientists using the metaphor of gates on memory. These gates control the flow of information. Thus, an input gate can open to allow information encountered at an arbitrary time to be placed in a state such that it is accessible at a later time. Conversely, when this input gate is closed, it can prevent transition to this state, protecting against distracting information. On the other end, an output gate can be opened to allow contextual information held in memory to be accessed at the right moment and used as a control signal.
To be useful for control, working memory gates should also be selective, allowing content addressable access to memory. Thus, if more than one piece of contextual information is maintained at a time, a selective gate can update and access specific information in memory, rather than gating everything or nothing.
Selective working memory gating is particularly important because we often have multiple contexts that are relevant to our behavior and evolve over different time scales. Consider, for example, that most everyday sequential behavior is hierarchical in nature, with goals and subgoals at different levels of abstraction. For instance, when I make a cup of coffee, I have an overall goal of making coffee, but also various subgoals like filling a carafe or adding grounds. I will update the subordinate goals while the superordinate one rides along in the background. Further, I may need to update subgoals contingent on my superordinate goal. Selective gating helps manage these complex dynamics among multiple elements of context.
Thus, to carry out a new task or unfamiliar rule, we need to quickly identify the right gating policies for that task. These gating policies will specify what information should be input our output from working memory and under what conditions. Returning to our cell phone example, our gating policy should update the context of driving as a context, and set conditions so we access this context for support when our phone buzzes. Thus, our gating policies realize our rule about using phones while driving within the complex dynamics of our actual drive.
In sum, finding the right working memory gating policies is part of how we are able to connect our goals with our behavior. Experimental evidence indicates that the efficiency with which we acquire and perform new tasks partly depends on our ability to locate and apply the right working memory gating policies, independent of the task rules that specify which stimulus cues what response. People’s behavior shows that they are capable of both transferring and generalizing gating policies to new tasks, leveraging their prior experience to behave flexibly. Thus, successfully controlling ourselves can depend not only on our general capacity for control but also on how well we have matched our policies for working memory gating with the task at hand.
In my next post tomorrow, I will discuss mechanisms for working memory gating in the brain, and how they help us address some of the basic challenges for cognitive control theory.
Thank you for your posts. I’m curious where (or if) you draw the line between ‘working memory’ and ‘episodic memory’. Because in my research, non-hippocampal ‘working memory’ tends to exist on time scales of seconds or less, and continuity beyond that time span (at least in neurotypical people) relies on episodic memory.
In fact, the way I would describe your hypothetical situation is that the ‘world-oriented’ necortical action network (i.e. senses, motor, and the decisions which drive immediate behavior) tends to drive our ‘automatic’ (well-learned) behavior, like driving or checking our phones. But that the default mode network, in concert with the hippocampus, helps supply the additional context (in your example, the fact that I’m driving when my phone rings). The DMN and hippocampus are also involved in the first learning of a new task, which is why we are so aware of that process.
Is that a distinction you are trying to make as well? How do you see episodic memory fit into your control scheme?
Thanks!
Thank you for your question.
I tend to conceptualize working memory in a way following Cowan, such that through retrieval we can activate a portion of long term memory for a brief period (which we call short-term memory), and we can attend and operate on a limited set of such activated items. What gating adds to this picture is a way to manage a fundamental computational tradeoff between stability and flexibility. An interaction between separate mechanisms for maintenance versus gating allows us to set up criteria by which we can both stably hold on to useful information, robust to distraction, while also flexibly updating information once it becomes obsolete. So, while it is certainly the case that control signals can arise from episodic memory, this internal source of information is still subject to this stability-flexibility dilemma if we wish to control ourselves. (For example, we don’t want every thought that jumps to mind to influence our actions any more than we want to be driven by a buzzing phone coming through the senses). There remains a gating demand in order to select and deploy portions of this information as a signal for control. As I will discuss tomorrow, we have focused on loops between networks that include lateral prefrontal cortex and the basal ganglia with regard to this kind of gating. Though, again, these gating loops can certainly interact with networks for episodic memory.
What’s the difference between working memory and short term memory?
That’s a difficult question. In the sense I use them above, short term memory is the active portion of long term memory. As these are active representations, they can influence behavior, but I do not need to be aware of them (as in, for example, priming). But for me to be able to operate or manipulate or maintain these representations against distraction, I need them to be held in the limited focus of my attention. Then, they are in working memory.
Thank you for your reply. I think I’m hearing that my question was orthogonal to what you’re talking about. I hear you using ‘working memory’ to mean the brain’s general buffer at the moment, during a process. My question was about which buffers specifically, but I understand now that it doesn’t really matter, regarding your thesis.
On a side note, I’m struck at how the pandemic creates a natural laboratory for questions about cognitive control. Because we get to witness the entire population of the world alter their normal daily habits, within this new condition of the virus.
Yes, the pandemic is interesting from a control perspective, as you say. It is also interesting to consider why people struggle maintain their behaviors over time from the perspective of cognitive control.
I’m also curious if your group has a working model to explain ‘driving mind’, by which I mean the weird moment when you’ve driven someplace familiar, and when you arrive, you realize you have no memory whatsoever of (at least part of) the drive. That seems like a potential cognitive control issue.
Yes, this is a characteristic of automatic actions when we don’t guide them using cognitive control. This is also why we often don’t notice when we make an error in a well-learned sequential task (like missing a step while making coffee or doubling a letter in a word while typing). Apart from the relationship of novelty, prediction error and attention to memory and awareness, I don’t have a strong account of this, however.