Elsevier

Neurocomputing

Volumes 58–60, June 2004, Pages 761-767
Neurocomputing

A recurrent neural network model of rule-guided delayed tasks

https://doi.org/10.1016/j.neucom.2004.01.124Get rights and content

Abstract

We developed a recurrent neural network model of rule-guided behavior to simulate neural activity in rule-guided tasks. Our model was constructed using neural system identification (Neurosci. 47 (4) (1992) 853) and a fully recurrent neural network model was optimized to perform a rule-guided delayed task. The response properties of the hidden units were compared to those of neurons to examine the degree to which the model accounts for the experimental data (Exp. Brain Res. 126 (1999) 315). The temporal patterns of the hidden units were consistent with the physiological results. The close similarity between the behavior of the model units and biological neurons shows that the brain uses mechanisms like those of the model, and that ample mutual connections in the prefrontal cortex are the basis for promoting flexible learning.

Introduction

The prefrontal (PF) cortex has long been suspected to play an important role in cognitive control [3]. Several studies have examined the role of the PF cortex in processing information about stimuli, rewards, errors, and so on. Recently, much attention has been directed to ‘rules’ that guide goal-directed behavior. The flexibility of our behavior is caused mainly by our ability to abstract such rules from a circumstance and apply them to other situations. White and Wise [7] studied single-neuron activities in the PF cortex, while a monkey performed a rule-guided task according to two different rules: the ‘conditional rule’ and the ‘spatial rule’. The task required the monkey to maintain fixation of the target location, which consisted of four light spots. For the conditional rule, the shape of the visual cue indicated the correct target. For the spatial rule, the location of the visual cue indicated the correct target. The rule to be applied in each trial was indicated by the color of a fixation spot that was presented before the visual cue (Fig. 1). Between one-third and one-half of the PF neurons showed activity differences that could be attributed to the rule. There was no significant regional segregation. These data support the hypothesis that the PF cortex plays a role in guiding behavior according to previously learned rules. Other studies have yielded similar results [1], [2], [6].

We have already proposed a neural network model of rule-guided behavior and simulated other physiological results [4]. However, the interpretation of the simulation results was limited, because the percentage of rule-selective units was small, which was inconsistent with the physiological results. Moreover, the temporal patterns of the rule-selective units differed slightly from those of biological neurons, especially for the patterns of buildup activities.

In this study, we developed a neural network model of rule-guided behavior that could accommodate the results of other physiological studies. In addition, we also compared the model units with neurons selective for information other than rules. Our model was constructed using neural system identification [9] and a fully recurrent neural network model was optimized to perform a rule-guided delayed task.

Section snippets

Modeling methods

The PF cortex is connected with virtually all sensory neocortical and motor systems. Physiological results indicate that there is little difference in the distribution of neurons preferring different types of information in the PF area. Therefore, we assumed that the PF area is a single uniform layer. Our model contains three layers: input, hidden, and output layers (Fig. 2). The input layer has three input modules: the first represents ‘what’ information (object module), the second ‘where’

Results

In a typical simulated behavioral trial, the fixation color was input in the first step. After a delay period (six steps), object and location information were input. After a second delay (three steps), object and location information were input again. After an additional delay (three steps), the response gate was loaded and motor output was produced (see Fig. 1). The network was successful in learning the rule-guided task. The response properties of the hidden units were compared with those of

Discussion

We simulated the results of a physiological experiment [7] using a recurrent network model with a simple input-and-output relation for executing a rule-guided task. Our model could explain the function of the PF cortex under the assumption that the object module corresponds to the inferior temporal cortex, the location module to the posterior parietal cortex, and the output layer to the motor-related area. Of course, it is difficult to directly connect RTRL with the learning mechanisms of the

Acknowledgements

This research was supported by the Project on Neuroinformatics Research in Vision through Special Coordination Funds for Promoting Science and Technology from the Ministry of Education, Culture, Sports, Science and Technology, the Japanese Government, and the Ministry of Education, Science, Sports and Culture, Grant-in-Aid for JSPS Fellows 14002026.

References (9)

  • D. Zipser

    Identification models of the nervous system

    Neuroscience

    (1992)
  • W.F. Asaad et al.

    Task-specific neural activity in the primate prefrontal cortex

    J. Neurophysiol.

    (2000)
  • E. Hoshi et al.

    Task-dependent selectivity of movement-related neuronal activity in the primate prefrontal

    J. Neurophysiol.

    (1998)
  • E.K. Miller et al.

    An integrative theory of prefrontal cortex function

    Annu. Rev. Neursci.

    (2001)
There are more references available in the full text version of this article.

Cited by (0)

View full text