Optimal information loading into working memory explains dynamic coding in the prefrontal cortex

Jake P. Stroud, Kei Watanabe, Takafumi Suzuki, Mark G. Stokes, Máté LengyelProceedings of the National Academy of Sciences 2023 · DOI

TL;DR

This paper proposes that "optimal information loading" – how the brain initially encodes information for working memory – explains the puzzling "dynamic coding" observed in the prefrontal cortex (PFC). By analyzing mathematical models and monkey brain recordings, they show that inputs orthogonal to the sustained memory state, rather than aligned with it, are key to this dynamic coding, unifying previous theories of brain function.

What Problem Does This Paper Address?

Working memory, our ability to hold information in mind, is crucial for complex cognition. While we know the prefrontal cortex (PFC) is vital for this, the neural mechanisms are debated. A key observation is "dynamic coding": neural activity changes significantly over time, even when the information being remembered remains stable. This contrasts with simpler models that predict steady, persistent activity. Specifically, researchers observe that information is easily decoded when training and testing a decoder at the same time point, or late in the delay period, but poorly decoded when training early and testing late, or vice versa. This suggests a dynamic shift in how information is represented. Previous models, like attractor networks (predicting stable states) or sequential models (predicting rapid transitions), couldn't fully explain this combination of early dynamic and late stable coding. The paper aims to bridge this gap by investigating a less-explored aspect: how information is initially "loaded" into working memory.

Methods

The authors employ a multi-pronged approach:

  • Mathematical Analysis: They use simplified linear dynamical systems (two-neuron models) to derive theoretical principles of information loading. This allows them to mathematically define and analyze concepts like the "persistent mode" (the stable state of memory) and the "most amplifying mode" (the input direction that maximally drives activity).
  • Numerical Simulations: They train task-optimized recurrent neural networks (RNNs) on working memory tasks. This allows them to explore how different training objectives and network architectures (symmetric vs. unconstrained connectivity) lead to different dynamics.
  • Neural Recordings: They analyze electrophysiological data from monkeys performing a working memory task. This provides real-world neural data to compare against their models.
  • Novel Metrics: They develop a specific metric based on the overlap of neural activity with "amplifying" and "persistent" modes to identify optimal information loading strategies. They also use cross-temporal decoding, a standard technique for analyzing dynamic neural codes.

Datasets

The study uses two main sources of data:

  • Simulated Data: Generated from the authors' own task-optimized recurrent neural network models, trained on a memory-guided saccade task.
  • Experimental Data: Electrophysiological recordings from the lateral prefrontal cortex (lPFC) of two monkeys performing a variable-delay memory-guided saccade task. This data was previously published (ref. 49).

Key Insights and Ideas

  • Orthogonality is Key: The most surprising finding is that optimal information loading into working memory involves initial inputs that are orthogonal (or near-orthogonal) to the sustained memory state (the "persistent mode"). This is counter-intuitive, as one might expect inputs to align with the target state.
  • "Most Amplifying Mode": This orthogonal input direction is identified as the "most amplifying mode," which efficiently drives the network towards its stable state while generating transient dynamics.
  • Unifying Theory: This concept of optimal information loading, particularly with unconstrained network connectivity, naturally explains both the early dynamic coding (transient activity) and the later stable coding (persistent activity) observed in the PFC. It bridges the gap between attractor and sequential network models.
  • Dynamic Coding as Functional: The paper argues that dynamic coding isn't just a byproduct but a functionally useful feature arising from optimal information loading, allowing for efficient memory encoding.
  • PFC is "Non-normal": The analysis of monkey data suggests that lPFC circuits exhibit "non-normal" dynamics, meaning their behavior is more complex than simple linear systems, which is consistent with the amplifying mode strategy.

Key Results

  • Model Predictions:
    • Symmetric attractor networks, when given inputs aligned with their persistent mode, show stable coding and poor dynamic coding.
  • Unconstrained attractor networks, when optimized for information loading, exhibit early dynamic coding and late stable coding, mirroring experimental data.
  • Task-optimized RNNs trained with a "just-in-time" cost function (requiring stable coding only before the response) show dynamics most similar to experimental data, including the characteristic overlap cross-over between amplifying and persistent modes.
  • Experimental Validation:
    • Monkey lPFC recordings show a clear "dynamic coding" pattern in cross-temporal decoding, consistent with the models.
  • The developed overlap metric reveals that lPFC activity initially overlaps strongly with "amplifying" modes and then shifts to "persistent" modes, a signature of optimal information loading.
  • The analysis of lPFC data suggests a high degree of "non-normality" in neural dynamics.

Limitations and Future Work

  • Model Simplifications: The authors acknowledge that their linear models are simplifications of biological neural circuits. While they extend findings to nonlinear networks, real biological systems are even more complex.
  • "Non-normality" in Data: While the data suggests strong non-normality, the models with randomly chosen weights had limited non-normality. This suggests that real PFC circuits might employ even more sophisticated strategies than those fully explored.
  • Specific Task: The study focuses on a memory-guided saccade task. Future work could explore if these principles generalize to more complex working memory tasks involving manipulation or decision-making.
  • Activity-Silent Mechanisms: The paper primarily focuses on dynamic coding driven by neural activity. While they briefly mention activity-silent mechanisms, a deeper integration could be explored.
  • Further Experimental Validation: While the overlap metric shows promise, further experimental validation across different tasks and brain regions would strengthen the findings.

Summary updated

Abstract

Working memory involves the short-term maintenance of information and is critical in many tasks. The neural circuit dynamics underlying working memory remain poorly understood, with different aspects of prefrontal cortical (PFC) responses explained by different putative mechanisms. By mathematical analysis, numerical simulations, and using recordings from monkey PFC, we investigate a critical but hitherto ignored aspect of working memory dynamics: information loading. We find that, contrary to common assumptions, optimal loading of information into working memory involves inputs that are largely orthogonal, rather than similar, to the late delay activities observed during memory maintenance, naturally leading to the widely observed phenomenon of dynamic coding in PFC. Using a theoretically principled metric, we show that PFC exhibits the hallmarks of optimal information loading. We also find that optimal information loading emerges as a general dynamical strategy in task-optimized recurrent neural networks. Our theory unifies previous, seemingly conflicting theories of memory maintenance based on attractor or purely sequential dynamics and reveals a normative principle underlying dynamic coding.
BibTeX
@inproceedings{stroud2023optimal,
  author = {Jake P. Stroud and Kei Watanabe and Takafumi Suzuki and Mark G. Stokes and Máté Lengyel},
  title = {{Optimal information loading into working memory explains dynamic coding in the prefrontal cortex}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2023},
  doi = {10.1073/pnas.2307991120},
}
Figure