This book transformed my understanding of cognitive control. It’s the most up-to-date and comprehensive introduction to the topic that I have seen. If you ever wondered how we can explain cognitive control without assuming a tiny controller pulling levers in the brain, then this book is for you.
Badre, D (2020): On task — How our brain gets things done. Princeton, NJ: Princeton University Press. Publisher Link
How do we manage to make a cup of coffee without planning every single step? How can we deal with new situations that we never encountered before and come up with good solutions? Why do we struggle to remember things when we get older? This and other questions are addressed in David Badre’s authoritative text on the science of cognitive control — a similar, commonly used concept is executive function. Across 10 chapters, Badre introduces the history of thinking about cognitive control, the evolutions of this field, and the current state-of-the-art. He skilfully manages to provide a conceptual understanding without shying away from getting into the computational models that have driven much of the progress in this field in recent years. I found it especially refreshing that the book was so up-to-date. Many books from eminent scientists reflect insights that have become textbook knowledge with only the occasional morsel of new and exciting developments sprinkled in. In contrast, this book is a true reflection of current thinking and provides an excellent overview for anyone who is interested in this field. That being said, I would not consider this book to be a popular science book. While it is very readable, the level of detail may be a bit too much for someone without any foundational knowledge in psychology, cognitive science, or related fields.
Please note: This is my summary of the content of the book. While I tried to capture the main points, this is necessarily a subjective and incomplete overview. I highly recommend that you pick up the book yourself to get the full details.
Chapter 1: What lies in the gap between knowledge and action?
When we formulate abstract goals like making a cup of coffee, we are not explicitly thinking about all the small steps that are necessary to accomplish this goal. Cognitive control is concerned with explaining how our brain plans and executes tasks.
Nowadays, cognitive control is inextricably linked to the prefrontal cortex (PFC). However, the function of the PFC was elusive for much of the history of psychology.
In his landmark 1890 psychology text, The Principles of Psychology, Harvard psychologist William James spent a chapter reviewing the known functions of the brain at the time, and he captured the perplexed state of the field concerning the prefrontal cortex. He called the frontal lobes a “puzzle” and stated, “The anterior frontal lobes, for example, so far as is yet known, have no definite functions. … neither stimulation nor excision of the prefrontal lobes produces any symptoms whatever.”
Descriptions of patients with damage to the PFC showed that they had difficulties carrying out complex but routine tasks, like preparing food for a dinner party, organising their belongings, or arranging travel to be in time for appointments.
… frontal patients were fluent in conversation, knowledgeable, and competent in their moment-to-moment interactions. Further, many of the problems these patients exhibited in their lives might be seen as annoying personality traits in lots of perfectly healthy people.
The difficulties exhibited by patients with PFC damage are hard to capture with laboratory tests. Even modern gold standard assessments explain only 18–20% of the variance in patient’s everyday difficulties.
A theory of cognitive control needs to explain:
- how knowledge is translated into action
1.1 selecting one option from many possibilities
1.2 planning across multiple time scales and levels of abstraction
2. how action policies and plans are learned
2.1 generalising learning to new situations
2.2 isolating the most important aspects of a problem
Chapter 2: The origin of human cognitive control
Cognitive control goes beyond inhibiting impulses, making rational choices, and switching between tasks. Cognitive control probably evolved to make complex planning possible. It enabled humans to select the best actions from a vast array of possible actions, and even to come up with entirely new solutions by combining the building blocks of actions in new ways.
Our human cognitive control system is general and has two basic ingredients. First, we are capable of conceiving of future situations and goals we have never experienced or considered before, within either our own lifetime or the lifetimes of our ancestors. Second, our control systems can plot out the complex actions necessary to achieve that future.
A generative and compositional action system made it even possible to imagine an entirely different future (episodic future thought) and devise plans to make it reality.
Cognitive control evolved to make the conceivable become realizable given the particular organization of our brain and our action systems.
At brain level, the increase in cognitive control in humans is reflected in the expansion of the prefrontal cortex and its increased connectivity with the rest of the brain.
Chapter 3: The stability-flexibility dilemma
The cognitive control system needs to be able to adapt action policies according to the context. For instance, you should pick up the phone when it rings, but not if you are driving a car. Some associations between a cue and an action are so strong that they are almost automatic. Other associations are weak. The weaker association will lose if the cognitive control system does not step in.
The cognitive control system is thought to achieve control by manipulating so-called gates:
First, a gate is needed to control what information from the world is allowed into networks supporting working memory. When the gate is open, new information enters the prefrontal cortex from perceptual systems to be maintained in working memory. … When the gate is closed, irrelevant information is kept out of prefrontal circuits, and what is there is protected.
These gates are probably implemented in the brain as connections between the frontal cortex, basal ganglia and thalamus that can either let impulses pass through or inhibit them.
Chapter 4: Hierarchies in the head
Plans consist of various levels of abstraction ranging from high-level goals, e.g. “get a cup of coffee”, to the lowest level of implementation, e.g. “move the hand forward by 2 cm towards the cup…”. There are different possibilities of how the cognitive control system may deal with these levels of abstraction. Early chaining theories suggested that one action triggers the next action in a sequence. However, these theories could not explain why we can accomplish a goal even when we fail at one of the subtasks, e.g. typing to the end of a sentence despite a typo halfway through. Karl Lashley proposed an alternative theory that assumed a central representation of an overall goal that is separate from the representation of the subgoals. However, this cannot explain some kinds of errors. For example, someone is making coffee every day by boiling water and then getting a coffee filter. One day, the person decides to make tea but absent-mindedly continues to get the coffee filters out (capture error). This suggests that the subgoal of boiling water is somehow linked to the next step in the sequence.
The model wasn’t just representing the subtask “pouring water”; it had a unique pattern for “pouring water while making coffee.” This way, the model kept the memory of the overall task active, even if some of the subgoals were lost or degraded. Unlike in early chaining models, the model kept track of the overall task, but it didn’t need a separate representation of that task to do so.
This partial representation of the overall task in the subtasks is also reflected at the brain level. Subtasks of the same overall goal show similar patterns of brain activation.
Lashley’s hierarchical model is correct for unfamiliar tasks, though. For unfamiliar tasks, there is a distinct abstract representation. The representation of the overall goal is used to put together a sequence by selecting appropriate automatic routines and monitoring the progress on the task. This hierarchical organisation makes behaviour flexible and adaptable, even for completely novel tasks. At the neural level, the hierarchical organisation is implemented through cortico-striatal-thalamic loops that gate working memory. More anterior areas tend to influence the gating of more posterior area.
Chapter 5: The tao of multitasking
Interleaving two tasks leads to decrements in performance and speed compared to doing both tasks separately. For instance, driving in a car simulation and simultaneously doing a cognitive task leads to much slower break times and lower performance on the cognitive tasks. The multi-tasking costs arise from shared bottlenecks in cognitive processing.
Our general ability to get things done comes from a system capable of building any task we can think of on the fly out of a library of abstract, compositional representations that come to mind readily when cued. However, the very generality of these representations makes it likely that any two complex tasks will rely on at least some overlapping representations.
The overlap in the representations required by both tasks created a conflict that needs to be resolved. This can either happen ‘bottom-up’, e.g. when we automatically attend when hearing our name, or ‘top-down’ when we select the input that matches the current goal.
With experience, we can reduce the overlap between representation and, thereby, decrease the multi-tasking cost. For instance, most people have no problem with walking and chewing gum at the same time.
As we perform a task repeatedly and automate that task, we may, through mechanisms of cortical plasticity, come to represent that task using more specific neural populations. A shift away from abstract, generalizable neural representations that are susceptible to competition and interference, and toward more task-specific representations that can be used in parallel, might hint at the kinds of changes that underlie the shift toward automaticity.
This model of multitasking implies that we cannot train a general multitasking ability — “there is nothing to train there”. However, people with better cognitive control will do better in multi-tasking situations because they will be better equipped to minimise the interference between tasks through top-down selection.
Chapter 6: Stopping
Inhibition is a central term in Psychology with a long tradition. It is generally understood as an ability that allows a person to control their behaviour by suppressing thoughts, memories, or movements. However, the term inhibition is also used to in different contexts with different meaning. For instance, at the neural level, inhibition is a physiological interaction by which one neuron’s activity leads to a suppression of activity in the other neuron. Further, it is possible to have inhibition as an outcome, i.e. something is not carried out, without having inhibition as a separate process.
Experimental evidence from the stop-signal task suggests that there is a separate process for inhibiting a motor response. This process is implemented in a frontal-basal ganglia network. Stopping is not selectively, i.e. activation of the inhibitory network leads to suppression of all motor responses. The same network is also involved in stopping cognitive processes. However, this does not imply that all psychological inhibitory processes are mediated by the stopping network. For instance, selecting between different responses via competition and gating is likely to support many processes that appear like inhibition at the psychological level.
… the degree of sensation seeking across people is not correlated with brain measures of the stopping network. They differ, rather, in the thickness of the cortex in areas of the frontoparietal system associated with working memory and selective gating. Thus, rather than having a problem setting thresholds to make difficult decisions, sensation seekers making the decision to abuse drugs may differently value various actions, like taking cocaine, used to make choices. As a result, they gate impulsive behaviors that lead to addiction. This outcome is similar to that of individuals with weak stopping control, but the cause and mechanism are distinct.
Chapter 7: The cost and benefits of control
Cognitive control and motivation are closely linked. For instance, when the incentives are increased, people perform better on tasks of cognitive control. At the brain-level, dopamine neurons in the striatum signal reward. Their firing enhances the stability of working memory representations in the cortex.
There seems to be a cost to exerting cognitive control. Given the choice between two tasks, people will pick the one that requires less cognitive control. Different theories have been proposed that try to explain this phenomenon. Some theories assume that cognitive control depletes some resource, e.g. ego depletion hypothesis. However, this does not explain why people can continue with a task when the incentives are increased. Further, there are several failed replication attempts for the original ego depletion studies.
Assuming your basic bodily needs of food, sleep, and water are taken care of, you could never do so many math problems, for example, that you couldn’t possibly do another, even if you wanted to. Rather, a reasonable heuristic would be if the motivation is there, so is the mental capacity.
So, why do we not like exerting cognitive control on the same task for a long time? It might be a mechanism to keep us from missing out on potentially better opportunities (opportunity cost hypothesis). The experience of mental effort and the relative comparison of rewards that can be gained from different courses of action is likely to be mediated by the anterior cingulate cortex.
Chapter 8: The information retrieval problem
The information retrieval problem concerns how a system recovers stored information with a value that exceeds the cost of retrieval itself.
There are two components to systems that successfully solve the information retrieval problem, namely being sensitive to the frequency at which items are accessed and being sensitive to the context in which they are relevant. Human memory fits both descriptions. It is easier to retrieve memories that are more often accessed and that are relevant in the current context. Cognitive control plays a role in retrieving the memory and processing the retrieved information afterwards. This is akin to putting together the right search string for a Google search and then going through the results to find the most relevant hits.
Post-retrieval control is also a prime means by which we offset some of the major causes of forgetting. We often think of forgetting as a decay process. Information gets put into memory, and then like segments going bad on an old hard drive or a picture fading in time, our memory just corrodes and deteriorates. However, this folk theory of forgetting is not quite right. Rather than being a passive time-dependent decay process, much of forgetting can be explained by active interference from other memories that block and corrupt the target memory.
Controlled retrieval is mediated by a network of areas in the frontal lobe associated with cognitive control and temporal areas important for memory, such as the hippocampus. In contrast, regions involved in post-retrieval control overlap with the frontoparietal control system.
If we view memory in terms of information retrieval, it also makes clear how we might improve our memories. If we want to learn something, we must use it, because use is what our memory system cares about.
In essence, we don’t use our memories to perform tasks; rather, our tasks are what we use to remember.
Chapter 9: Cognitive control over the lifespan
Cognitive control takes a long time to develop with increases that extend into the third decade of life. Some theories describe children as if their frontal lobes are not operational and highly vulnerable to environmental influences — Badre calls this the Death Star model. However, it is unlikely that humans would have evolved with such a glaring vulnerability in their system. Further, the Death Star model does not explain why cognitive control takes so long to develop. The alternative theory suggests that the long development reflects the malleability of cognitive control. Because of the abstract, complex, and adaptable nature of cognitive control, a lot of experience is needed to adjust the cognitive control system to the requirements of the environment.
The kinds of tools we use to accomplish tasks today, like computers, smartphones, and combustion engines, were not on the landscape when our ancestors were flint knapping their way through dinner. Thus, our control systems adapt and are customized to the world we are in, and that type of optimizing requires data.
This view suggests that children should be exposed to diverse experiences to train their cognitive control, i.e. environmental enrichment. In particular, unstructured activities that place a higher demand on cognitive control are likely to be beneficial.
This view of cognitive control development, then, places a special emphasis on learning and experience, and in particular on the diversity of experience we have during childhood. To build useful and abstract gating policies that apply in lots of situations later in life, we need to try to control ourselves in lots of different settings.
This view also has implications for cognitive control problems. They may arise when there is a mismatch between the current environment and the environment that shaped the tuning of the cognitive control system over development.
At the other end of the lifespan, the picture looks less positive.
… there is evidence that all aspects of cognitive control will show some age-related decline. This is particularly the case in novel settings, like traveling; or with new open-ended problems; or with tasks with which we have little experience or practice, like figuring out that new smart TV.
This is not only an academic concern. Measures of cognitive control are associated with the ability to live independently in older adults. Part of the decline is explained by brain ageing in the systems supporting cognitive control. However, other systems may also be involved. Older adults show higher engagement of cognitive control systems during challenging tasks, which may indicate that they use cognitive control to compensate for deficits in other systems. Unfortunately, there is no easy fix for this. The scientific evidence suggests that currently available brain training games marketed to combat cognitive decline do not translate to real-world improvements.
Chapter 10: Postscript — Getting things done that matter
Many of factors that govern cognitive control in humans apply more widely to many complex systems. For instance, to tackle climate change, society must devise policies to translate the knowledge about climate change into action.
Whether stability versus flexibility, generality versus specificity, informativity versus accuracy, or simultaneity versus opportunity, these trade-offs are found not just in our brain but in any system that is highly general and needs to control the states in which it finds itself.