Stability of the fittest: organizing learning through retroaxonal signals

Kenneth D. HarrisTrends 2008 · DOI

TL;DR

This paper proposes a novel framework for adult learning, suggesting that strengthening a neuron's output synapses stabilizes recent changes in its input synapses. This "retroaxonal signaling" mechanism, inspired by computational models like backpropagation, could explain how neural representations adapt to behavioral demands, moving beyond simple Hebbian learning.

What Problem Does This Paper Address?

The brain's ability to adapt and learn is remarkable, allowing neural representations to become finely tuned to the demands of specific behavioral tasks. While Hebbian learning (neurons that fire together wire together) provides a foundational principle, it struggles to explain the sophisticated, task-specific nature of these representations. Computational models like "backpropagation" (backprop) have achieved impressive results by incorporating a "training signal" that flows backward from target cells to inform synaptic adjustments. However, the biological plausibility of such a mechanism in the brain has been questioned due to the presumed speed mismatch between fast synaptic plasticity and slower retroaxonal signaling. This paper aims to bridge this gap by proposing a framework where retroaxonal signals, operating on slower timescales, can still guide the formation of behaviorally adapted neural representations.

Methods

The paper doesn't present new experimental methods but rather a theoretical framework integrating existing research. The core of their proposal is a hypothesis about the interplay between:

  • Synaptic Plasticity: Changes in the strength of synapses, particularly input synapses, leading to altered neuronal firing patterns.
  • Output Synapse Strengthening: If these altered firing patterns are beneficial (e.g., lead to successful behavior), the neuron's output synapses are strengthened.
  • Retroaxonal Signaling: This strengthening of output synapses triggers a signal that travels backward along the axon to the neuron's cell body (soma).
  • Synaptic Consolidation: The retroaxonal signal then acts to stabilize the recent changes in the input synapses. If the output synapses are not strengthened, the retroaxonal signal is absent, and the input synapse changes decay.

The authors draw parallels to computational models like backpropagation and biological phenomena like "synaptic tagging," where a transient potentiation can be consolidated by a later signal. They also discuss potential molecular mechanisms, highlighting the role of neurotrophins and the transcription factor CREB.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Retroaxonal Signals as Stabilizers, Not Instructors: The key insight is that retroaxonal signals don't need to be fast and instructive (like in backprop) to be useful. Instead, they can act as a slower, selective "consolidation" mechanism, stabilizing beneficial synaptic changes.
  • "Stability of the Fittest" in Adult Learning: The paper extends the concept of "neuronal Darwinism" from development to adult learning. Instead of cell survival, it's the stability of beneficial synaptic changes that determines which representations persist.
  • Bridging Computational and Biological Models: The framework offers a way to reconcile the success of backpropagation-like algorithms with the biological constraints of neuronal signaling, particularly the slower speed of retroaxonal communication.
  • Candidate Molecular Mechanisms: The authors identify neurotrophins (like BDNF) and CREB activation as plausible molecular players in this retroaxonal stabilization process.

Key Results

This paper is theoretical and proposes a hypothesis, rather than presenting experimental results. The "results" are the logical implications of their proposed framework:

  • Behaviorally Adapted Representations: The proposed mechanism would lead to neural representations that are specifically tailored to the demands of behavioral tasks, as output synapse strengthening would be driven by behavioral relevance.
  • Selective Retention of Information: Information that is relevant to producing behavior would be preferentially retained, while irrelevant information would decay.
  • Plausibility within Biological Timescales: The proposed mechanism operates within the ~1-hour window for synaptic tagging consolidation, making it biologically plausible despite the slower nature of retroaxonal signaling compared to synaptic transmission.

Limitations and Future Work

  • Circumstantial Evidence: The authors acknowledge that the evidence presented is circumstantial, relying on the convergence of findings from molecular mechanisms, computational models, and electrophysiological recordings.
  • Experimental Verification Needed: The hypothesis requires direct experimental investigation to confirm the proposed retroaxonal signaling and stabilization mechanisms in adult learning.
  • Role of Neuromodulators: While retroaxonal signals are proposed for cell-specific selection, the paper suggests that neuromodulatory systems might provide a complementary, temporally precise signal, especially during heightened alertness. Further research could explore the interaction between these two systems.
  • Specific Molecular Pathways: While neurotrophins and CREB are highlighted, further research is needed to fully elucidate the specific molecular pathways involved in retroaxonal signaling and synaptic consolidation in adult learning.

Summary updated

BibTeX
@inproceedings{harris2008stability,
  author = {Kenneth D. Harris},
  title = {{Stability of the fittest: organizing learning through retroaxonal signals}},
  booktitle = {Trends},
  year = {2008},
  doi = {10.1016/j.tins.2007.12.002},
}
Figure