Coding and learning of behavioral sequences

Ofer Melamed, Wulfram Gerstner, Wolfgang Maass, Misha Tsodyks, Henry MarkramTrends 2003 · DOI

TL;DR

This paper proposes a novel neural mechanism involving asymmetric receptive fields and oscillatory inhibition that can convert a rate code into a temporal code. This allows the brain to learn and represent behavioral sequences across vastly different timescales using a fixed millisecond-scale learning rule, a long-standing puzzle in neuroscience.

What Problem Does This Paper Address?

The nervous system's ability to learn and execute behavioral sequences that unfold over a wide range of timescales – from milliseconds to seconds – using a fixed, millisecond-based learning rule is a significant challenge. For instance, a piano player can execute rapid sequences of notes much faster than they can play a slow melody, yet the underlying neural mechanisms for learning these sequences are thought to operate on a consistent, rapid timescale. This paper tackles the question of how the brain can reconcile these disparate temporal scales for learning and representation.

Methods

The paper explores theoretical and experimental findings related to neural coding and learning. It discusses concepts like Synfire Chains (SFCs) as a potential framework for millisecond-resolution sequence coding. A key focus is on the mechanism of "phase precession" observed in hippocampal place cells. The authors highlight a proposed explanation for phase precession involving the interplay of asymmetric receptive fields and oscillatory inhibition. This mechanism is posited to transform a "rate code" (where information is encoded by the firing rate of neurons) into a "temporal code" (where information is encoded by the precise timing of spikes). The paper also touches upon Spike-Timing Dependent Plasticity (STDP) as a plausible learning rule that can leverage this temporal coding.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Bridging Time Scales: The core insight is that the brain might not need separate mechanisms for fast and slow sequences. Instead, a single, rapid mechanism can be used to represent information across different timescales.
  • Rate Code to Temporal Code Conversion: The paper emphasizes how oscillatory inhibition, when combined with asymmetric receptive fields, can effectively convert a rate-based representation into a precise temporal code. This is crucial because millisecond-level learning rules like STDP operate on precise spike timing.
  • Phase Precession as a Temporal Compressor: Hippocampal phase precession is presented as a biological example of this temporal compression. It allows sequences of events that occur over longer durations to be represented within a single, short theta cycle, making them amenable to millisecond-scale learning.
  • Asymmetric Hebbian Plasticity: The development of asymmetric receptive fields is linked to asymmetric Hebbian learning rules, suggesting that the learning process itself shapes the neural representations in a way that facilitates temporal coding.

Key Results

The paper doesn't present new experimental results but rather synthesizes existing findings and proposes a unifying theoretical framework. The key "results" are conceptual:

  • A model where asymmetric excitation and oscillatory inhibition can order the precise timing of spikes in neurons with different receptive fields.
  • This mechanism allows for the squeezing of activity into a standardized time frame, enabling a fixed learning rule to scale synapses based on spatial location and temporal activation.
  • This mechanism is suggested to be generalizable to receptive fields in sensory cortices where periodic background signals are present.

Limitations and Future Work

  • Resetting of Symmetry: A significant open question is why the asymmetry of place fields, observed during training, appears to be reset the following day, while phase precession persists. This suggests that phase precession might be generated by a different learning mechanism or that the relationship between asymmetry and precession needs further clarification.
  • Order of Learning: It remains unclear whether sequences are learned before or after the asymmetry of place fields evolves.
  • Generalizability: While the proposed mechanism is suggested to be general, further experimental validation across different brain regions and sensory modalities is needed.
  • Alternative Explanations: The paper acknowledges that other lines of research suggest that bridging time scales might be less problematic due to inherent long time constants in neural circuits, which could naturally support temporal information extraction by readout neurons. This highlights the need for further research to compare and integrate these different perspectives.

Summary updated

BibTeX
@inproceedings{melamed2003coding,
  author = {Ofer Melamed and Wulfram Gerstner and Wolfgang Maass and Misha Tsodyks and Henry Markram},
  title = {{Coding and learning of behavioral sequences}},
  booktitle = {Trends},
  year = {2003},
  doi = {10.1016/j.tins.2003.10.014},
}
Figure