The Effect of Catecholamines on Performance: From Unit to System Behavior

David Servan-Schreiber, Harry Printz, Jonathan D. CohenNeurIPS 1989 · OpenReview

Introduction

A long-standing paradox in neuroscience surrounds the action of catecholamines—neurotransmitters like dopamine and norepinephrine associated with arousal and stress. At the level of individual neurons, these chemicals increase the "gain" or responsivity of cells to both excitatory and inhibitory inputs. However, from a Signal Detection Theory (SDT) perspective, simply increasing the slope of a neuron’s response function does not inherently improve its ability to distinguish a signal from background noise.

Despite this theoretical limitation at the single-cell level, behavioral data clearly shows that drugs increasing catecholamine levels (such as amphetamines or methylphenidate) significantly improve a human subject's ability to detect faint signals in noisy environments. This paper bridges this gap, demonstrating how changes in individual cell responsivity translate into improved system-wide performance through a neural network "chain effect."

The Paradox of Single-Neuron Gain

To understand the mechanism, researchers first modeled a single neuron using a logistic activation function: $f_G(x) = 1 / (1 + e^{-(Gx+B)})$, where $G$ represents the gain modulated by catecholamines.

Logistic activation functions demonstrating different gain levels

Through mathematical analysis, the authors established the Constant Optimal Performance Theorem. This theorem proves that for a single unit, increasing the gain ($G$) does not improve signal detection. While a steeper activation curve might change the optimal threshold for firing, the inherent overlap between the signal and noise distributions remains mathematically identical. If the brain consisted of isolated, single-layer sensors, stimulants would offer no perceptual benefit.

The Power of the Chain Effect

The breakthrough in this research comes from analyzing a multi-layer system—a "chain" of units where the output of one serves as the input to the next, with noise injected at each processing step.

When gain is increased in an early layer of a network, it pushes the outputs closer to the extremes (0 or 1), effectively widening the separation between the means of the signal and noise distributions. When subsequent noise is added in the next layer, this wider separation makes the signal much harder to degrade. Crucially, this improvement only occurs because noise is present between layers of processing. In a perfectly noise-free brain, the catecholamine-induced gain would provide no systemic advantage.

Computer Simulation and Key Results

To test this chain effect against empirical data, the researchers built a recurrent three-layer neural network (comprising 12 input, 30 hidden, and 10 output units). They trained this network using backpropagation to perform the Continuous Performance Test (CPT), a standard behavioral task requiring subjects to identify repeating letters in a sequence.

To simulate the effect of a central nervous system (CNS) stimulant, the researchers simply increased the gain parameter ($G$) of the units in the network from 1.0 to 1.1. They then compared the network's performance to existing clinical data of human subjects performing the CPT under the influence of methylphenidate (Ritalin).

Neural network architecture and simulated vs. empirical CPT performance

The simulation results were strikingly aligned with human behavioral data:

  • Human Trials: Clinical data showed that methylphenidate dropped the "miss" rate (failing to see a target) from 11.7% to 5.5%, while false alarms (seeing a target that isn't there) remained nearly flat.
  • Model Simulation: By increasing the network's gain, the model's misses dropped from 13.0% to 6.6%, with false alarms remaining perfectly stable (0.75% to 0.78%).

This pattern—reducing misses without inflating false alarms—is a specific hallmark of human psychopharmacology that the model successfully replicated. Furthermore, the chain effect proved robust whether the gain was increased in the hidden layer, the output layer, or both.

Limitations and Future Work

While the model elegantly resolves the single-neuron paradox, it relies on certain simplifications. The simulation assumes Gaussian noise with a zero mean that is independent of the gain, whereas real biological noise may be multiplicative or far more complex. Additionally, the study utilizes a relatively small, three-layer recurrent network. While sufficient for modeling the CPT task, future research is needed to determine how these findings scale to more complex cognitive functions like language processing or high-level reasoning.

Finally, the authors focused exclusively on "gain" as the primary mechanism for catecholamine action. In biological systems, these neurotransmitters likely have multifaceted effects, such as altering the temporal dynamics of neurons or modulating synaptic plasticity, which offer exciting avenues for future modeling.

Summary updated

Abstract

At the level of individual neurons. catecholamine release increases the responsivity of cells to excitatory and inhibitory inputs. We present a model of catecholamine effects in a network of neural-like elements. We argue that changes in the responsivity of individual elements do not affect their ability to detect a signal and ignore noise. However. the same changes in cell responsivity in a network of such elements do improve the signal detection performance of the network as a whole. We show how this result can be used in a computer simulation of behavior to account for the effect of eNS stimulants on the signal detection performance of human subjects.
BibTeX
@inproceedings{servanschreiber1989effect,
  author = {David Servan-Schreiber and Harry Printz and Jonathan D. Cohen},
  title = {{The Effect of Catecholamines on Performance: From Unit to System Behavior}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1989},
  url = {https://openreview.net/forum?id=ByNMgOZ_-S},
}

Loading related content...

Figure