Elsevier

Neurocomputing

Volumes 38–40, June 2001, Pages 573-579
Neurocomputing

Cortex–basal ganglia interaction and attractor states

https://doi.org/10.1016/S0925-2312(01)00413-1Get rights and content

Abstract

We propose a set of hypotheses about how the basal ganglia contribute to information processing in cortical networks and how the cortex and basal ganglia interact during learning and behavior. We introduce a computational model on the level of system of networks. We suggest that the basal ganglia control cortical activity by pushing a local cortical network into a new attractor state, thereby selecting certain attractors over others. The ideas of temporal difference learning and convergence of corticostriatal fibers from multiple cortical areas within the striatum are combined in a modular learning system capable of acquiring behavior with sequential structure.

Introduction

The basal ganglia are large subcortical nuclei that are highly developed in primates and that are strongly interconnected with the neocortex. In humans, these nuclei function in relation to motor and psychomotor control, as indicated by their involvement in neurologic and neuropsychiatric disorders such as Parkinson's disease and obsessive–compulsive disorder. What the functions of the basal ganglia are under normal circumstances remains unknown, but increasing evidence across a range of species from bird to human suggests that the basal ganglia may function in sensory-motor learning and in the acquisition of habits and behavioral repertoires [5]. In this paper we present the initial phase of a model of interactions between the basal ganglia and neocortex suggesting that the states of local networks in the neocortex can be set and shifted by activity in cortico–basal ganglia circuits.

Characteristics of their anatomy and physiology suggest that the basal ganglia influence the executive functions of the frontal lobes. The basal ganglia, mainly via their main input nucleus, the striatum, receive inputs from nearly all of the cerebral cortex and project back, via multi-synaptic links involving the thalamus, mainly to frontal cortex. These connections are often described as cortico–basal ganglia loops [4]. Some of these loops have the interesting property of appearing relatively closed. Thus areas of the frontal cortex project to the part of the basal ganglia system that in turn projects via the thalamus to the same frontal cortical areas. Other loops, namely, those arising from cortex not in the frontal lobes, are not closed (but instead take the form of, for example, parietal cortex–basal ganglia–thalamus–frontal cortex pathways). It appears that not all of the cortical areas in such non-closed loops are themselves directly interconnected. Because the pathways are multi-synaptic, the details of these loops have been difficult to determine experimentally. Even as analyzed with new methods, however, the amount of basal ganglia outflow to non-frontal cortex appears to be relatively restricted.

These considerations suggest that a major function of the basal ganglia is to influence the frontal cortex based on inputs gathered from a much larger sample of cortical areas. A second notable property of cortico–basal ganglia connectivity is that there is extensive remapping within any one functionally defined loop [6]. For example, somatotopically corresponding sites in different areas of somatosensory and motor cortex send projections to the striatum that precisely overlap each other (convergence) but send these projections to multiple focal sites in the striatum (divergence). In the outflow from the striatum (to the pallidum), information from such input matched but dispersed striatal sites can converge again, suggesting a parallel with divergent–reconvergent adaptive mixture of expert models [6]. Thus it may be useful to think of functional units of cortico–basal ganglia processing, each carrying out remapping functions on its particular cortical inputs.

A third interesting feature of cortico–basal ganglia loops is that there are three output modes for these loops, the “direct pathway” with two GABAergic connections in a row producing disinhibition of the thalamus, an “indirect pathway” with two GABAergic links followed by a glutamatergic link increasing inhibition of the thalamus, and a “striosomal pathway” which sends output to the dopaminergic substantia nigra compacta. Winner-take-all competition is thought to occur on the part of opposing direct and indirect pathways, serving to produce “decisions” sent on to the executive areas of the frontal cortex. The striosomal pathway may modulate the dopamine system, discussed below. No complete neuron-for-neuron mapping of any one loop has been achieved experimentally, and it is possible that some cortical inputs may engage only one mode.

There is strong experimental evidence suggesting that these cortico–basal ganglia loops are modifiable. For example, both LTP and LTD have been demonstrated in corticostriatal pathways [3]. A principal modulator of this plasticity is the dopaminergic nigrostriatal system, which is thought to provide a reinforcement-based teaching signal to the striatum [9]. Most models of basal ganglia plasticity and their functions in learning and memory emphasize this dopaminergic input (e.g. [2]), and to varying degrees the interaction between direct and indirect pathways (e.g. [1]). In the model we present here, we take into account the dopamine-based reinforcement signal as a biasing mechanism, but place our emphasis on what the basal ganglia might do to influence dynamically neural activity in the neocortex. The core of the model is the suggestion that the states of activity in any cortical region can be modeled as attractor states, and that a function of basal ganglia input is to shift these attractor states dynamically based on “decisions” made in the basal ganglia based on plasticity in cortico–striatal synapses.

Section snippets

Model

In this early stage of the model, we have chosen to strip it down to the minimal elements necessary for learning to occur, rather than integrating detailed anatomical knowledge.

Our hypotheses are as follows:

. Schematic diagram of attractor state hypothesis, showing multiple attractor states in a cortical network and shifts from one state to another (horizontal arrow) resulting from basal ganglia inputs. Note that the direct (right arrow) and indirect (left arrow) pathways do not necessarily

For further reading

[7], [10]

Acknowledgments

This work was supported by Javits award NIH NS25529 to AMG.

Mikael Djurfeldt has a M.Sc. in Engineering Physics from KTH, Stockholm. He is pursuing a Ph.D. at the SANS group at the Department of Numerical Analysis and Computing Science, KTH, and at the Graybiel laboratory, Dept. of Brain and Cognitive Sciences, MIT. His main research interest concerns information processing in the basal ganglia, and, more specifically, the interplay between the cerebral cortex of the frontal lobe and the striatum.

References (10)

There are more references available in the full text version of this article.

Cited by (21)

  • Competitive and cooperative interactions between medial temporal and striatal learning systems

    2020, Neuropsychologia
    Citation Excerpt :

    These studies suggest that the hippocampus influences the PFC during declarative learning, possibly depriving the nondeclarative system of PFC processing resources. Another theory proposes that the striatum also controls PFC activity as learning progresses from forming simple associations to automaticity (Ashby and Maddox, 2011; Djurfeldt et al., 2001). This theory is supported by intracellular recordings from primates, where the learning-related change in firing rate was faster for striatal than for PFC neurons, suggesting that the striatum and its connections “train” the PFC during initial learning (Pasupathy and Miller, 2014).

  • Increases in functional connectivity between prefrontal cortex and striatum during category learning

    2014, Neuron
    Citation Excerpt :

    Support for this comes from human imaging studies showing that both the PFC and STR are engaged during category learning (Reber et al., 1998; Seger et al., 2000; Vogels et al., 2002). Also, computational and neurophysiological studies suggest more rapid changes in the STR than PFC during learning, as if the BG was “training” the cortex (Djurfeldt et al., 2001; Pasupathy and Miller, 2005). We recently provided more direct support in monkeys trained to learn new categories (Antzoulatos and Miller, 2011).

  • The basal ganglia: Learning new tricks and loving it

    2005, Current Opinion in Neurobiology
View all citing articles on Scopus

Mikael Djurfeldt has a M.Sc. in Engineering Physics from KTH, Stockholm. He is pursuing a Ph.D. at the SANS group at the Department of Numerical Analysis and Computing Science, KTH, and at the Graybiel laboratory, Dept. of Brain and Cognitive Sciences, MIT. His main research interest concerns information processing in the basal ganglia, and, more specifically, the interplay between the cerebral cortex of the frontal lobe and the striatum.

Örjan Ekeberg is an Associate Professor in Computer Science at the Royal Institute of Technology in Stockholm since 1994. His research interests ranges from artificial to biological neural networks with an emphasis on integrating the two lines of research. Current work focuses on the development of techniques for integrated neuro-mechanical simulations.

Ann Graybiel is the Walter A. Rosenblith Professor of Neuroscience at the Massachusetts Institute of Technology. Her research focuses on the functions of the basal ganglia and cortico–basal ganglia loops. She and her colleagues use ensemble neuronal recording methods and gene-based methods in rodents and primates to analyze cortico–basal ganglia circuits and their impact on behavior.

View full text