Predictive pursuit emerges in high-dimensional recurrent neural networks

1. Problem Statement

The paper addresses a gap in understanding the computational principles underlying predictive pursuit—the ability of an organism to track and anticipate the future location of a moving target. While recent behavioral experiments in freely moving rodents have characterized the neural dynamics associated with visually guided pursuit, the specific mechanisms allowing for internal prediction and anticipatory behavior remain poorly understood. The core challenge is determining how neural systems generate predictions of future target states to execute efficient pursuit in complex, dynamic environments.

2. Methodology

To investigate these mechanisms, the authors employed a Recurrent Neural Network (RNN) modeling approach:

• Model Architecture: They developed a high-dimensional RNN model trained to perform a visually guided pursuit task. The agent in the model must track a moving target and adjust its trajectory to maintain pursuit.
• Training Regimen: The model was exposed to stereotyped trajectories of the target to simulate learning from repeated environmental patterns.
• Dimensionality Manipulation: A critical experimental variable was the rank of the network. The authors trained models of varying ranks (low-dimensional vs. high-dimensional) to test the hypothesis that sufficient neural resources (high dimensionality) are required for predictive capabilities.
• Analysis Techniques:
  • Behavioral Analysis: Comparing model pursuit strategies against empirical data from rodents.
  • Neural Representation Decoding: Analyzing the internal state of the RNN to identify specific units encoding target position.
  • Ablation Studies: Systematically removing specific functional units to establish causal links between neural representations and task performance.

3. Key Contributions

• Emergence of Predictive Behavior: The study demonstrates that predictive pursuit is not hard-coded but emerges in RNNs through exposure to stereotyped trajectories, leading to increased anticipatory behaviors.
• Identification of Egocentric Coding: The authors identified specific units within the RNN that encode the egocentric location of the target (position relative to the agent). This finding provides a computational analog to "egocentric target neurons" observed in biological animals.
• Causal Validation: Through ablation experiments, the paper establishes a causal role for these egocentric units; their removal significantly degrades pursuit performance, confirming their necessity for efficient tracking.
• Dimensionality Hypothesis: The work proposes and validates the hypothesis that high-dimensional neural codes are a prerequisite for predictive pursuit, distinguishing it from simpler vision or motor tasks that may be solvable with lower-dimensional representations.

4. Key Results

• Anticipatory Performance: The RNN successfully generated internal predictions of the target's future location. The accuracy of these predictions and the degree of anticipatory behavior correlated positively with the amount of exposure to stereotyped target trajectories.
• Alignment with Biology: The emergent pursuit strategy of the RNN was highly aligned with the behavior observed in real rodents, validating the model's biological relevance.
• Functional Specialization:
  • Low-Rank Models: While capable of strong pursuit performance, these models failed to exhibit significant anticipatory behavior. They encoded egocentric target location but lacked complex spatial representations.
  • High-Rank Models: Only models with sufficiently high rank (high dimensionality) exhibited robust anticipatory behavior. Crucially, these high-dimensional networks were the only ones to develop representations of allocentric (world-centered) self and target locations.
• Necessity of High Dimensionality: The results indicate that while basic pursuit can be achieved in lower-dimensional networks, the predictive component of the task specifically requires the computational capacity of high-dimensional networks.

5. Significance

This paper provides a foundational computational framework for understanding predictive coding in motor control and navigation.

• Theoretical Advancement: It challenges the notion that simple motor tasks can be fully explained by low-dimensional dynamics, suggesting that complex cognitive tasks like prediction require high-dimensional neural manifolds.
• Biological Insight: By linking RNN emergent properties (egocentric coding, allocentric emergence) to biological observations, the study offers a mechanistic explanation for how the brain might support the complex computations required for dynamic pursuit in the wild.
• Future Directions: The findings suggest that future models of biological intelligence must account for network dimensionality to accurately simulate predictive behaviors, moving beyond standard vision or memory task paradigms.