Oscillatory Phase Coding Emerges in Adaptive RNNs for Flexible Working Memory

arXiv:2611.06932 · q-bio.NC, cs.AI · Submitted 2026-09-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Oscillatory Phase Coding Emerges in Adaptive RNNs for Flexible Working Memory".

Marcus: Here's the gist of the paper. Oscillatory phase coding (OPC) emerges in recurrent neural networks when trained with adaptation to perform sequence working memory tasks with variable delays The memorized items recur at distinct phases of a shared rhythm, with oscillation amplitude encoding item identity and phase encoding serial order This study reveals how OPC arises from network structures and task demands, offers insights into the implementation and computational mechanisms of OPC in biological systems, and inspires us to develop computational models for human-like cognitive functions.

Ines: Emergence of Oscillatory Phase Coding The recurrent neural networks (A-RNNs) spontaneously develop oscillatory phase coding after training on a sequence working memory task with flexible delays The memorized items are reactivated at distinct phases in an internally generated rhythm, with oscillation amplitude encoding item identity and phase encoding serial order.

Marcus: Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information.

Ines: Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately.

Paper summary: Ines: So to wrap up, the paper "Oscillatory Phase Coding Emerges in Adaptive RNNs for Flexible Working Memory" shows that oscillatory phase coding develops naturally when you train adaptive recurrent neural networks to handle sequence working memory with variable delays. The core finding is that the network spontaneously organizes items into a rhythm where amplitude carries identity and phase carries order.

Marcus: The authors found this emergence depends on having adaptation in neural dynamics, using variable delay in retention, and flexibility in reading out information. They identified specific cell types—content cells for identity and time cells for sequence—and showed how their recurrent connections are organized to support this structure across different model sizes.

Yuki: From a population genetic perspective, this suggests that the underlying computational principles of sequence processing might be more general than we think, potentially applicable across different species if these mechanisms are conserved. The focus on absolute versus item-relative phase tuning gives us a concrete hypothesis about how biological systems might represent memory structure.

Ines: Exactly. The title points to OPC emerging in adaptive RNNs, but the authors show it's not just a feature of any RNN; it's tied directly to the task demands—the variable delay and adaptation. It’s about how structure emerges from dynamic interaction rather than being hard-coded.

Marcus: The implication for the wider field is that we have a candidate mechanism—a neural circuit model—for sequence working memory that operates through this phase coding. It offers a computational blueprint for what might be happening in biological systems when they need to juggle complex, ordered information over time.

Yuki: What this means for us listening right now is that it gives us a new lens to look at how the brain handles juggling multiple pieces of information—the identity and the order—which is something we all do constantly. It connects abstract computation to observable neural dynamics.

Conclusion: Ines: So we're wrapping up on this paper about oscillatory phase coding emerging in adaptive recurrent neural networks for flexible working memory. Essentially, these networks develop a rhythm when trained to juggle sequences with variable delays, where the rhythm's amplitude tells you what item it is and the timing tells you its order.

Marcus: Yeah, so they showed how this structure pops up naturally from the network dynamics when you introduce that variable delay in retention and adaptation into the system. It’s not something they explicitly program; it just happens under those specific training conditions.

Yuki: From a population genetics viewpoint, this suggests that the fundamental computational principles for handling sequential information might be quite robust across different biological contexts, which is interesting if we think about how memory systems evolve in organisms with varying cognitive demands.

Ines: Exactly. The authors identified specific cell types—content cells for identity and time cells for sequence—and showed how their connections are organized to support this rhythmic structure across different network sizes. It gives us a concrete model of how the brain might physically organize information during working memory tasks.

Marcus: I mean, the stability of this phase coding is quite strong; only when you have that variable delay in retention does it exceed a really high measure called R phi, which they found to be over zero point nine. That suggests this mechanism isn't just some fluke that works in one setup; it’s robust enough to be considered a real candidate for biological function.

Yuki: But we have to remember the caveat here—the study is based on these artificial recurrent networks, so we need to be careful not to extrapolate these findings directly back onto complex biological neural circuits without more specific data from living systems. The sample size and the exact neuron types used matter a lot in those comparisons.

Ines: True, that's the limitation, obviously. But what this work really changes for us is providing a computational blueprint for flexible working memory that we can test against existing biological models, which is a big step forward in how we try to understand cognition.

Marcus: It gives us something tangible to look at when we think about what specific circuits might be responsible for sequence processing in the brain, moving beyond just abstract ideas and giving us a mechanism that could actually be simulated.

Yuki: So, it moves the discussion from "can this work?" to "how does this relate to known biological constraints and evolutionary pressures?" which is where we need to go next.

Peking University

q-bio.NC, cs.AI

Submitted: 2026-09-28

Updated: 2026-11-06

Comments: 37 pages, 16 figures

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 91/100

The gist: Oscillatory phase coding (OPC) emerges in recurrent neural networks when trained with adaptation to perform sequence working memory tasks with variable delays The memorized items recur at distinct…

Key concepts

Oscillatory Phase Coding (OPC)
A pattern in the neural network where memorized items recur at specific points in a shared rhythm. The size of the oscillation encodes what the item is, while its phase encodes its position in the sequence.
Adaptation State ($ heta_t$)
A slow-changing state within the A-RNN that evolves more slowly than neural activity. This state feeds back into network dynamics, helping to capture changes in neural behavior during the learning process for variable delays.
Content and Time Cells
Specific types of neurons identified in the trained network. Content cells are phase-locked to a particular item's reactivation, suggesting they carry item identity. Time cells form a traveling wave providing a shared temporal reference across all items.

Terminology

Summary

Oscillatory phase coding (OPC) emerges in recurrent neural networks when trained with adaptation to perform sequence working memory tasks with variable delays The memorized items recur at distinct phases of a shared rhythm, with oscillation amplitude encoding item identity and phase encoding serial order This study reveals how OPC arises from network structures and task demands, offers insights into the implementation and computational mechanisms of OPC in biological systems, and inspires us to develop computational models for human-like cognitive functions. p. 0

Where the paper says it

Together, this study reveals how OPC arises from network structures and task demands, offers insights into the implementation and computational mechanisms of OPC in biological systems, and inspires us to develop computational models for human-like cognitive functions. 1 INTRODUCTION Oscillation is a pervasive feature of neural activity across brain regions and is thought to play important functional roles (Buzsáki & Draguhn, 2004; Lisman & Jensen, 2013; Malhotra et al.

Emergence of Oscillatory Phase Coding

The recurrent neural networks (A-RNNs) spontaneously develop oscillatory phase coding after training on a sequence working memory task with flexible delays The memorized items are reactivated at distinct phases in an internally generated rhythm, with oscillation amplitude encoding item identity and phase encoding serial order. Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information. p. 0

Where the paper says it

After training, we found that OPC emerges in the RNN spontaneously: the memorized items recur at distinct phases of a shared rhythm, with oscillation amplitude encoding item identity and phase encoding serial order. Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information.

Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately. p. 0
Where the paper says it

Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information. Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately.

Recurrent connectivity analysis revealed phase-organized connections among time cells and structured connections within digit-selective content-cell groups. p. 0
Where the paper says it

Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately. Recurrent connectivity analysis revealed phase-organized connections among time cells and structured connections within digit-selective content-cell groups.

How the Training Signal is Built

The A-RNN architecture involves sensory encoders, an Adaptive RNN (A-RNN), and dual readout pathways. The network dynamics are governed by equations describing recurrent activity, adaptation states, and external inputs. p. 2

Where the paper says it

• Cue Encoder (MLP): A multilayer perceptron encodes the ordinal cue during the retrieval stage and projects it to the RNN. 2) Adaptive Recurrent Neural Network (A-RNN): The recurrent network integrates external inputs and the activities of hidden neurons, with adaptation in the dynamics of single neurons (Figure 1d).

Spike-frequency adaptation is captured by the adaptation state θt, which evolves more slowly than neural activity and feeds back to the network dynamics through the term −θt−1. p. 3
Where the paper says it

One mechanism involves slowly accumulating potassium currents which provide negative feedback to reduce neuronal excitability (Madison & Nicoll, 1984; Schwindt et al. In our model, these adaptive dynamics are captured by the adaptation state θt, which evolves more slowly than the neural activity (αθ < αh) and feeds back to the network dynamics through the term −θt−1.

The dual-readout pathways produce content scores (pt) and a temporal gate score (gt), where gt is controlled by αt, which peaks at the phase when the oscillation of the target item reaches its maximum. p. 3
Where the paper says it

• Gating Head: This pathway reads out the timing of extracting the target item specified by the ordinal cue. The timing is represented by a gating variable αt ∈ [0, 1], which peaks at the phase when the oscillation of the target item reaches its maximum.

Where the Gains Come From

The stability of oscillatory phase coding is robust across different training conditions, with only the variable-delay A-RNN exceeding a robust Rϕ of 0.9. The content and time cells are identified through concentration measures, where time-cell candidates meet Rabs i ≥ 0. p. 6

Where the paper says it

(b,c) Absolute (b) and content-relative (c) phase concentration across trials. Colored bars mark neurons meeting the classification criteria; the inset in (b) shows the preferred phases of time-cell candidates.

7 and content cells meet Rcontent i ≥ 0. p. 0
Where the paper says it

Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information. Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately.

7. Recurrent connectivity analysis shows that within-digit blocks contain coherent organization, with the fraction f1(B) measuring how much of the block can be described by its strongest common pattern. p. 6
Where the paper says it

(e) Recurrent connectivity kernel of time cells on the phase ring, showing an approximately cosine shape. (f) Leading-mode fraction within and across digit-selective content-cell groups, normalized by the mean shuffled fraction.

Pruning redundant reasoning tokens

The proposed neural circuit model reproduces OPC through interactions between content units and an adaptive time ring. The time cells form a traveling wave which provides a shared temporal reference for all memorized items. p. 1

Where the paper says it

We further carried out analyses to categorize neuron types in the trained A-RNNs, and we identified two main neuron types: content and time cells. Content cells encode the item identity and their phases track the serial position of the item, while time cells form a traveling wave which provides a shared temporal reference for all memorized items.

Content cells remain phase-locked to the reactivation of a particular item, suggesting that they carry item identity. p. 6
Where the paper says it

Phase references, classification criteria, and connectivity analyses are given in Appendix C. these neurons as content cells; their phases track the reactivation of a particular item, suggesting that they carry item identity.

The model predicts that changing an item’s serial position will shift content-cell phases while preserving the temporal reference provided by time cells. p. 1
Where the paper says it

According to our model, time cells maintain consistent oscillatory phases across sequences, whereas content cells remain phase-locked to the reactivation of their associated items. Thus, our model predicts that changing an item’s serial position will shift content-cell phases while preserving the temporal reference provided by time cells.

Learning and Delay Generalization Across Model Sizes

Adaptive networks with H hidden neurons are compared with non-adaptive networks with H + 5 neurons, showing that A-RNNs reach higher mean epoch60 validation accuracy at every tested size. The robust Rϕ measure shows that the variable-delay A-RNN exceeds 0. p. 4

Where the paper says it

Robust Rϕ measures the stability of oscillatory phase coding through cross-trial consistency in phase differences between item readouts at different ordinal positions, evaluated over temporal windows (Appendix D). Only the variable-delay A-RNN exceeds 0.

9 (Figure 3c, bottom). p. 4
Where the paper says it

Every trial uses the same temporal gate, regardless of which digits were presented or which one is cued (Appendix D). After training, the content-readout activities still oscillate alongside slower changes (Figure 3a(ii),c), but the phase differences between item readouts vary more across trials than in the main A-RNN (Figure 3c, bottom).

The analysis of recurrent connectivity patterns in Figure 4e,f recur across the nine models with developed time and content populations. p. 6
Where the paper says it

4 A NEURAL CIRCUIT MODEL OF OPC INSPIRED BY THE TRAINED A-RNN Guided by these activity and connectivity patterns, we propose a neural circuit model of OPC (Figure 5a) with two reciprocally coupled populations: one content unit per digit-selective content-cell group, representing its population-averaged activity, and a population of time cells forming a ring attractor network with adaptation, which provides a temporal oscillation reference through periodic traveling waves. Both populations combine recurrent excitation with adaptation.

The recurrent weights for content cells show within-digit structure is stronger in all nine models. p. 6
Where the paper says it

To relate activity to structure, we examined recurrent connections within each population. The time-cell kernel is approximately cosine-shaped on the phase ring—positive for nearby phases and negative near opposite phases—and disappears under phase shuffling (Figure 4e); content cells show structured connectivity within digit-selective groups (Figure 4f).

The proposed biological test of absolute and item-relative phase tuning suggests that neurons tuned to absolute phase should retain their preferred phases across changes in item identity and order. Neurons with item-relative phase tuning should maintain a consistent phase offset from a particular memory item’s reactivation across permutations. p. 1

Where the paper says it

According to our model, time cells maintain consistent oscillatory phases across sequences, whereas content cells remain phase-locked to the reactivation of their associated items. Thus, our model predicts that changing an item’s serial position will shift content-cell phases while preserving the temporal reference provided by time cells.

The study supports the representational prediction by observing the dissociation between absolute and item-relative phase tuning. p. 1
Where the paper says it

Our study also generated experimentally testable predictions regarding the dissociation between memory content and serial order in biological circuits. According to our model, time cells maintain consistent oscillatory phases across sequences, whereas content cells remain phase-locked to the reactivation of their associated items.

The paper also generates experimentally testable predictions regarding the dissociation between memory content and serial order in biological circuits. The model translates the content–time organization of the trained A-RNN into a candidate circuit mechanism for biological OPC. p. 8

Where the paper says it

2; Figure 10a–c). The model thus translates the content–time organization of the trained A-RNN into a candidate circuit mechanism for biological OPC.

The study may also inspire the development of AI models for realizing flexible working memory. p. 8
Where the paper says it

The study also generates testable predictions: changing an item’s serial order in a sequence will shift the phase of the corresponding content cells while preserving the phases of time cells. Our study may also inspire the development of AI models for realizing flexible working memory.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can spontaneously develop oscillatory phase coding when trained recurrent neural networks (RNNs) are adapted to perform sequence working memory tasks with variable delays, as OPC emerges in the RNN spontaneously.

  2. The improved system can achieve disentangled representations for item identity and serial order because oscillation amplitude encodes item identity and phase encodes serial order, allowing for a disentangled representations scheme.

  3. The system can selectively read out cued items by modulating temporal gating, as the temporal gate learns to select the oscillation phase of the cued item, which is evident in that the largest element of yb predicts the item identity.

  4. The system's learning strategy can be optimized by incorporating neural adaptation, as ablation studies identified neural adaptation and its learnable strength as key elements required for OPC.

  5. The improved system can generate a neural circuit model that reproduces OPC through interactions between content units and an adaptive time ring, providing a simple account of how OPC could operate in biological circuits.

  6. The system can be designed to predict the dissociation between memory content and serial order, as the model predicts that changing an item’s serial position will shift content-cell phases while preserving the temporal reference provided by time cells.

  7. The system can utilize specialized neuron types, such as content cells that encode the item identity and their phases track the serial position of the item, to separate these two memory dimensions.

  8. The system can implement a circuit model where time cells form a traveling wave which provides a shared temporal reference for all memorized items, while content cells are phase-locked to the reactivation of a particular item.

Abstract

Oscillatory phase coding (OPC) is widely observed in neural systems and has been implicated in cognitive functions such as memory and sequence processing. However, its computational role and neural substrates remain largely unclear. Here, we studied OPC by training recurrent neural networks (RNNs) with adaptation to perform sequence working memory tasks with variable delays. After training, we found that OPC emerges in the RNN spontaneously: the memorized items recur at distinct phases of a shared rhythm, with oscillation amplitude encoding item identity and phase encoding serial order. Ablation studies identified three key elements required for OPC: adaptation in neural dynamics, variable delay in memory retention, and flexibility in reading out information. Neuron-type analysis revealed the emergence of content and time cells in the trained RNN, and they process item identity and sequence order separately. Recurrent connectivity analysis revealed phase-organized connections among time cells and structured connections within digit-selective content-cell groups. Inspired by these findings, we proposed a neural circuit model which reproduces OPC through interactions between content units and an adaptive time ring. Together, this study reveals how OPC arises from network structures and task demands, offers insights into the implementation and computational mechanisms of OPC in biological systems, and inspires us to develop computational models for human-like cognitive functions.

Sources

Related papers