Hebbian Consolidation during Sleep Drives the Semantization of Episodic Sequences: A Neurocomputational Model

Technical Summary: Hebbian Consolidation during Sleep Drives the Semantization of Episodic Sequences

Problem Statement
While it is widely accepted that sleep consolidates episodic memories and transforms them into generalized, semantic representations, the specific neural mechanisms driving this "semantization" remain debated. Existing neurocomputational models within the Complementary Learning Systems (CLS) framework often rely on error-driven learning, gradient descent, or abstract optimization procedures that lack biological plausibility. Furthermore, few models explicitly address how the replay of temporally ordered episodic sequences (rather than isolated items) during sleep can drive the abstraction of context-bound autobiographical memories into decontextualized cortical representations, while simultaneously distinguishing the computational roles of non-REM (SWS) and REM sleep stages.

Methodology
The authors present a biologically grounded neurocomputational model that integrates a hippocampal episodic memory system with a cortical sensory-semantic-lexical network. The model relies exclusively on Hebbian and anti-Hebbian plasticity rules, avoiding external objective functions or error-correction algorithms.

• Architecture:

  • Hippocampus (HC): Modeled with CA3 and CA1 layers using a theta-gamma code. Episodes are represented as collections of features oscillating in gamma synchrony, with sequential episodes occupying different theta phases. Features in the HC are strictly orthogonal and autobiographical (e.g., a specific cat meowing at a specific time).
  • Cortical Network: Comprises three layers:
    1. Sensory Cortex (SC): Encodes features in a compressed, non-orthogonal form (general features). It receives input from the HC and external sensory inputs.
    2. Semantic Network (SN): An autoassociative network where units represent general concepts (e.g., "cat"). It learns statistical correlations among features through prolonged training. It includes short-term synaptic fatigue to allow transitions between concepts.
    3. Lexical Network (LN): A heteroassociative network linking semantic features to word forms.
  • Plasticity: Synapses within the HC (CA3 auto-associative, CA1-CA3 hetero-associative) and within the SN are trained before sleep. Crucially, feedback synapses from the SN to the SC are initially null but become plastic during sleep.

• Simulation Protocols:

  • Pre-Sleep Training: The network learns two distinct episodic sequences (each with four episodes of three features) in the HC and learns a taxonomy of objects in the SN/LN.
  • Non-REM Sleep Simulation: The HC is isolated from external input but reactivates stored sequences randomly ("replay") due to high noise. Acetylcholine (ACh) levels are low, suppressing new HC encoding but allowing replay. The replay drives activity in the SC and SN. A critical mechanism is introduced: a physiological delay (10–12 ms) between the SC and SN. This delay creates a temporal coincidence where the presynaptic activity of a semantic concept (from episode $t$) coincides with the postsynaptic activity of sensory features from the subsequent episode ($t+1$). This triggers Hebbian potentiation of heteroassociative synapses from the SN to the SC.
  • REM Sleep Simulation: The HC is disconnected from the cortex. ACh levels are high, transmitting theta rhythms to the cortex. Increased noise and altered neurotransmitter levels (low norepinephrine) lead to fragmented, recombined activity patterns in the cortex, simulating dream-like associations.

Key Results

1. Semantization via Non-REM Replay: The model demonstrates that Hebbian potentiation during non-REM sleep creates new heteroassociative connections from the Semantic Network to the Sensory Cortex. These connections link the semantic representation of one episode to the sensory features of the next.
2. Generalized Retrieval: Post-consolidation, the model exhibits two retrieval modes:
   • Hippocampal Retrieval: Requires an exact, specific autobiographical cue to recall the sequence (context-dependent).
   • Cortical Retrieval: The sequence can be retrieved from the cortex alone using any combination of general features that represent the objects in the first episode, even if those specific features differ from the original autobiographical experience. This indicates the memory has become "semantized" and decontextualized.
3. REM Fragmentation: REM simulations produce fragmented and recombined patterns where concepts from different sequences are linked based on shared features (e.g., "daughter" replacing "son"), simulating creative but disordered associative thinking.
4. Sensitivity Analysis: The model successfully recovers sequences in the majority of trials when the sensory-semantic delay is set to physiological values (10–12 ms). Deviations from this range result in incomplete recovery, highlighting the importance of precise timing for the Hebbian mechanism.

Significance and Claims
The authors claim this framework provides a parsimonious, physiologically plausible account of how sleep transforms episodic sequential experiences into semantized representations without relying on non-biological learning rules.

• Mechanistic Insight: The study proposes that replay during non-REM sleep acts not merely as a stabilization signal but as a restructuring signal. It organizes pre-existing semantic representations into sequences by reinforcing specific cortical synapses, thereby changing which cues can access a memory and shifting the retrieval substrate from the hippocampus to the cortex.
• Distinction of Sleep Stages: The model offers a unified explanation for the divergent roles of sleep stages: non-REM sleep drives the ordered abstraction and generalization of sequences, while REM sleep facilitates creative, fragmented recombination of these representations.
• Testable Predictions: The paper generates specific predictions for experimental validation:
  1. Non-REM replay should selectively strengthen cortical interactions linking semantic representations of one event to sensory features of subsequent events.
  2. Post-sleep retrieval should be possible from generalized object-level cues that were not part of the original autobiographical episode.
  3. Sleep should specifically improve memory for sequences of episodes rather than just associations among individual items.
  4. Non-REM and REM stages should differentially affect retrieval: non-REM supporting ordered reinstatement, and REM favoring weaker, fragmentary transitions.

The authors acknowledge limitations, including the use of only two sequences, the absence of global synaptic downscaling (though they argue it is compatible with their findings), and the lack of feedback connections from the cortex to the hippocampus. They suggest these as avenues for future work.