Dynamical Latent State Computation in the Posterior Parietal Cortex

Kaushik J. Lakshminarasimhan, Eric Avila, Xaq Pitkow, Dora E. AngelakibioRxiv 2022 · DOI

TL;DR

This paper investigates how the brain tracks hidden states of the world, like our position, during navigation. By recording from monkey brains and using advanced computational models, they found that the posterior parietal cortex (PPC) dynamically computes and updates this "world model" through complex neural interactions, which is crucial for guiding behavior.

What Problem Does This Paper Address?

Many everyday tasks, from driving to navigating a familiar room, require us to keep track of "hidden states" – information not directly available to our senses, like our precise location or the speed of a car we can't see. Understanding how the brain computes and maintains these dynamic, continuous-valued latent states is a major challenge in neuroscience. While we have good models for sensory perception and motor control, the neural mechanisms underlying the inference and tracking of these hidden world states remain largely unknown. This paper aims to bridge that gap by exploring how the posterior parietal cortex (PPC), a brain region implicated in spatial processing and working memory, handles this complex computation during naturalistic navigation.

Methods

The researchers employed a multi-pronged approach:

  • Behavioral Experimentation: Monkeys were trained to navigate a virtual environment using optic flow cues to reach a hidden target, without explicit landmarks. This created a task where inferring self-motion and updating spatial position were essential.
  • Neural Recordings: They simultaneously recorded activity from a large number of neurons in the PPC of these monkeys during the navigation task.
  • Computational Modeling:
    • Generalized Additive Model (GAM): This statistical model was used to analyze how individual neurons' activity related to various task variables (sensory input, latent states, motor output, etc.).
  • Recurrent Neural Network (RNN): A sophisticated RNN was trained to perform the same navigation task as the monkeys. This allowed for a direct comparison of neural representations learned by the biological brain and the artificial network.
  • Decoding Analysis: Linear decoders were trained to predict task variables from population neural activity, and the accuracy of these predictions was correlated with behavioral performance.
  • Task Manipulations: The reliability of sensory input, the gain of motor control, and the latent state dynamics were manipulated during experiments to observe their effects on neural activity and interactions.

Datasets

The study used data collected from three rhesus macaques performing a virtual reality navigation task. This dataset, comprising neural recordings and behavioral data from over 120,000 trials, was used for both training and evaluating computational models. No new benchmark datasets were introduced.

Key Insights and Ideas

  • PPC as a Dynamic World Model: The findings strongly suggest that the PPC doesn't just passively receive information but actively computes and maintains a dynamic internal model of the world, specifically tracking the animal's spatial position relative to a goal.
  • Neural Interactions are Key: The study highlights the critical role of inter-neuronal coupling (how neurons influence each other) in enabling this dynamic world model. Changes to task conditions significantly altered these interactions, impacting how latent states were represented.
  • Latent States are Special: While sensory and motor representations remained relatively stable across task manipulations, the representation of the latent state (position) was significantly affected, indicating its unique computational role.
  • RNNs as a Window into Neural Computation: The remarkable similarity between the representations learned by the trained RNN and the monkey PPC provides strong evidence that the neural mechanisms in PPC are shaped by the computational demands of the task, potentially driven by architectural constraints.

Key Results

  • Sequential Neural Dynamics: Neurons in the PPC exhibited sequential activity patterns at the population level, rather than persistent activity, suggesting a dynamic updating process.
  • Decoding Latent States: The monkeys' spatial position (latent state) could be dynamically decoded from population PPC activity with high accuracy.
  • Behavioral Prediction: The accuracy of decoding the latent state was directly correlated with the monkeys' navigation performance and behavioral errors on individual trials.
  • Task Manipulations Alter Coupling and Latent Representation: Manipulating sensory reliability, motor gain, or latent dynamics significantly changed neural coupling and the representation of the latent state, while sensory and motor representations remained largely stable.
  • RNN Recapitulates Findings: A trained RNN model showed similar sequential dynamics, tuning to task variables, and sensitivity to task manipulations as observed in the monkey PPC.

Limitations and Future Work

  • Causal Interventions: While the study used task manipulations, it acknowledges the need for direct causal intervention experiments (e.g., inactivating specific brain regions) to definitively establish the role of PPC in computing latent states.
  • Inheritance of Information: The study suggests that PPC contributes to latent state computation but doesn't definitively rule out that some information might be inherited from other brain areas. Future work will explore this through brain-wide recordings and causal interventions.
  • Stability of Representations: The study examined the effects of task manipulations but did not investigate the long-term stability of neural representations over extended periods.
  • Autonomous Strategies: The role of purely planning-based (autonomous) strategies, independent of sensory feedback, was not fully explored, and future work could compare PPC's contribution with that of frontal brain regions involved in planning.

Summary updated

Abstract

Summary Success in many real-world tasks depends on our ability to dynamically track hidden states of the world. To understand the underlying neural computations, we recorded brain activity in posterior parietal cortex (PPC) of monkeys navigating by optic flow to a hidden target location within a virtual environment, without explicit position cues. In addition to sequential neural dynamics and strong interneuronal interactions, we found that the hidden state – monkey’s displacement from the goal – was encoded in single neurons, and could be dynamically decoded from population activity. The decoded estimates predicted navigation performance on individual trials. Task manipulations that perturbed the world model induced substantial changes in neural interactions, and modified the neural representation of the hidden state, while representations of sensory and motor variables remained stable. The findings were recapitulated by a task-optimized recurrent neural network model, suggesting that neural interactions in PPC embody the world model to consolidate information and track task-relevant hidden states.
BibTeX
@misc{lakshminarasimhan2022dynamical,
  author = {Kaushik J. Lakshminarasimhan and Eric Avila and Xaq Pitkow and Dora E. Angelaki},
  title = {{Dynamical Latent State Computation in the Posterior Parietal Cortex}},
  year = {2022},
  doi = {10.1101/2022.01.12.476065},
  howpublished = {bioRxiv},
}
Figure