Neural Network Analysis of Distributed Representations of Dynamical Sensory-Motor Transformations in the Leech

NeurIPS 1989

TL;DR

This paper uses a recurrent neural network model to analyze the local bending reflex in leeches, demonstrating that this complex behavior is likely controlled by a distributed network of interneurons rather than a few specialized ones. The model successfully replicates experimental data, offering insights into how biological neural circuits can achieve robust sensory-motor transformations through distributed representations.

What Problem Does This Paper Address?

Understanding how biological neural circuits process sensory information and generate motor commands is a fundamental challenge in neuroscience. Specifically, researchers are interested in how complex behaviors, like the leech's local bending reflex, are orchestrated by networks of neurons. A key question is whether specific neurons are dedicated to particular aspects of a behavior (localist representation) or if the function emerges from the collective activity of many neurons with more general roles (distributed representation). Previous work had identified many neurons involved in the leech's local bending reflex, but their precise roles and how they interact to produce the coordinated movement remained unclear, particularly concerning the potential for distributed processing.

Methods

The authors employed a recurrent back-propagation algorithm to train a dynamical neural network model. This model was designed to mimic the known structure and physiological properties of the leech's local bending circuit. Key aspects of the model include:

  • Neuron Representation: Neurons were modeled as single electrical compartments with defined input resistance and time constants, capturing their dynamic electrical behavior.
  • Synaptic Connections: Both chemical and electrical synapses were incorporated. Chemical synapses included physiological delays implemented by "s-units" with varying time constants, reflecting the observed temporal dynamics of synaptic potentials.
  • Training: The network was trained to reproduce experimentally recorded input-output patterns of the local bending reflex. This involved stimulating model sensory neurons (P cells) and comparing the resulting motor neuron outputs to actual biological data.
  • Model-Data Comparison: After training, the model's interneurons were analyzed by simulating physiological experiments (recording input-output connections) to compare their properties to identified biological interneurons.

Datasets

The study utilized experimentally derived input-output patterns of the leech's local bending reflex for training and evaluation. These patterns were obtained from previous physiological recordings of sensory neuron (P cell) stimulation and subsequent motor neuron responses. The paper also references specific datasets for synaptic strengths and temporal properties derived from intracellular recordings of leech neurons.

Key Insights and Ideas

  • Distributed Representation is Key: The model strongly supports the idea that the local bending reflex is mediated by a distributed representation of sensory information across a population of interneurons, rather than by a few highly specialized neurons.
  • Emergent Roles: The training process allowed model interneurons to develop functional roles that, while not perfectly matching any single identified biological interneuron, collectively produced the correct behavior. This suggests that the observed diversity in interneuron properties might be a consequence of their participation in a distributed processing scheme.
  • Explaining Counterintuitive Aspects: The distributed model helps explain why some interneurons appear to have both appropriate and inappropriate effects on motor neurons. These seemingly contradictory roles can be reconciled when viewed as part of a larger population whose net effect is the desired behavior.
  • Model-Data Similarity: The qualitative similarity between the simulated and actual interneurons in terms of connectivity and temporal response properties validates the modeling approach and its ability to capture biological neural dynamics.

Key Results

  • The trained neural network model successfully reproduced the input-output function of the local bending reflex, closely matching the desired output patterns after 6,000-10,000 training epochs.
  • Model interneurons, like their biological counterparts, received substantial input from multiple sensory neurons and exerted significant influence on many motor neurons.
  • The temporal response properties of model interneurons showed agreement with actual leech interneurons, exhibiting fast, slow, and mixed temporal components, which arose from the relative strengths of different synaptic units.
  • The model demonstrated that a population of interneurons with distributed roles could mediate the local bending reflex without requiring interneurons specifically tuned to each form of bending (dorsal, ventral, lateral).

Limitations and Future Work

  • Incomplete Biological Data: The authors acknowledge that the identified local bending interneurons represent only a subset of the total interneurons involved in the reflex. The model's training algorithm did not produce exact matches between all model and actual interneurons, partly due to this incompleteness.
  • Future Directions: The paper suggests that more precise matches could be achieved by using two pools of model interneurons: one to represent identified neurons and another to represent unidentified ones. This latter pool could then generate testable physiological predictions about the connectivity of these unknown interneurons. The approach could also be extended to more complex biological systems with more realistic neuron models and synaptic connections.

Summary updated

BibTeX
@inproceedings{1989neural,
  title = {{Neural Network Analysis of Distributed Representations of Dynamical Sensory-Motor Transformations in the Leech}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1989},
}
Figure