Motor Learning Mechanism on the Neuron Scale

Peilei Liu, Ting Wang

TL;DR

This paper proposes a novel computational model of motor learning operating at the neuron scale, bridging the gap between molecular evidence and existing computational models. It posits that fundamental learning mechanisms like dendritic competition and synaptic plasticity can emerge from local rules governing neuron and synapse behavior, offering a unified framework for understanding motor control and sensory-motor integration.

What Problem Does This Paper Address?

Despite significant progress in understanding motor learning, a gap persists between existing computational models and physiological data. Current models often operate at a higher "module scale" (e.g., internal forward models) or are difficult to biologically implement (e.g., CMAC). This paper aims to address this by developing a quantitative model of motor learning at the neuron scale, inspired by self-organization principles. The authors seek to explain how complex motor behaviors and sensory-motor integration arise from the local interactions of individual neurons and synapses, providing a more biologically plausible explanation for phenomena like conditional reflex and "grandmother cell" coding.

Methods

The authors propose a neuron model (N1 and N2) and a synapse model (S1 and S2) that operate on local rules.

  • Neuron Model (N1, N2):
    • N1 describes the firing frequency ($f$) and hyperpolarized degree of a neuron based on its own state and synaptic inputs. It incorporates a "hormone factor" ($h$) and a "fatigue" term ($b_i = c_3 - a_i$).
  • N2 models the dynamics of the "fatigue" factor ($a_i$) over time, distinguishing between depolarized and hyperpolarized states. This fatigue mechanism is crucial for explaining channel recovery and the distinct behavior of motor neurons compared to sensory neurons.
  • Synapse Model (S1, S2):
    • S1 describes Long-Term Potentiation (LTP), where synaptic strength ($w$) increases based on presynaptic activity and a "hormone factor" ($h$). It also models synaptic decay rate ($r$).
  • S2 describes Long-Term Depression (LTD), where synaptic strength decreases. This is presented as a devaluation of LTP and is linked to the lack of back-propagating action potentials (BAPs) in motor neurons, contributing to channel fatigue.
  • Dendritic Competition: A model of dendritic competition is presented, where retrograde messengers released by dendritic spikes transiently depress other dendrites on the same axonal branch. This leads to a "winner-take-all" mechanism, where the dendrite with the highest firing frequency strengthens its synapse, while others decay. This is proposed as a key mechanism for conditional reflex and lateral inhibition.

The core idea is that by defining these local rules for neurons and synapses, a complex neural network and its learning mechanisms can emerge automatically.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and proposes a framework rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Synaptic Strength as Probability: Neuron firing frequency and synaptic strength are re-interpreted as probabilistic estimates of features based on stimulus history.
  • Lateral Inhibition as Statistical: Lateral inhibition is also viewed as having statistical implications, similar to how contradictory answers reduce reliability.
  • Dendritic Competition as Foundation: Dendritic competition via retrograde messengers is identified as the basis for conditional reflex and "grandmother cell" coding, crucial for motor learning and sensory-motor integration.
  • Motor vs. Sensory Neuron Differences: Motor neurons exhibit channel fatigue when depolarized and recovery when hyperpolarized, leading to a different computational model than sensory neurons, which rely more on temporal summarization of postsynaptic signals. LTD in motor neurons is linked to this fatigue.
  • Emergent Network Properties: The proposed local rules for neurons and synapses, combined with dendritic competition, can lead to the automatic emergence of a functional neural network and learning mechanisms.
  • Sandglass Structure for Integration: Sensory-motor integration is conceptualized as a "sandglass" structure, with inward trees for sensory input and outward trees for motor commands, potentially mediated by Purkinje cells (PCs) or Pyramid cells.
  • Conditional Reflex as Learning Foundation: The paper posits that the competition between neurons firing simultaneously, leading to strengthened connections to motor units, forms the physiological basis of conditional reflex and motor learning.

Key Results

The paper presents a theoretical framework and a set of models rather than quantitative experimental results. The "results" are primarily conceptual:

  • A unified model for excitory and inhibitory motor neurons is proposed, differing mainly in details.
  • The model explains how dendritic competition leads to a "winner-take-all" scenario, which is crucial for learning.
  • It provides a biological explanation for conditional reflex as a consequence of this competition and synaptic plasticity.
  • The model suggests that sensory-motor integration can be achieved through "sandglass" structures.
  • It differentiates the computational roles of motor neurons (fatigue-based) and sensory neurons (temporal summarization-based).

Limitations and Future Work

  • Lack of Empirical Validation: The paper is primarily theoretical and does not present empirical data or simulations to validate the proposed models against experimental findings.
  • Parameter Tuning: The models involve numerous constants ($c_i$) whose specific values and biological relevance are not fully detailed, suggesting that parameter tuning would be a significant aspect of future work.
  • Generalization of "Grandmother Cell": While the "grandmother cell" concept is mentioned, its precise implementation and generalization across different motor tasks require further exploration.
  • Formation of Initial Pathways: The paper touches upon unsupervised learning for the formation of initial sensory-motor pathways but suggests this as a direction for future research.
  • Role of Hormones: The influence of hormones (like dopamine) is mentioned as a meta-mechanism for adjusting firing frequency and synaptic growth, but a detailed model of this interaction is not fully developed.

Future work would likely involve implementing these models in simulations, testing them against experimental data, and further refining the parameterization and biological interpretations.

Summary updated

Abstract

Based on existing data, we wish to put forward a biological model of motor system on the neuron scale. Then we indicate its implications in statistics and learning. Specifically, neuron firing frequency and synaptic strength are probability estimates in essence. And the lateral inhibition also has statistical implications. From the standpoint of learning, dendritic competition through retrograde messengers is the foundation of conditional reflex and grandmother cell coding. And they are the kernel mechanisms of motor learning and sensory motor integration respectively. Finally, we compare motor system with sensory system. In short, we would like to bridge the gap between molecule evidences and computational models.
BibTeX
@misc{liu2014motor,
  author = {Peilei Liu and Ting Wang},
  title = {{Motor Learning Mechanism on the Neuron Scale}},
  year = {2014},
  eprint = {1407.7027},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1407.7027},
}

Loading related content...

Figure