A Computational Model of Prefrontal Cortex Function

Todd S. Braver, Jonathan D. Cohen, David Servan-SchreiberNeurIPS 1994 · OpenReview

This classic 1994 paper from the early days of connectionist modeling provides a bridge between neurobiology, computation, and clinical psychology. It offers a unified theory of how the prefrontal cortex (PFC) manages the "executive" functions of the brain.

The Dual Role of the Prefrontal Cortex

For decades, researchers viewed the prefrontal cortex as having two distinct roles: active memory (holding information "online" to guide behavior) and inhibition (stopping reflexive or inappropriate responses). However, it wasn't clear how these two functions interacted or if they were actually two sides of the same coin.

Furthermore, existing computational models often ignored the temporal dynamics of these processes—how long a stimulus lasts or how long a delay is. The authors sought to create a model that could account for both memory and inhibition within a realistic time-based framework, specifically to understand the cognitive deficits seen in schizophrenia.

Methods: Task and Model

The researchers developed a two-pronged approach, pairing a behavioral task with a continuous-time recurrent neural network.

1. The CPT-AX Task They used a variation of the Continuous Performance Test. Subjects see a sequence of letters and must press a "target" button only when an X is preceded by an A.

  • Memory Probe: The delay between the cue (A) and the probe (X) was manipulated from 750ms (short) to 5000ms (long).
  • Inhibition Probe: A-X trials were made very frequent (70%). This creates a "prepotent" tendency to hit the target button whenever an X appears, even if the cue wasn't an A (a B-X trial).

2. The Computational Model To simulate human performance, the authors designed a recurrent neural network consisting of three interconnected modules.

Network Architecture
  • Input Layer: Represents the stimuli (A, B, X, Y).
  • Memory Layer (PFC): Features self-excitatory connections to allow activity to persist (active memory) and lateral inhibition to create competition.
  • Output Layer: Determines the response (Target vs. Non-target).

Unlike many models that move in discrete "steps," this model uses differential equations to simulate processing over actual milliseconds, allowing it to be constrained by real-world reaction times. The model was trained on the 8 task conditions (A-X, A-Y, B-X, B-Y at short and long delays) and compared against behavioral data collected from 36 healthy human subjects.

Key Insights and Results

With only four free parameters, the model achieved a remarkably close fit to 24 different data points (accuracy and reaction times) from the human subjects. The paper argues that "inhibition" is actually a byproduct of "active memory." If the network cannot remember the context (the "A"), it cannot inhibit the reflexive response to the "X."

Dopamine as "Gain" and the Schizophrenia Prediction The authors simulate the effect of dopamine by adjusting a "gain" parameter ($\gamma$). High gain sharpens the signal-to-noise ratio, making representations robust. Low gain—hypothesized to occur in schizophrenia—causes representations to become "blurry" and decay quickly.

This leads to a striking double dissociation prediction for schizophrenic patients.

Model Predictions for Reduced Gain

As shown in the long delay condition, a "low gain" model fails significantly on B-X trials because it forgets the "B" and reflexively hits "Target" on the "X". Counterintuitively, it actually makes fewer errors on A-Y trials than the normal model. Because it forgets the "A," it isn't "primed" to expect a target, and thus doesn't mistakenly hit "Target" when the "Y" appears. This "superior" performance in a specific context is a rare and highly testable clinical prediction.

Limitations and Future Work

While groundbreaking, the model is highly abstracted. It uses only a few units to represent entire populations of neurons and does not account for the complex laminar structure of the cortex. Additionally, it uses a simple constant to account for the time it takes to physically execute a motor response, oversimplifying the motor system. Finally, at the time of publication, the specific "better than normal" performance on A-Y trials for schizophrenics remained a novel prediction requiring further empirical validation in clinical populations.

Summary updated

Abstract

Accumulating data from neurophysiology and neuropsychology have suggested two information processing roles for prefrontal cor(cid:173) tex (PFC): 1) short-term active memory; and 2) inhibition. We present a new behavioral task and a computational model which were developed in parallel. The task was developed to probe both of these prefrontal functions simultaneously, and produces a rich set of behavioral data that act as constraints on the model. The model is implemented in continuous-time, thus providing a natural framework in which to study the temporal dynamics of processing in the task. We show how the model can be used to examine the be(cid:173) havioral consequences of neuromodulation in PFC . Specifically, we use the model to make novel and testable predictions regarding the behavioral performance of schizophrenics, who are hypothesized to suffer from reduced dopaminergic tone in this brain area.
BibTeX
@inproceedings{braver1994computational,
  author = {Todd S. Braver and Jonathan D. Cohen and David Servan-Schreiber},
  title = {{A Computational Model of Prefrontal Cortex Function}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1994},
  url = {https://openreview.net/forum?id=ByVmjO-_-r},
}
Figure