Decisions can differ depending on the context that surrounds them. Analyses of the prefrontal cortex region of the monkey brain indicate that a dynamical process at the neuronal population level controls this behaviour. See Article p.78
Neurons typically receive inputs from many thousands of others. This dense connectivity suggests that brain functions such as perception, cognition and motor control result from the concerted activity of populations of neurons, rather than from single-cell activity. Nevertheless, neuroscientists have long made discoveries by recording from, and thinking about, one neuron at a time. But more-recent developments in electrophysiological recording1 and imaging2 of the dynamics of hundreds of neurons simultaneously, alongside advances in data-analytic and theoretical tools3,4, highlight the importance of understanding the relationship between the dynamics of large neural populations and the computations those dynamics embody5,6,7,8,9,10,11. In a tour de force reported on page 78 of this issue, Mante et al.12 use electrophysiology together with sophisticated behavioural and neural modelling to add a population-dynamics view of the computations underlying flexible, context-dependent behaviour.
How you respond to a glance up a busy street should be very different depending on whether you intend to cross the road or hail a taxi. In the first case, you should pay attention to the motion of the cars, in the second case to their colour. Such stimulus-feature selection is an example of cognitive control13 — our brains' ability to flexibly adapt how we process information from moment to moment, depending on current goals. Mante and colleagues wanted to know how the brain decides which aspect of a stimulus should guide a behaviour.
The authors designed an elegant task to explore this flexibility (Fig. 1a). They trained monkeys to look at a screen showing randomly moving dots, each dot randomly coloured either red or green. In trials in the colour context, the monkeys had to decide whether there were more red dots or more green dots, regardless of the dots' direction of motion. In the motion context, the animals decided whether most of the dots were moving left or right, regardless of the dots' colour. In both contexts, the same set of stimuli, containing motion and colour signals, was used.
The team then recorded from the monkeys' prefrontal cortex (PFC), a brain region thought to be crucial for flexible, context-dependent behaviour and which also controls the motor activity that animals use to report their decisions. Consistent with previous work14,15, the activity of individual PFC neurons depended in complex, time-varying ways on multiple aspects of the task, including the motion and colour signals in the stimulus, the context and the monkey's decision. In what initially seemed a bewildering array of responses, these dependencies varied greatly across neurons.
To look at the data from a population viewpoint, the authors put together all the individual recordings for a given stimulus, and defined the neural population's response as the dynamical trajectory followed in a high-dimensional space in which each dimension represents the activity of one of the recorded neurons. They then used linear regression to find the axes (corresponding to a direction) in this space that best represented responses to the motion and the colour in the stimulus, and the monkeys' decision (Fig. 1b). They could therefore estimate separately how each signal's representation in the PFC developed over time, and how it depended on the context.
Which features of the data explained the context dependence of the behaviour? Not changes in the decision axis, which pointed in the same direction in both contexts. Nor could the behaviour be explained by irrelevant inputs being blocked from reaching the PFC, or by the inputs being changed in some context-dependent way, because the stimulus axes and the stimulus signals (motion and colour) on those axes remained remarkably constant across the two contexts. Yet, somehow, the decision signal was driven by the motion signal in the motion context, and by the colour signal in the colour context.
To investigate how the switch in response to the different contexts came about, Mante et al. used modelling. They took model neurons with nonlinear input–output functions, and densely connected them in a recurrent network. They provided the network with inputs representing the motion and colour signals and, separately, the current context signal. The activity of one of the model neurons was used to indicate the decision of the network.
The team then 'trained' the network by finding the strengths of the neuronal connections required to produce the appropriate, context-dependent decision in each trial. Starting the training from different initial random connection strengths led to different connection patterns that solved the problem. But cutting-edge analyses of how the model networks were solving the problem revealed that all the solutions shared two dynamical principles.
First, a line of closely spaced, stable, fixed points lay along the decision axis, with the two possible outcomes at opposite ends of the line. During each trial, stimulus-driven evidence for or against a decision accumulated along that line by gradually nudging the system from one fixed point to the next. Second, in dynamical systems, some inputs can cause a long-lasting perturbation in the system's trajectory, whereas other inputs can be short-lived, and rapidly decay away. In Mante and co-workers' model networks, the critical feature that depended on context was which of the inputs was short-lived. In the colour context, for example, signals from the irrelevant input (motion) were rapidly quenched and had no lasting impact on the neural trajectory, whereas colour signals persisted to drive the decision (Fig. 1c).
In other words, the context dependence was not achieved by controlling or gating the flow of information from one neuronal population to another. Instead, it was achieved within a single, densely interconnected cell population by controlling the directions along which the population dynamics quenched the neural activity. This dynamical principle can only be understood — in fact, only exists — at the population level. And because the neural trajectories in the model closely resembled the neural trajectories in the data, Mante et al. propose this principle as the neural basis of flexible, context-dependent processing.
This fascinating study raises some intriguing questions. First, how do we go beyond correlations and causally test the authors' proposed principle? With existing perturbation methods, we cannot align manipulations with the critical axes that define the population's dynamical behaviour. Second, is the region of PFC that the authors recorded from even necessary for the context-dependent behaviour investigated here? The behavioural task used is new, so this question has never been addressed. Third, the conclusions on the behaviour of neural populations are based on data from many separate single-neuron recordings. Will the same principles hold when multi-neuron recordings are used to directly see the population neural trajectories in single trials? Finally, what is the specific mechanism that the network uses to change its quenching directions? These questions aside, Mante et al. have led the field with novel and deep ideas that should drive much discussion and thinking.
Pfeiffer, B. E. & Foster, D. J. Nature 497, 74–79 (2013).
Dombeck, D. A., Khabbaz, A. N., Collman, F., Adelman, T. L. & Tank, D. W. Neuron 56, 43–57 (2007).
Yu, B. M. et al. J. Neurophysiol. 102, 614–635 (2009).
Goldman, M. S. Neuron 61, 621–634 (2009).
Laurent, G. Nature Rev. Neurosci. 3, 884–895 (2002).
Briggman, K. L., Abarbanel, H. D. I. & Kristan, W. B. Jr Science 307, 896–901 (2005).
Machens, C. K., Romo, R. & Brody, C. D. J. Neurosci. 30, 350–360 (2010).
Harvey, C. D., Coen, P. & Tank, D. W. Nature 484, 62–68 (2012).
Churchland, M. M. et al. Nature 487, 51–56 (2012).
Shenoy, K. V., Sahani, M. & Churchland, M. M. Annu. Rev. Neurosci. 36, 337–359 (2013).
Rigotti, M. et al. Nature 497, 585–590 (2013).
Mante, V., Sussillo, D., Shenoy, K. V. & Newsome, W. T. Nature 503, 78–84 (2013).
Miller, E. K. & Cohen, J. D. Annu. Rev. Neurosci. 24, 167–202 (2001).
Roy, J. E., Riesenhuber, M., Poggio, T. & Miller, E. K. J. Neurosci. 30, 8519–8528 (2010).
Clark, K. L., Noudoost, B. & Moore, T. J. Neurosci. 32, 10907–10914 (2012).
About this article
Distributed representations of action sequences in anterior cingulate cortex: A recurrent neural network approach
Psychonomic Bulletin & Review (2018)