Ponulak and Hopfield (2013) proposed a hippocampus-like network of spiking place cells that solves path planning in a single parallel event: a wavefront of spiking activity initiated at a target propagates through a previously learned cognitive map while temporally asymmetric anti-STDP ('reverse' STDP) writes a transient synaptic vector field (SVF) whose field lines converge toward the target; the agent then follows this field to the goal, after which all planning synapses are reset to baseline before the next planning episode. The paper demonstrated the protocol only by simulation and explicitly left its network-level properties unquantified, noting that a statistical-mechanics analysis (dynamics, storage capacity, phase boundaries as functions of network size, activity level, noise, and place-field size) existed only for the same-topology storage network of Hopfield (2010), analyzed by Monasson and Rosay (2013). The open problem is to derive that missing quantitative theory for the planning protocol itself: the conditions under which wavefront propagation remains stable (the wave reaches every cell of the connected map with a single spike and does not self-sustain) and the written SVF faithfully encodes shortest-path information (following it from any cell yields a monotone decreasing distance-to-target trajectory that terminates at the target); the storage capacity and phase boundaries as functions of network size, activity level, noise, place-field size, and number of stored environments, including the effect of repeated write/erase planning cycles on multi-map storage; and the tradeoffs among these parameters, with validation against simulation of the original model.

Background

Efficient path planning — finding a shortest or optimal route through a known environment — is a computational problem shared by animals and robots. In the rodent brain, spatial memory is supported by the hippocampus, whose 'place cells' fire whenever the animal enters a specific spatial location, so that a population of place cells forms an internal map of an environment. Classical algorithmic solutions propagate an expanding 'wavefront' of distance labels outward from the goal (equivalent to Dijkstra or A* search), then read the route back by steepest descent; Ponulak and Hopfield (2013) asked how this computation could be carried out rapidly and in parallel by biologically plausible spiking dynamics.

The substrate is a recurrent network of place cells modeled by adapting leaky integrate-and-fire neurons (a standard spiking-neuron model in which the membrane potential integrates input and decays toward rest, with spike-frequency adaptation), of the type introduced by Hopfield (2010) for 'mental exploration'. During an exploration phase, the agent learns the map: pairs of place cells whose fields are adjacent along feasible trajectories become connected by spike-timing-dependent plasticity (STDP), the biologically observed rule in which the sign and magnitude of synaptic change depend on the relative timing of pre- and post-synaptic spikes. Planning then works as follows. Exciting the place cell(s) at a chosen target launches an expanding wavefront of single-spike activity through the network. While the wavefront passes, the polarity of plasticity is reversed — 'anti-' or 'reverse-' STDP, controlled by a parameter d∈{−1,0,1}d \in \{-1, 0, 1\} in the learning rule (with dd linked by the authors to neuromodulators, which are known experimentally to flip the sign of STDP; Seol et al., 2007). The reverse-STDP modification left by the passage of the wave writes a 'synaptic vector field' (SVF): a pattern of directed synaptic asymmetries whose field lines, by construction of an expanding circular wave, converge when traced backward toward the wave's origin — that is, toward the target. The agent then moves by letting its sensory location and the SVF-biased activity align, and the vector field stably guides it to the goal. The convergent geometry is why anti-STDP is preferred over normal STDP: converging field lines attenuate neural noise, whereas diverging ones amplify it, and convergence additionally permits mental rehearsal of the route from the current location. Whenever a new planning episode is needed, all synapses are reset to the baseline state. The network can store several environments (multiple maps in overlapping place-cell populations) and can select among simultaneously presented targets of different value in one parallel wave.

The paper demonstrated all of this by simulation: it did not attempt to quantify the results. For the storage dynamics of the same type and topology of network (without the planning protocol), a quantitative theory already existed: Monasson and Rosay (2013) applied statistical-mechanics mean-field tools (approximations that replace the interacting network by self-consistent averages) to a binary-unit attractor network storing multiple spatial maps (an attractor network being a recurrent network whose dynamics settles into self-sustaining activity patterns, here one localized activity pattern per stored map) and derived the storage capacity and the phase diagram as functions of network size, activity level, response noise, and place-field size — finding localized 'clump' activity states below critical noise and load, crosstalk-driven delocalized states beyond them, capacity of order 10−310^{-3} bits per synapse for 2D maps, and the notable effect that moderate noise slightly increases storage capacity; an earlier capacity analysis for a similar topology is Battaglia and Treves (1998). This theory, however, covers the Hebbian storage and retrieval of maps — not the transient, out-of-equilibrium planning protocol (wavefront propagation with anti-STDP writing, full synaptic reset between episodes, multi-map retrieval under repeated planning), whose operating limits were therefore left entirely open. On the empirical side, awake sharp-wave ripples (SWRs — brief high-frequency population bursts in the hippocampus at rest) replay place-cell sequences depicting paths toward remembered goals (Pfeiffer and Foster, 2013), making them a candidate biological substrate for the model's waves; but the model's computational claims stand so far on simulation alone.

Problem Statement

Derive an analytical, quantitative characterization of the network-level properties of the Ponulak–Hopfield wavefront path-planning protocol — the spiking place-cell network of Hopfield (2010) type operated with the planning schedule of Ponulak & Hopfield (2013) (map storage by exploration-driven STDP; target-initiated expanding wavefront of spiking activity during which the STDP polarity is reversed, d=−1d = -1, writing a synaptic vector field; reset of planning synapses to baseline after each episode, the exact reset of the source protocol (the repeated-cycle deliverable (ii) additionally uses a scoped modelling extension in which the reset has a finite erase fidelity ε∈[0,1]\varepsilon \in [0, 1] (ε=1\varepsilon = 1 being the source's exact reset), leaving a residual modification that can accumulate across episodes); multiple stored environments). The requested results, all three of which are jointly required to serve the single objective of a predictive theory of the protocol's operating regimes, are: (i) the conditions — as functions of network size, level of neural activity, spiking noise, place-field size, and number of stored environments — under which the target-triggered wavefront propagates stably (it reaches every cell of the connected map with a single spike and does not self-sustain) and the anti-STDP-written synaptic vector field faithfully encodes shortest-path (distance-to-target) information (following it from any cell yields a monotone decreasing distance-to-target trajectory that terminates at the target); (ii) the storage capacity and the phase boundaries of the protocol as functions of network size, level of neural activity, spiking noise, place-field size, and the erase fidelity ε\varepsilon, i.e., how many environments and how many successive write/erase planning episodes can be supported before crosstalk between stored maps or residual inter-episode modification accumulated across planning episodes destroys map retrieval or misdirects the vector field, including the transitions between well-localized, crosstalk-dominated, and noise-dominated regimes; and (iii) the tradeoffs among these parameters — the exchange relations between the wavefront-stability and SVF-fidelity conditions of (i) and the storage capacity and phase-boundary positions of (ii) as network size, activity level, noise, place-field size, and the erase fidelity ε\varepsilon are varied. The analysis should use a controlled approximation scheme, state its approximation regimes explicitly, and be validated against direct simulation of the original model (adapting leaky integrate-and-fire neurons, the STDP learning rule whose amplitude asymmetry Aji>AijA_{ji} > A_{ij} (source-paper notation) makes exploration-phase plasticity net-positive, the d∈{−1,0,1}d \in \{-1, 0, 1\} schedule, and the reset rule, as specified in the source paper's Methods). The repeated-cycle results of (ii), which rest on the scoped finite-erase-fidelity extension above, are instead validated against simulation of the source model extended by that finite-erase-fidelity reset. The biological-plausibility, motor-implementation, and engineering-extension questions raised in the same Discussion are adjacent open problems and are not part of this completion boundary.

References

  1. Ponulak, F., & Hopfield, J. J. (2013). Rapid, parallel path planning by propagating wavefronts of spiking neural activity. Frontiers in Computational Neuroscience 7:98. DOI 10.3389/fncom.2013.00098. https://doi.org/10.3389/fncom.2013.00098 (source paper)
  2. Hopfield, J. J. (2010). Neurodynamics of mental exploration. Proceedings of the National Academy of Sciences USA 107(4):1648–1653. DOI 10.1073/pnas.0913991107. https://doi.org/10.1073/pnas.0913991107
  3. Monasson, R., & Rosay, S. (2013). Cross-talk and transitions between multiple spatial maps in an attractor neural network model of the hippocampus: phase diagram (I). arXiv:1304.1457; published as Physical Review E 87:062813. DOI 10.1103/PhysRevE.87.062813. https://arxiv.org/abs/1304.1457
  4. Battaglia, F. P., & Treves, A. (1998). Attractor neural networks storing multiple space representations: a model for hippocampal place fields. Physical Review E 58(6):7738–7753. DOI 10.1103/PhysRevE.58.7738. https://doi.org/10.1103/PhysRevE.58.7738
  5. Seol, G. H., Ziburkus, J., Huang, S., Song, L., Kim, I. T., Takamiya, K., Huganir, R. L., Lee, H.-K., & Kirkwood, A. (2007). Neuromodulators control the polarity of spike-timing-dependent synaptic plasticity. Neuron 55(6):919–929. DOI 10.1016/j.neuron.2007.08.013. https://doi.org/10.1016/j.neuron.2007.08.013
  6. Pfeiffer, B. E., & Foster, D. J. (2013). Hippocampal place-cell sequences depict future paths to remembered goals. Nature 497:74–79. DOI 10.1038/nature12112. https://doi.org/10.1038/nature12112
  7. Zennir, M. N., Benmohammed, M., & Martinez, D. (2017). Robust path planning by propagating rhythmic spiking activity in a hippocampal network model. Biologically Inspired Cognitive Architectures 20:47–58. DOI 10.1016/j.bica.2017.02.001. https://doi.org/10.1016/j.bica.2017.02.001
  8. Khajeh-Alijani, A., Urbanczik, R., & Senn, W. (2015). Scale-free navigational planning by neuronal traveling waves. PLoS ONE 10(7):e0127269. DOI 10.1371/journal.pone.0127269. https://doi.org/10.1371/journal.pone.0127269
  9. Butler, K., & Cruz, L. (2025). Neuronal traveling waves form preferred pathways using synaptic plasticity. Journal of Computational Neuroscience 53(1):181–198 (first published online 27 December 2024). DOI 10.1007/s10827-024-00890-2. https://doi.org/10.1007/s10827-024-00890-2
  10. Kang, L., & DeWeese, M. R. (2019). Replay as wavefronts and theta sequences as bump oscillations in a grid cell attractor network. eLife 8:e46351. DOI 10.7554/eLife.46351. https://doi.org/10.7554/eLife.46351
  11. Zhong, W., Lu, Z., Schwab, D. J., & Murugan, A. (2020). Nonequilibrium statistical mechanics of continuous attractors. Neural Computation 32(6):1033–1068. DOI 10.1162/neco_a_01280. https://doi.org/10.1162/neco_a_01280
  12. Espino, H., Bain, R., & Krichmar, J. L. (2024). A rapid adapting and continual learning spiking neural network path planning algorithm for mobile robots. IEEE Robotics and Automation Letters 9(11):9542–9549. DOI 10.1109/LRA.2024.3457371. https://doi.org/10.1109/LRA.2024.3457371
  13. Chorney, W. (2025). Generalized wavefront path planning. In: 2025 3rd International Conference on Mechatronics, Control and Robotics (ICMCR), pp. 25–29. DOI 10.1109/ICMCR64890.2025.10963265. https://doi.org/10.1109/ICMCR64890.2025.10963265
  14. Storesund, S., Valset Aars, K., Dietrich, R., & Waniek, N. (2025). Predictive spike timing enables distributed shortest path computation in spiking neural networks. arXiv:2509.10077. https://arxiv.org/abs/2509.10077
  15. von Seeler, O., Offenberg, E. C., Michaelis, C., Kulvicius, T., Luboeinski, J., Lehr, A. B., & Tetzlaff, C. (2026). Road to scalability for efficient graph search on massively parallel neuromorphic hardware. arXiv:2606.28907. https://arxiv.org/abs/2606.28907
  16. Meessen, Ch. (2026). Soliton-like waves in a two-dimensional recurrent spiking neural network with weighted spike-timing-dependent plasticity. arXiv:2606.21432. https://arxiv.org/abs/2606.21432
  17. Widloski, J., & Foster, D. J. (2025). Replay without sharp wave ripples in a spatial memory task. Nature Communications 16:10287. DOI 10.1038/s41467-025-65181-5. https://doi.org/10.1038/s41467-025-65181-5