Compressed Cortical Input Separates Control from Dynamics in Striatum

1. Problem Statement

The dorsolateral striatum (DLS) is critical for a wide range of time-sensitive behaviors, including action chunking (grouping sequences into single units), duration estimation, and motor timing. However, a significant theoretical gap exists: no single existing framework successfully explains how the DLS simultaneously supports these diverse temporal functions. Current models often fail to unify the anatomical reality of massive cortical convergence with the functional diversity observed in DLS-dependent behaviors. The core challenge is to identify a computational principle that links the anatomical structure of corticostriatal projections to these varied temporal dynamics.

2. Methodology

The authors propose and test a novel computational framework using a corticostriatal neural network model. The methodology involves the following key components:

• Network Architecture:
  • Cortical Module: A recurrent neural network (RNN) representing the cortex.
  • Striatal Module: A separate recurrent neural network representing the striatum.
  • The Bottleneck: The two modules communicate via a low-dimensional, noisy bottleneck. This bottleneck simulates the "massive convergence" of cortical projections onto striatal neurons, where high-dimensional cortical information is compressed into a lower-dimensional signal before reaching the striatum.
• Training Regime: The entire system is trained using Reinforcement Learning (RL). The agents are tasked with solving three distinct DLS-associated tasks:
  1. Action chunking.
  2. Duration estimation.
  3. Motor timing.
• Perturbation Analysis: The authors perform causal interventions by perturbing the compressed cortical signal to observe the resulting changes in behavior and internal neural dynamics.

3. Key Contributions

The paper introduces a unifying computational motif that bridges information theory and dynamical systems theory in the context of basal ganglia function:

• Separation of Control and Dynamics: The primary contribution is the demonstration that compression forces a functional separation. The cortex provides low-dimensional control signals (essentially "what" to do or "when" to switch), while the striatum generates stable, time-encoding dynamics (the "how" and the temporal evolution of the action).
• Anatomical-Functional Link: It establishes a direct link between the anatomical fact of massive cortical convergence (compression) and the emergence of specific temporal behaviors.
• Unified Framework: It offers a single model capable of explaining action chunking, duration estimation, and motor timing, resolving the fragmentation in previous literature.

4. Key Results

The model successfully reproduced complex behavioral phenomena and neural signatures across the three tasks:

• Emergent Behaviors:
  • Action Chunking with Slipping: The system naturally developed chunked behaviors. Crucially, it exhibited "action slipping," where errors occurred at the boundaries of chunks, mirroring biological observations.
  • Intensity-Biased Duration Judgements: The model demonstrated that duration estimation is not absolute but is modulated by stimulus intensity, a known biological phenomenon.
  • Stereotyped Motor Timing: The system generated consistent, stereotyped timing programs for motor sequences.
• Neural Dynamics:
  • The compressed cortical signal acted as a control input that modulated the striatal dynamics without dictating the full trajectory.
  • The striatum maintained stable, time-encoding dynamics even when the cortical input was noisy or compressed.
• Causal Perturbation Effects:
  • When the compressed cortical signal was perturbed, the behavior shifted (e.g., timing errors or chunking failures).
  • However, the sequential structure of striatal activity was preserved. This suggests that the striatum maintains an internal temporal scaffold (dynamics) that is robust, while the cortex provides the flexible control signals that drive specific behavioral outputs.

5. Significance

This work is significant for several reasons in neuroscience and computational modeling:

• Theoretical Unification: It successfully unifies information-theoretic perspectives (focusing on compression and dimensionality reduction) with dynamical systems perspectives (focusing on stable attractor states and temporal encoding).
• Mechanistic Insight: It provides a mechanistic explanation for why the striatum has such massive convergence from the cortex. The compression is not just a biological constraint but a functional necessity that separates control from execution dynamics.
• Clinical Relevance: By linking specific computational motifs to DLS functions, the findings offer new hypotheses for understanding disorders affecting the basal ganglia (such as Parkinson's or Huntington's disease) where timing and sequencing deficits are prominent. It suggests that therapeutic interventions might need to target the integrity of the cortical control signal or the stability of striatal dynamics differently.
• Generalizability: The proposed "compression-driven separation" motif may serve as a general principle for understanding how other brain regions with massive convergent inputs process complex temporal information.