Cholinergic suppression of transmission may allow combined associative memory function and self-organization in the neocortex

Michael E. Hasselmo, Milos CekicNeurIPS 1995

Overview

This paper proposes a fascinating biological mechanism for how the brain manages two fundamentally different types of learning simultaneously. The authors suggest that the neuromodulator acetylcholine acts as a biological "mode switch" in the cortex. By selectively suppressing feedback synapses during learning while leaving feedforward connections intact, the brain can build new representations (self-organization) and link them together (associative memory) without the two processes catastrophically interfering with each other.

The Tension: Self-Organization vs. Associative Memory

In neural network theory, there is a well-known tension between two primary modes of learning:

  • Self-organization (Unsupervised Learning): This requires a neuron's activity to be driven by its own modifiable synapses so it can adapt to incoming data and become a "feature detector."
  • Associative Memory (Supervised Learning): This typically requires a neuron's activity to be driven by an external "teacher" or afferent input, temporarily ignoring its current synaptic state to store a new target pattern.

If a standard recurrent network attempts to do both simultaneously, the existing feedback connections (the "internal chatter") tend to overwhelm the new incoming information. This interference makes it incredibly difficult to learn complex, non-linear mappings—such as the classic XOR problem—using biologically plausible local learning rules. The central question of this paper is: How does the neocortex successfully balance these two conflicting modes?

A Dual-Pronged Methodology

To answer this, the authors employed a combination of biological "wetware" experiments and computational modeling.

Neural network architecture demonstrating self-organizing feedforward and associative feedback modes.
  1. Computational Modeling: They built a two-region neural network utilizing a Hebbian learning rule. The key innovation is an activation rule containing a suppression variable ($c$). During the "learning phase," $c$ is high, which mathematically silences the feedback connections. During the "recall phase," $c$ drops to zero, allowing the network to function as a fully recurrent system.
  2. Brain Slice Physiology: To ground this model in biology, they used rat somatosensory neocortex slices to test the effects of carbachol, a drug that mimics acetylcholine. They specifically compared synaptic transmission in Layer I (which primarily receives feedback from other cortical areas) and Layer IV (which primarily receives feedforward input from the thalamus).

Key Insights: Acetylcholine as a Gatekeeper

The most significant insight from this work is that acetylcholine (ACh) provides a physiological mechanism for "clamping" the network into a learning state.

By suppressing feedback, ACh allows the feedforward synapses to self-organize into feature detectors based purely on sensory input. Meanwhile, the suppressed feedback synapses "observe" this activity and learn the association (effectively the transpose of the feedforward connectivity).

Crucially, this cholinergic suppression is not global but laminar-specific. It heavily targets the feedback-rich Layer I but spares the feedforward-heavy Layer IV. This suggests the brain possesses a built-in architectural mechanism for separating "new data" from "old memories" during the encoding process.

Results and Performance

Physiological Confirmation

In the presence of carbachol, synaptic potentials in Layer I (feedback) were significantly suppressed, while potentials in Layer IV (feedforward) showed little to no effect. This successfully confirmed the biological basis for the computational model's "mode switching."

Electrophysiological recordings showing suppression of feedback synaptic potentials in Layer I by carbachol.

Computational Success

The model successfully learned the XOR problem—a benchmark that is famously not linearly separable and typically requires biologically implausible Backpropagation to solve in standard networks.

Using purely local Hebbian rules, the four units in the model's second region became selective for the four specific input patterns of XOR during the learning phase. Once learning was complete and the mathematical "suppression" was removed, the network could be presented with just the "stimulus" part of the pattern. The activity would flow forward to the hidden units and then back through the newly associated feedback connections to correctly "fill in" the missing "response" units.

Limitations and Future Work

While foundational, the paper notes a few limitations:

  • Simplistic Dynamics: The model uses a simplified version of inhibition and assumes long-range connections are purely excitatory, whereas real cortical inhibition is highly nuanced.
  • Binary States: The model handles the discrete (on/off) inputs of the XOR problem well. However, the limits of "partial suppression" in noisier, continuous real-world environments remain to be fully explored.
  • Scaling: The paper demonstrates convergence on small-scale logical benchmarks. Future work is needed to determine whether this laminar-specific suppression scales effectively to deep hierarchies with many interacting cortical regions.

Summary updated

BibTeX
@inproceedings{hasselmo1995cholinergic,
  author = {Michael E. Hasselmo and Milos Cekic},
  title = {{Cholinergic suppression of transmission may allow combined associative memory function and self-organization in the neocortex}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1995},
}
Figure