The dimensionality of neural representations for control

David Badre, Apoorva Bhandari, Haley Keglovits, Atsushi Kikumoto, David Badre, Apoorva Bhandari, Haley Keglovits, Atsushi Kikumoto

Abstract

Cognitive control allows us to think and behave flexibly based on our context and goals. At the heart of theories of cognitive control is a control representation that enables the same input to produce different outputs contingent on contextual factors. In this review, we focus on an important property of the control representation's neural code: its representational dimensionality. Dimensionality of a neural representation balances a basic separability/generalizability trade-off in neural computation. We will discuss the implications of this trade-off for cognitive control. We will then briefly review current neuroscience findings regarding the dimensionality of control representations in the brain, particularly the prefrontal cortex. We conclude by highlighting open questions and crucial directions for future research.

Keywords: cognitive control; executive function; frontal lobes; neural computation; neural representation.

Conflict of interest statement

Conflict of Interest statement for: The dimensionality of neural representations for control ‘Declaration of interest: none’.

Figures

Figure 1.
Figure 1.
Schematic illustration of representational dimensionality and its computational properties. Each panel plots the response of a toy population of three neurons to stimuli that vary in shape (square or circle) and color (red or black). Axes represent the firing rates of single neurons and collectively define a multi-dimensional firing rate space for the population. Each point within this space represents the population response (i.e., activity pattern) for a given input (identified by colored shapes). Distance between points reflects how distinct responses are, and the jittered cloud of points reflect the trial-by-trial variability in responses to a given input. (a-c) show a low dimensional representation. Though the population is 3-dimensional (i.e., has three neurons), the representation defined by the population responses to the four stimuli defines a lower 2-dimensional plane (traced by solid black lines). A linear ‘readout’ of this representation is implemented by a ‘decision hyperplane’ (yellow) that divides the space into different classes, such as color and shape. The readout can be visualized by projecting (highlighted for red circles) the responses into a readout subspace. (a) In the shape subspace (peach), square and circle stimuli produce distinct, well-separated responses that generalize over different colors. (b) The color subspace (grey) separates red and black but generalizes over shape. (c) illustrates that it is impossible to linearly read out integrated classes (like red-square or black-circle vs black-square or red-circle) in this low-dimensional representation. A non-linear decision surface is required. This problem can be solved by a high-dimensional representation where the response patterns span 3 dimensions (d-f). As before, shape (d) and color (e) information can be linearly read out, though with poorer generalization along the irrelevant dimensions (reflected in the distance between clouds in the irrelevant dimension). Importantly, classes based on color-shape conjunctions (f) can also be linearly read out. Thus, high-dimensional representations are more expressive, making a wider variety of classes linearly separable. A population with a diversity of non-linear mixed selective neurons will have a higher representational dimensionality.
Figure 2.
Figure 2.
(a-c) Illustrations of control representations of different dimensionality across modulatory and transmissive architectures for cognitive flexibility. (a) A modulatory model of cognitive control based on the widely recognized Stroop model from Cohen et al. [1]. Localist units represent patterns of activity of a neural population. Input layers with the word and ink color feedforward to a hidden layer which feeds forward to a response layer at the top. A task cue input feeds forward to a low dimensional context layer. This is the control representation that biases the task-relevant response pathway to enable flexibility. (b) A transmissive architecture that achieves flexibility with a high dimensional hidden layer that mixes all inputs into separate pathways. Thus, the hidden layer is an expressive control representation that makes any combination of cue and stimuli available for direct readout by the response layer. (c) A modulatory model with a high dimensional context layer. This expressive control representation can use any combination of task cue and stimuli as a contextual input to bias the route from stimulus to response. (d)-(e) Illustrations of the tradeoffs of dimensionality for control. (d) A simplified representation of the architecture in 2a without the context layer and with an added output, following [32]. The combination of two input types and two ouput types makes four “tasks”. Arrows show pathways from input to output with active unites highlighted and distinguished by fill pattern. A low dimensional control representation can link an input to both verbal and button press responses. This is useful for fast learning; a new task involving auditory input can leverage this low dimensional representation to link to all available outputs. However, if two tasks are performed at once, there is cross-talk interference (red paths) (e) A higher dimensional control representation in the hidden layer, akin to 1b, separates the inputs across the tasks. This allows multitasking without interference, but a new task involving auditory input must learn entirely new pathways.

References

    1. Cohen JD et al. (1990) On the control of automatic processes: A parallel distributed processing account of the Stroop effect. Psychological Review 97, 332–361.
    1. Cohen JD (2017) Cognitive control: core constructs and current considerations In The Wiley Handbook of Cognitive Control (Egner T. ed), pp. 1–28, Wiley.
    1. Fusi S et al. (2016) Why neurons mix: high dimensionality for higher cognition. Curr Opin Neurobiol 37, 66–74.

    2. This review discusses recent theoretical developments on representational dimensionality and empirical findings in animals. They introduce the compelling computational link between mixed selectivity of neurons and dimensionality, and its implications for the separability/generalizability trade-off.

    1. Frank MJ and Badre D (2012) Mechanisms of hierarchical reinforcement learning in corticostriatal circuits 1: computational analysis. Cereb Cortex 22 (3), 509–26.
    1. Miller EK and Cohen JD (2001) An integrative theory of prefrontal cortex function. Annual Review of Neuroscience 24, 167–202.
    1. O’Reilly RC and Frank MJ (2006) Making working memory work: a computational model of learning in the prefrontal cortex and basal ganglia. Neural Comput 18 (2), 283–328.
    1. Duncan J (2013) The structure of cognition: attentional episodes in mind and brain. Neuron 80 (1), 35–50.
    1. Hommel B (2004) Event files: feature binding in and across perception and action. Trends Cogn Sci 8 (11), 494–500.
    1. Schumacher EH and Hazeltine E (2016) Hierarchical task representation: Task files and response selection. Current Directions in Psychological Science 25 (6), 449–54.
    1. Kikumoto A and Mayr U (in press) Conjunctive representations that integrate stimuli, responses, and rules are critical for action selection. Proc Natl Acad Sci U S A.

    2. This EEG study in humans shows that action selection engages a conjunctive representation that integrates multiple task relevant features, consistent with an event (task) file. It highlights the significance of a transmissive control representation in action selection.

    1. Rainer G et al. (1998) Selective representation of relevant information by neurons in the primate prefrontal cortex. Nature 393 (6685), 577–9.
    1. Stokes MG et al. (2013) Dynamic coding for cognitive control in prefrontal cortex. Neuron 78 (2), 364–75.
    1. Pischedda D et al. (2017) Neural Representations of Hierarchical Rule Sets: The Human Control System Represents Rules Irrespective of the Hierarchical Level to Which They Belong. J Neurosci 37 (50), 12281–12296.
    1. Woolgar A et al. (2016) Coding of Visual, Auditory, Rule, and Response Information in the Brain: 10 Years of Multivoxel Pattern Analysis. J Cogn Neurosci 28 (10), 1433–54.
    1. Rigotti M et al. (2013) The importance of mixed selectivity in complex cognitive tasks. Nature 497 (7451), 585–90.

    2. The first experimental study to directly estimate dimensionality of control representations in macaque PFC during a working memory task. The authors measured maximal dimensionality for the task, which collapsed when the monkeys made errors. The study defines and validates methods for estimating dimensionality from neural recordings.

    1. Maass W (2016) Searching for principles of brain computation. Curr Opin Behav Sci 11, 81–92.

    2. A broad review of computational models and empirical research across domains. Four key principles governing the operation of neural circuits are identified. A key theme developed is the importance of a diversity of computational elements in representations, and the stable emergence of computational function from complex, highly dynamic and varying neural circuits.

    1. Panzeri S et al. (2017) Cracking the neural code for sensory perception by combining statistics, intervention, and behavior. Neuron 93 (3), 491–507.
    1. Cohen U et al. (2020) Separability and geometry of object manifolds in deep neural networks. Nat Commun 11 (1), 746.
    1. Ganguli S and Sompolinsky H (2012) Compressed sensing, sparsity, and dimensionality in neuronal information processing and data analysis. Annu Rev Neurosci 35, 485–508.
    1. Boser BE et al., A training algorithm for optimal margin classifiers, Proceedings of the fifth annual workshop on Computational learning theory, 1992, pp. 144–152.
    1. Jun JK et al. (2010) Heterogenous population coding of a short-term memory and decision task. J Neurosci 30 (3), 916–29.
    1. Mante V et al. (2013) Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature 503 (7474), 78–84.
    1. Bernardi S et al. (2019) The geometry of abstraction in hippocampus and prefrontal cortex. bioRxiv, 1–22.

    2. This study investigates the neural representations of abstract control representations during reversal learning, when hidden contextual information is used across trials. Representations in the ACC, PFC, and hippocampus generalize to multiple contexts, a form of abstraction. Further, across epochs of a trial, the dimensionality and degree of abstraction of the representation changes between and within trials.

    1. Stringer C et al. (2019) High-dimensional geometry of population responses in visual cortex. Nature 571 (7765), 361–365.

    2. The first study to provide a realistic estimate of representational dimensionality using a complex, naturalistic task space and simultaneous recording of over 10,000 neurons using two-photon calcium imaging in mice visual cortex. They find high-dimensional, but smooth representations, which would offer high separability with robustness to noise.

    1. Farrell M et al. (2019) Recurrent neural networks learn robust representations by dynamically balancing compression and expansion. bioRxiv, 564476.
    1. Badre D et al. (2010) Frontal cortex and the discovery of abstract action rules. Neuron 66 (2), 315–26.
    1. Collins AG and Frank MJ (2013) Cognitive control over learning: creating, clustering, and generalizing task-set structure. Psychol Rev 120 (1), 190–229.
    1. Duncan J et al. (2017) Complexity and compositionality in fluid intelligence. Proceedings of the National Academy of Sciences 114 (20), 5295–5299.
    1. Chatham CH and Badre D (2015) Multiple gates on working memory. Curr Opin Behav Sci 1, 23–31.
    1. Musslick S et al., Multitasking capability versus learning efficiency in neural network architectures, Cognitive Science Society, London, 2017, pp. 829–34.

    2. Using formal analysis and neural network simulations, this study shows that the use of shared low-dimensional representations trades off the ability to simultaneously process two tasks for improved learning efficiency of novel tasks.

    1. Musslick S and Cohen JD, A mechanistic account of constraints on control-dependent processing: Shared representation, conflict and persistence, 41st Annual Meeting of the Cognitive Science Society, 2019.
    1. Sagiv Y et al., Effiency of learning vs. processing: Towards a normative theory of multitasking, 40th Annual Meeting of the Cognitive Science Society, 2018.
    1. Bartolo R et al. (2020) Dimensionality, information and learning in prefrontal cortex. PLoS Comput Biol 16 (4), e1007514.
    1. Machens CK et al. (2010) Functional, but not anatomical, separation of “what” and “when” in prefrontal cortex. J Neurosci 30 (1), 350–60.
    1. Brincat SL et al. (2018) Gradual progression from sensory to task-related processing in cerebral cortex. Proc Natl Acad Sci U S A 115 (30), E7202–E7211.
    1. Parthasarathy A et al. (2017) Mixed selectivity morphs population codes in prefrontal cortex. Nat Neurosci 20 (12), 1770–1779.
    1. Parthasarathy A et al. (2019) Time-invariant working memory representations in the presence of code-morphing in the lateral prefrontal cortex. Nat Commun 10 (1), 4995.

    2. This study reports that a population of mixed-selective neurons drastically changes the geometry of firing patterns (i.e., morphing) against a distractor during a working memory task. Yet, it preserves a temporally stable subspace that allows readout of a memory encoded prior to the onset of a distractor.

    1. Tang C et al. (2019) Independent activity subspaces for working memory and motor preparation in the lateral prefrontal cortex. . bioRxiv, 756072.
    1. Murray JD et al. (2017) Stable population coding for working memory coexists with heterogeneous neural dynamics in prefrontal cortex. Proc Natl Acad Sci U S A 114 (2), 394–399.
    1. Broadbent DE (1977) Levels, hierarchies, and the locus of control. Quarterly Journal of Experimental Psychology 29, 181–201.
    1. Tang E et al. (2019) Effective learning is accompanied by high-dimensional and efficient representations of neural activity. Nat Neurosci 22 (6), 1000–1009.
    1. Mack ML et al. (2020) Ventromedial prefrontal cortex compression during concept learning. Nat Commun 11 (1), 46.
    1. Brody BA et al. (1987) Sequence of central nervous system myelination in human infancy. I. An autopsy study of myelination. J Neuropathol Exp Neurol 46 (3), 283–301.
    1. Ahlheim C and Love BC (2018) Estimating the functional dimensionality of neural representations. Neuroimage 179, 51–62.
    1. Diedrichsen J et al. (2013) A multivariate method to determine the dimensionality of neural representation from population activity. Neuroimage 76, 225–35.
    1. Kriegeskorte N and Kievit RA (2013) Representational geometry: integrating cognition, computation, and the brain. Trends Cogn Sci 17 (8), 401–12.
    1. Bhandari A et al. (2018) Just above Chance: Is It Harder to Decode Information from Human Prefrontal Cortex Blood Oxygenation Level-dependent Signals? J Cogn Neurosci, 1–26.
    1. Bhandari A et al., Measuring prefrontal representational geometry: fMRI adaptation vs pattern analysis, Cognitive Computational Neuroscience conference proceedings, Berlin, Germany, 2019.
    1. Rigotti M and Fusi S (2016) Estimating the dimensionality of neural responses with fMRI Repetition Suppression. arXiv preprint arXiv:1605.03952.
    1. Barron HC et al. (2016) Repetition suppression: a means to index neural representations using BOLD? Philos Trans R Soc Lond B Biol Sci 371 (1705).

Source: PubMed

Sponsors et collaborateurs

Les conditions médicales

Interventions en matière de drogue

3
S'abonner