Working Memory with Polychronous Chains

summary

The gist

Working memory — the ability to store and recall precise temporal patterns of neural activity — remains an open challenge for spiking neural networks (SNNs) The paper proposes a recurrent SNN of…

In short

Researchers created a recurrent Spiking Neural Network (SNN) with 1024 neurons and 41 heterogeneous synaptic delays to model working memory. By training it using surrogate-gradient backpropagation through time, the network successfully stores and retrieves complex temporal patterns. The results show that this architecture effectively mimics biological hippocampal function, demonstrating that varied synaptic delays are an efficient way to build scalable working memory in neuromorphic systems.

Key concepts

Recurrent HD-SNN
This is a specific type of SNN where each neuron receives input from other neurons across multiple preceding time steps (D=41). The 'Heterogeneous Delay' means these connections have different, varied time lags. This structure allows the network to maintain a temporal context necessary for remembering sequences.
Spiking Motifs
Instead of storing single patterns, the memory is stored as a chain of overlapping short sequences called motifs. Each motif has a length equal to the delay (D=41) and uniquely predicts what will happen in the next time step, allowing for progressive retrieval.
Surrogate-Gradient Backpropagation Through Time
This is the training method used to teach the network. It simulates how learning happens over long sequences of spiking activity by using a simplified gradient calculation. This allows the network to learn complex temporal dependencies end-to-end, optimizing its performance against specific memory targets.

Terminology used across episodes

This episode discusses

The paper

Working Memory with Polychronous Chains · Read on arXiv

Institut de Neurosciences de la Timone · Aix Marseille University · CNRS

Working memory --- the ability to store and recall precise temporal patterns of activity --- remains an open challenge for spiking neural networks (SNNs). A recurrent SNN of N=1024 neurons is proposed, in which each synapse is equipped with D = 41 heterogeneous delays, parameterised as a weight tensor W in R N times N times D and trained end-to-end with surrogate-gradient backpropagation through time. Each stored motif is represented as a sequential chain of overlapping Spiking Motifs: contiguous context windows of length D that uniquely predict the activity at the next time step. A closed-form Hebbian initialisation, derived by deconvolving the LIF membrane response and targeting a membrane potential value above the firing threshold, already achieves high F 1 scores before any gradient step on a benchmark of M=16 motifs of duration T=1000 steps. With learning, the network reaches a mean F 1 score of 0.80 over the motifs, and retrieval degrades gracefully when the trigger window is corrupted, with an F 1 score of 0.53 at 25% bit-flip noise. The network shows attractor-like retrieval dynamics consistent with hippocampal motif completion and with low-dimensional collective dynamics of population activity on manifolds. These results demonstrate that heterogeneous synaptic delays are an efficient and scalable substrate for working memory in SNNs, with direct implications for neuromorphic edge deployment.

Transcript

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

Ines: I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "Working Memory with Polychronous Chains".

Marcus: Here's the gist of the paper. Working memory — the ability to store and recall precise temporal patterns of neural activity — remains an open challenge for spiking neural networks (SNNs) The paper proposes a recurrent SNN of N = 1024 neurons in which each synapse is equipped with D = 41 heterogeneous delays, parameterised as a weight tensor W ∈ R N×N×D and trained end-to-end with surrogate-gradient backpropagation through time Working Memory Model The core concept involves representing each stored pattern as a sequential chain of overlapping Spiking Motifs: contiguous context windows of length D that uniquely predict the activity at the next time step The network is designed to retrieve memory progressively, where each motif of length D allows to predict the next spikes in the memory to retrieve This allows the network to generate a chain of overlapping Spiking Motifs, where each motif provides the context for predicting the next spikes in the sequence The architecture is a recurrent Heterogeneous-Delay SNN (HD-SNN) which consists of a population A of N = 1024 leaky integrate-and-fire (LIF) neurons implemented in snnTorch Each neuron j ∈ A integrates inputs from all other neurons across all preceding D = 41 time steps The spike history of all N neurons over the past D steps is concatenated into a flattened context vector of dimension N · D at each time step Training and Initialization The network is trained end-to-end with surrogate gradients

19: using a fast-sigmoid surrogate with sharpness parameter α = 15 The loss is defined as L = 1 − F1, such that minimizing this loss (with a maximum of 1 and minimum of 0) maximises the mean F1 score across all M target patterns The analytical weight initialisation alone (Eq.

Ines: 5) already achieves F1 = 1.0 before any gradient step, demonstrating that the closed-form Hebbian solution is sufficient for perfect recall on this benchmark The Gram-matrix approximation CC⊤ ≈ N · D · pA · I reduces the pseudo-inverse to a simple crosscorrelation, yielding a ≈5× speed-up with identical accuracy Performance Analysis and Robustness The network consistently maximised the mean F1 score across all M = 16 target patterns to a perfect score of 1.0, with N = 1024 neurons, D = 41 delays, T = 1000 time steps, pA = 10−3, and β = 0.7 Robustness to spike noise (Figure 4, left) shows that at pflip = 0.25, the network still achieved precision = 1.000, recall = 0.937, and F1 = 0.967 Robustness to trigger-window duration (Figure 4, middle) indicates that at 75% of the full trigger duration (≈ 30 steps out of D = 41), the network achieved precision = 1.000, recall = 0.757, and F1 = 0.862 Robustness to partial neuron coverage (Figure 4, right) shows that with N/8 = 128 neurons silenced — meaning 87.5% of the population remained active — the network still achieved perfect recall (F1 = 1.000) Capacity and Scaling Limits The memory limit is determined by parameter count rather than interference, and the capacity scales as N2 · D, linear in the delay depth The loss decreases monotonically as D grows from 3 to 127 ms, confirming that deeper delay windows provide a larger, more orthogonal context and that longer maximum delays directly increase the capacity of the recurrent HD-SNN Pattern duration T (Figure 5, middle) shows that the loss grows monotonically with T, from L ≈ 0.004 at T = 64 to L ≈ 0.08 at T = 2048 This is the empirical signature of the two compounding difficulties identified in the capacity analysis: longer sequences impose a larger open-loop rollout, and the error propagation recurrence accumulates over T −D steps Future Directions The paper suggests that introducing adaptive threshold dynamics or more complex single-neuron models, as studied in Szatmáry and Izhikevich

25: , should increase the diversity of intrinsic timescales available to the network and improve both capacity and retrieval fidelity Partitioning the population into inputoutput neurons and a larger pool of latent “reservoir” neurons, as in delay-based reservoir computing

20: , could enrich the internal spiking dynamics without increasing the number of trained parameters proportionally A natural extension is an online learning rule that updates the weights continuously as new patterns are presented, without catastrophic forgetting of previously stored sequences The most immediate application is to replace the supervised F1 loss with a self-supervised or contrastive objective that trains the network to predict upcoming spikes from the recent spike history, without access to ground-truth labels Discussion on Heterogeneous Delays The compression of the entire pattern is made possible precisely by the temporal structure introduced by the D delay channels: rather than requiring one parameter per stored bit, the network exploits the sequential dependencies between consecutive context windows of length D to represent the entire pattern as a chain of overlapping Spiking Motifs This provides concrete evidence that heterogeneous delays are a computational asset rather than a biological nuisance The robustness experiments further strengthen this interpretation, showing that the network tolerates up to 25% bit-flip noise in the trigger window This progressive degradation profile is consistent with the network operating as an attractor Acknowledgements This work was performed using HPC resources from GENCI-IDRIS (Grant 2025–AD010314955R2) The authors declare no competing interests regarding this manuscript The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results For the purpose of open access, the authors have applied a CC-BY public copyright licence to any Author-Accepted Manuscript version arising from this submission The implementation is available at https://github.com/laurentperrinet/MNESIS The recurrent HD-SNN was implemented using the snnTorch

8: library with custom extensions for heterogeneous delay handling The implementation leverages PyTorch with the Metal Performance Shader (MPS) backend for accelerated computation on Apple Silicon (M3 Ultra chip featuring 32 CPU cores and 80 GPU cores, with 512 GB of unified memory architecture) and further tested with the CUDA backend on the Jean Zay supercluster at GENCI Training typically converged within 128 gradient steps, requiring approximately 5 minutes of computation time per training The loss is defined as L = 1 − F1, such that minizing this loss (with a maximum of 1 and minimum of 0) maximises the mean F1 score across all M target patterns The cosine learning-rate schedule

13: helped guide the network away from these extremes towards a balanced, sparse-firing regime Dropout regularization with probability 0.

Paper summary: Ines: So, we're talking about this paper "Working Memory with Polychronous Chains," which proposes a recurrent SNN of one thousand twenty-four neurons where every synapse has forty-one different time delays <ref:2604.14096#pg1>. The main idea is that memory isn't stored as individual bits, but as a chain of overlapping spiking motifs where each motif has the length D equals forty-one steps and it uniquely predicts what happens next > <ref:2604.14096#pg1,chain of overlapping spiking motifs>

Marcus: That structure, this recurrent heterogeneous delay SNN, they claim can reach a mean F1 score of zero point eight over sixteen target patterns when the memory duration is one thousand steps >

Yuki: They start with a closed-form Hebbian initialization derived from deconvolving the LIF membrane response to get started, and this initialization already achieves an accuracy as measured by the F1-score relative close to that expected when noise is present <ref:2604.14096#pg1> <ref:2604.14096#pg1>

Ines: What this really shows is that heterogeneous synaptic delays are an efficient way to build working memory in SNNs, moving beyond needing one parameter per stored bit because they exploit those sequential dependencies between the context windows >

Marcus: The performance analysis shows the network hits a mean F1 score of zero point eight over all sixteen motifs, and when you add twenty-five percent bit-flip noise, it degrades gracefully down to an F1 score of zero point five three at that level <ref:2604.14096#pg1> <ref:2604.14096#pg1>

Yuki: It’s interesting to see how they connect this work to the broader literature on sequence processing, like the observations Ikegaya et al. made about recurring spatiotemporal spike sequences in the neocortex >

Ines: They argue that these results demonstrate attractor-like retrieval dynamics that are consistent with hippocampal motif completion and also with low-dimensional collective dynamics of population activity on manifolds >

Marcus: The paper suggests we should consider introducing adaptive threshold dynamics or more complex single-neuron models to increase the diversity of intrinsic timescales available to the network, which would improve both capacity and retrieval fidelity <ref:2604.14096#pg1> <ref:2604.14096#pg1>

Yuki: And they mention partitioning the population into input-output neurons and a larger pool of latent reservoir neurons as in delay-based reservoir computing, which could enrich internal spiking dynamics without increasing the number of trained parameters proportionally <ref:2604.14096#pg1> <ref:2604.14096#pg1>

Ines: The immediate application they point to is using a self-supervised or contrastive objective instead of the supervised F1 loss to train the network to predict upcoming spikes from recent spike history, without needing ground truth labels <ref:2604.14096#pg1> <ref:2604.14096#pg1>

Marcus: So, looking at "Working Memory with Polychronous Chains," they are showing that this recurrent HD-SNN structure is a viable substrate for working memory in SNNs by using those heterogeneous delays to compress the pattern into a chain of overlapping motifs >

Ines: And the authors conclude that this provides concrete evidence that these heterogeneous delays are a computational asset rather than just being some biological nuisance >

Yuki: From my perspective, what this means is that we need to focus on how these precise temporal dependencies are encoded and retrieved in the cortex because the paper suggests those relationships are fundamental to sequence processing <ref:2604.14096#pg3> <ref:2604.14096#pg3>

Marcus: Exactly. They’re not just saying it works; they’re showing a method where capacity scales linearly with delay depth, which is a strong structural finding for how this system can handle memory loads >

Ines: And for us, the implication is that we should look at building models that incorporate these precise relative timings across a population of neurons in a way that allows them to manage those temporal dependencies efficiently <ref:2604.14096#pg1> <ref:2604.14096#pg1>

Yuki: So, ultimately, this paper suggests we need more complex models of single neurons or maybe even reservoir computing approaches to enrich the internal spiking dynamics without just training massive weight tensors <ref:2604.14096#pg3> <ref:2604.14096#pg3>

Conclusion: Ines: So, we're wrapping up this look at "Working Memory with Polychronous Chains," which is basically about how spiking neural networks can handle working memory using these heterogeneous delays >

Marcus: Right, it’s about taking a big memory problem and solving it by making every synapse have different time lags so the network stores a whole sequence in one go >

Yuki: The authors are using a recurrent SNN with one thousand twenty-four neurons and forty-one different delays, and they show that this setup can actually learn to recall sequences quite well, hitting an F1 score of zero point eight over sixteen different memory patterns >

Ines: They're arguing that those delays aren't just random noise; they are a computational asset because they let the network represent the whole pattern as a chain of overlapping motifs >

Marcus: And when you test it, like adding twenty-five percent bit-flip noise to the trigger window, it doesn't crash; it still maintains an F1 score of zero point five three >

Yuki: That suggests that if we look at how memory works in real species, like the hippocampus, these kinds of sequential dependencies are really what allow for that kind of recall mechanism >

Ines: So, the authors conclude that this structure offers a scalable way to build working memory in SNNs without needing an insane number of parameters for every single bit >

Marcus: Exactly. It’s about showing how complexity can be managed through temporal structure rather than just brute force parameter counts >

Yuki: And what this means for us is that we need to look at those precise timing relationships across a population of neurons if we want to model sequence processing in biological systems accurately >

More episodes

← Home