Scientific Understanding of Consciousness
Prefrontal Cortex (PFC) tuned to Mixtures of Multiple Task-Related Aspects
Nature 497, 585–590 (30 May 2013)
The importance of mixed selectivity in complex cognitive tasks
Center for Theoretical Neuroscience, Columbia University College of Physicians and Surgeons, New York, New York 10032, USA
Mattia Rigotti, Omri Barak & Stefano Fusi
Center for Neural Science, New York University, New York, New York 10003, USA
Mattia Rigotti, Xiao-Jing Wang & Nathaniel D. Daw
Department of Psychology, New York University, New York, New York 10003, USA
Mattia Rigotti & Nathaniel D. Daw
The Picower Institute for Learning and Memory & Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA
Melissa R. Warden & Earl K. Miller
Department of Bioengineering, Stanford University, Stanford 94305, California, USA
Melissa R. Warden
Department of Neurobiology, Kavli Institute for Neuroscience, Yale University School of Medicine, New Haven, Connecticut 06525, USA
Present address: Department of Physiology, Technion Medical School, Haifa, 31096, Israel.
Single-neuron activity in the prefrontal cortex (PFC) is tuned to mixtures of multiple task-related aspects. Such mixed selectivity is highly heterogeneous, seemingly disordered and therefore difficult to interpret. We analysed the neural activity recorded in monkeys during an object sequence memory task to identify a role of mixed selectivity in subserving the cognitive functions ascribed to the PFC. We show that mixed selectivity neurons encode distributed information about all task-relevant aspects. Each aspect can be decoded from the population of neurons even when single-cell selectivity to that aspect is eliminated. Moreover, mixed selectivity offers a significant computational advantage over specialized responses in terms of the repertoire of input–output functions implementable by readout neurons. This advantage originates from the highly diverse nonlinear selectivity to mixtures of task-relevant variables, a signature of high-dimensional neural representations. Crucially, this dimensionality is predictive of animal behaviour as it collapses in error trials. Our findings recommend a shift of focus for future studies from neurons that have easily interpretable response tuning to the widely observed, but rarely analysed, mixed selectivity neurons.
Neurophysiology experiments in behaving animals are often analysed and modelled with a reverse engineering perspective, with the more or less explicit intention to identify highly specialized components with distinct functional roles in the behaviour under study. Physiologists often find the components they are looking for, contributing to the understanding of the functions and the underlying mechanisms of various brain areas, but they are also bewildered by numerous observations that are difficult to interpret. Many cells, especially in higher-order brain structures like the prefrontal cortex (PFC), often have complex and diverse response properties that are not organized anatomically, and that simultaneously reflect different parameters. These neurons are said to have mixed selectivity to multiple aspects of the task. For instance, in rule-based sensory-motor mapping tasks (such as the Wisconsin card sorting test), the response of a PFC cell may be correlated with parameters of the sensory stimuli, task rule, motor response or any combination of these. The predominance of these mixed selectivity neurons seems to be a hallmark of PFC and other brain structures involved in cognition. Understanding such neural representations has been a major conceptual challenge in the field.
We showed that the mixed selectivity that is commonly observed in PFC responses can be interpreted as a signature of high-dimensional neural representations. One advantage of high dimensionality is that information about all task-relevant aspects and their combinations is represented in a way that is easily accessible to linear classifiers, such as simple neuron models. The information is distributed across multiple neurons in an ‘explicit’ format that allows a readout neuron to implement an arbitrary classification of its inputs. Previous studies have already shown that a linear readout is often sufficient to decode particular task aspects or to perform specific tasks. Here, by showing that the neural representations are high-dimensional, we demonstrate that any binary choice task involving the 24 experimental conditions that we analysed could be performed by a linear readout.
Although high dimensionality is not strictly necessary for generating rich dynamics and performing complex tasks, it is known to greatly simplify the design of local neural circuits. It is often necessary to expand the dimensionality of the neuronal representations of the external sensory input and the internal state. In recent models, the dimensionality of the neuronal representations is expanded by mixing in a nonlinear way the different sources of information in a population of randomly connected neurons. The resulting neuronal representations are high-dimensional like those observed in PFC, and consistent with high dimensionality, the neurons show mixed selectivity which is diverse across time (that is, in different epochs of the trials) and space (that is, across different neurons). Random connectivity in small brain regions has been suggested on the basis of anatomical reconstructions and recently observed in the connections from the olfactory bulb to the olfactory cortex.
We showed that the recorded mixed selectivity can be useful for the activity to be linearly read out. It is legitimate to ask whether these considerations would still be valid if we consider more complex nonlinear readouts. For example, some of the transformations which increase the dimensionality of the neural representations could be implemented at the level of individual neurons by exploiting dendritic nonlinearities. Our results do not exclude the functional importance of such dendritic processes. They do, however, tend to argue against a scenario where all important nonlinear transformations are carried out at the level of single neurons, a situation where dimensionality expansion could happen in a ‘hidden way’, and the observable representations provided by the neuronal firing rates could therefore be low-dimensional.
Finally, the particular form of redundancy inherited from high-dimensional representations allows neural circuits to flexibly and quickly adapt to execute new tasks, just as it allows them to implement arbitrary binary classifications by modifying the weights of a readout neuron. We showed an example of this flexibility by training a simulated neuron to perform a new virtual task based on the recorded activity. High dimensionality might therefore be at the basis of the mechanisms underlying the remarkable adaptability of the neural coding observed in the PFC and, as such, be an important element to answer fundamental questions that try to map cognitive to neurophysiological functions.
In conclusion, the measured dimensionality of the neural representations in PFC is high, and errors follow a collapse in dimensionality. This provides us with a motivation to shift the focus of attention from pure selectivity neurons, which are easily interpretable, to the widely observed but rarely analysed mixed selectivity neurons, especially in the complex task designs that are becoming progressively more accessible to investigation.
[end of paraphrase]
Return to — Prefrontal Cortex