{"id":"706d954d-6b37-4ad5-94a5-feaef9861b87","arxiv_id":"2507.08320","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"NeurOptimiser uses populations of spiking neurons to run heuristic search, solving BBOB benchmark functions up to 40 dimensions with estimated low power consumption.","lead":"The authors present NeurOptimiser, a framework that runs evolutionary search through spiking neuron dynamics on Intel's Lava platform. It is a proof of concept that spike-triggered heuristics can optimize standard benchmark functions, with estimated power consumption below conventional CPUs at small scale.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is not supported by the implemented architecture: Section 4.3's High-Level Selector and Neighbour Manager are central, non-spiking coordinators, so 'fully spike-based' and 'fully decentralised' fail even before considering power.","rationale":"Chose this over the power-model concern because the strongest claim is about the paradigm (fully spike-based, fully decentralised), and the architecture itself contradicts it. The power analysis is an important secondary issue, but even a measured Loihi power number would not rescue the claim if the system's search and coordination are actually performed by central CPU processes on real-valued data. The reader's verdict is conditional and already mentions these issues, so I do not propose changing the verdict; re-scoping the claims and/or demonstrating a real Loihi mapping would bring the paper in line with its evidence. I agree only partially with the reader's identified weakest assumption because the reader singled out energy/timing, whereas I find the architectural mismatch more load-bearing.","tokens_in":27150,"tokens_out":8417,"duration_ms":108182,"concrete_test":"Run the released Lava code (zenodo 15858610) with the High-Level Selector and Neighbour Manager disabled, replacing global-best broadcast and neighbourhood aggregation with purely local spike-based interactions, and measure BBOB convergence. Separately instrument the process-level message traffic to count bytes exchanged through PPP/fff_p versus through SSS/AAA. If convergence collapses without the central coordinator, or if most information flows through real-valued arrays, the 'fully decentralised, fully spike-based' claim is falsified. A second decisive check is to map Algorithms 6 and 8 onto Loihi 2 without host-side CPU processes; if they cannot be mapped, the framework as described is not a neuromorphic implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section 7) is a 'fully spike-based', 'fully decentralised' optimisation framework whose coordination arises 'without external orchestration'. The implementation in Section 4 contradicts this. Algorithm 6 (High-Level Selector) receives every NHU's PPP and f_p, computes argmin over all units, and broadcasts global best ggg as a real-valued vector; Algorithm 8 (Neighbour Manager) centrally builds and redistributes neighbourhood matrices P_n and F_n. Algorithm 2 (Selector) evaluates f on the CPU and maintains floating-point ppp, and Sender/Receiver (Algorithms 4-5) exchange real-valued position and fitness arrays. The DE/current-to-rand spike-triggered rule (Eq. 21b) is applied to real-valued positions; spikes are event triggers around CPU heuristics, not the substrate of the search. The paper itself states in Section 6 that all experiments were CPU simulations with Loihi only 'targeted'. Thus the load-bearing condition that spiking dynamics serve as the search engine and that coordination is native/decentralised is not established by the described system; this is an internal inconsistency in the strongest claim, not a matter of benchmarking consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces NeurOptimiser, a framework that combines spiking neuron models with metaheuristic search. A population of Neuromorphic Heuristic Units (NHUs) encodes candidate solutions into neuron states, applies a dynamic rule h_d and a spike-triggered rule h_s (fixed reset or DE/current-to-rand/1), and communicates through a spike matrix, a Neighbour Manager, and a High-Level Selector. The authors implement the system in Intel's Lava framework with Loihi 2 as target, evaluate on noiseless BBOB functions in 2--40 dimensions against BBOB 2009 and RANDOMSEARCH, and estimate power consumption from Loihi per-event costs. They claim this is a fully spike-based, fully decentralised, low-energy optimisation framework.","tokens_in":27507,"tokens_out":6370,"duration_ms":69635,"significance":"The paper is a useful proof of concept: it shows that spike-triggered heuristic rules can drive a population-based search on standard benchmark functions, it ships reproducible code and data, and it benchmarks against external references rather than fitted baselines. The internal-dynamics visualisations (Figs. 6--9) are a strength. However, the significance as claimed in the abstract and conclusion ('fully spike-based', 'fully decentralised', 'milliwatt-level') is not supported by the described implementation, since several coordinating components operate on real-valued position/fitness matrices and all results come from CPU simulation rather than Loihi 2 deployment. The contribution is better framed as a CPU-simulated, spike-triggered heuristic framework with a Loihi target.","major_comments":[{"comment":"The claim that coordination arises 'without external orchestration' (Section 7) is contradicted by the architecture. Algorithm 6 collects every NHU's PPP and f_p, performs a global argmin over all units, and broadcasts the real-valued global best ggg; Algorithm 8 centrally builds and redistributes neighbourhood matrices P_n and F_n from all units. These are central coordination processes operating on floating-point data, not native spiking mechanisms.","section":"Section 4.3 (Algorithms 6 and 8)"},{"comment":"The 'fully spike-based' claim is not supported: candidate positions are real-valued; the Selector evaluates f on the host CPU and maintains ppp in floating point; Sender/Receiver exchange real-valued position and fitness arrays; and the DE mutation in Eq. (21b) is an arithmetic update on real-valued state vectors, with spikes only determining whether the rule fires. Spikes act as event triggers around CPU heuristics rather than as the substrate of the search, so the strongest formulation in Section 7 ('spiking dynamics ... search engine') is an overstatement.","section":"Section 4.2 (Algorithms 2, 4, 5) and Eq. (21b)"},{"comment":"The power feasibility claim is based on an analytical estimate, not a measurement. The formula E_step = 23.6 N_syn + 89.7 n pJ multiplies Loihi per-event costs from [41] by the algorithm's event count and divides by an assumed Delta_t_sim = 0.5 ms; no Loihi 2 chip was used. Moreover, the worst-case estimate reported in Figure 13 is 1.35 W, which contradicts the 'milliwatt-level' wording in the abstract and Section 7. The data support sub-watt estimates for small configurations, but not the stated milliwatt-level feasibility claim.","section":"Section 6 (Figure 13)"},{"comment":"All experiments are CPU simulations; Loihi 2 is only 'targeted'. Statements in Section 6 ('we anticipate further gains') and Section 7 ('suitability ... for practical NC hardware deployments') go beyond the evidence. The authors should confine conclusions to spike-triggered heuristics in simulation and describe the Loihi deployment as future work, or add a measured deployment.","section":"Sections 5 and 6"}],"minor_comments":[{"comment":"Two equations are listed but the text says they correspond to three strategies ('current-to-best, rand-to-best, and current-to-rand'); align the number of equations with the description.","section":"Section 3.3, Eqs. (21a)--(21b)"},{"comment":"The spike-triggered rule h_s in Eq. (13) appears to return a complex scalar ϑ1 + ϑ2 r1 e^{i2πr2}, while the state space V is real and two-dimensional; clarify the intended construction, for example as a vector in polar coordinates.","section":"Section 3.2, Eq. (13)"},{"comment":"The first experiment is described as using 1000×d steps, but the captions of Figures 6--9 say 1000 steps; reconcile this inconsistency.","section":"Section 5 vs. Figures 6--9"},{"comment":"The random value r introduced in the encoding must be retained for deterministic invertibility; specify where this per-step random value is stored and how T^{-1} accesses it, since the algorithm as written does not keep r in the state.","section":"Section 3.4, Eq. (22)"},{"comment":"There is a typo ('Averate') in the caption of Figure 13, and the caption does not clearly map the parenthetical power values to the (n, d) settings, making the figure harder to read.","section":"Figure 13 caption"},{"comment":"The header row for f1 appears garbled ('f111 12 12 ...') in the preprint; ensure the table renders correctly.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is a genuine engineering contribution, but its three headline claims — fully spike-based, fully decentralised, milliwatt-level — are not supported by the implementation described in Section 4. The good news is that the gap is fixable: the framework itself is reusable and the experiments are honest.\n\nWhat is actually new: the modular design of Neuromorphic Heuristic Units with plug-in dynamics h_d and spike-triggered heuristics h_s, a bidirectional mapping between problem space and neuromorphic space, and a full Lava implementation evaluated on BBOB up to 40D. That is the first concrete general instantiation of Talbi's Nheuristics concept, and it goes beyond the problem-specific QUBO/TSP spiking solvers and CPU-driven LavaBO. The public code and data are a real plus. The characterization experiments (phase portraits, spike activity, population dynamics) are thoughtful, and the paper does not hide that performance is far from state of the art at 20D/40D.\n\nSoft spots, in order of severity. The claim that coordination arises 'without external orchestration' and that the system is 'fully spike-based' is contradicted by the architecture itself. Algorithm 6 (High-Level Selector) gathers every unit's p and f_p, computes argmin, and broadcasts a real-valued global best; Algorithm 8 (Neighbour Manager) centrally builds and redistributes real-valued neighbourhood matrices; the Selector evaluates f on the CPU. Spikes trigger events, but the search state and the global coordination are ordinary floating-point data. This is not a matter of benchmarking choice; the abstract and conclusion overstate what the described system does. The power estimates are also analytical: Loihi per-event costs from the literature multiplied by event counts, with a worst case of 1.35 W, which does not match 'milliwatt-level' in the abstract. No Loihi measurement is reported. The confirmatory variants were selected after the initial screen, so the comparison has a mild selection bias; not fatal, but worth noting.\n\nIf the authors re-scope the claims — replace 'fully spike-based' with 'spike-triggered heuristics in a Lava-based framework', and present power as an estimate with the worst-case number — the contribution is solid. The experiments do support the weaker claim that spike-triggered heuristics can search low-dimensional BBOB functions with structured population dynamics.\n\nWho this is for: people working on neuromorphic optimisation and edge computing; also anyone building on the Nheuristics idea. It deserves a serious referee, but the referee should require the re-scoping and a clearer statement that the system is not yet decentralised in the strong sense. I would send it to review, not desk-reject.","headline":"Real engineering contribution with public code, but the central claims about full spike-basing, decentralisation, and milliwatt power are ahead of the implemented architecture; re-scope and it is worth engaging seriously.","tokens_in":28010,"tokens_out":2695,"would_cite":true,"duration_ms":29455,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that NeurOptimiser, a fully spike-based, asynchronous framework, solves BBOB problems up to 40 dimensions using coordination that emerges from local spiking events rather than a central program, with milliwatt-scale power…","keywords":["Neuromorphic Computing","Event-Driven Asynchronous Optimisation","Neuromorphic Metaheuristics","Spiking Neural Networks","Evolutionary Algorithms","Differential Evolution","BBOB benchmark","Loihi 2"],"falsifier":"Run the same NeurOptimiser configurations on a Loihi 2 chip and measure energy per optimisation step; if measured power for the n=90, d=40, m=89 worst case is not in the roughly watt-level range of the estimate, the power-feasibility claim fails. A cheaper check is to instrument the CPU simulation to record total emitted spikes and compare that count with the n(n-1)md synaptic-event formula used in the estimate.","tokens_in":26961,"feed_emoji":"⚡","tokens_out":7999,"duration_ms":87483,"temperature":0.7,"pith_summary":"The paper tries to show that a population-based optimiser can be built entirely from spiking neuron events, with no central clock or CPU orchestrator driving the search. Its NeurOptimiser encodes candidate solutions into the internal states of spiking neuron units; when a unit's membrane potential crosses a threshold, the resulting spike triggers a perturbation heuristic, and those spikes propagate to neighbours to share promising positions. The authors implement this in a neuromorphic software framework targeting the Loihi 2 chip and test it on the BBOB benchmark up to 40 dimensions. They report convergence that matches or beats random search in low and moderate dimensions and roughly matches reference optimisers on separable and well-conditioned problems, with milliwatt-level power estimates derived from Loihi event costs.","feed_headline":"Spiking neurons alone can run evolutionary search","feed_subtitle":"NeurOptimiser runs evolution entirely on spike events, with no central coordinator, and estimates milliwatt-level power on Loihi 2.","key_machinery":"The central object is the Neuromorphic Heuristic Unit (NHU): one spiking neuron per coordinate of a candidate solution, each following the generalised transition $v^{t+1} \\leftarrow h_d(v^t)$ if $\\Phi(v^t, \\ldots) \\neq 1$ and $v^{t+1} \\leftarrow h_s(v^t)$ otherwise, with $\\Phi$ the spiking condition, typically $|v_{1,j}| \\geq \\vartheta_j$. Here $h_d$ supplies the exploration dynamics (linear systems, LIF, or Izhikevich models), $h_s$ acts as the perturbation heuristic (stochastic reset, reset toward a best state, directional displacement, or differential-evolution mutation), and the bidirectional map $T$ translates between problem-space coordinates and neuromorphic states. The mechanism works because the spiking condition is simultaneously a selection predicate and a communication event: a spike both resets or perturbs the unit and propagates through a bulk spike-contraction layer to activate neighbouring units, so search and coordination are the same event.","core_discovery":"On its own terms, the paper's central claim is that the search operator of a metaheuristic can be realised by spiking neuron dynamics rather than by an external program. Each candidate solution is a set of $d$ neurons; their internal states evolve under a dynamic rule $h_d$, and when the spiking condition $\\Phi$ triggers, a spike-triggered rule $h_s$ applies a heuristic perturbation, such as a reset toward the best-so-far state or a differential-evolution-style mutation. Local and global coordination, including neighbourhood information, global best sharing, and spike propagation, emerge from asynchronous message passing encoded as spikes rather than from a central loop. The authors argue that this is the first complete and reproducible integration of neuromorphic heuristic-based optimisers, and they report that on BBOB up to 40 dimensions the heterogeneous variants converge reliably on separable, unimodal, and moderately conditioned problems, with runtime linear in population size and dimension, per-unit steps under 12 ms, and power estimates from milliwatts to about 1.35 W in the tested worst case.","pith_inferences":["Editorial inference: the paper attributes the high-dimensional ECDF plateaus to constrained evaluation budgets rather than algorithmic stagnation; a direct test would run the same variants with budgets comparable to the BBOB 2009 reference and check whether the plateaus lift.","The analytical power model assumes worst-case dense connectivity, so a testable extension is to measure actual emitted spikes on the hardware or in simulation and recompute the per-step energy with measured synaptic-event counts, which could bring the estimate below the milliwatt levels already reported.","A natural next step the paper leaves implicit is adaptive spike-threshold tuning tied to local convergence, which could reduce superfluous spiking and further lower the energy estimate.","The framework defines a design space for choosing $h_d$, $h_s$, and $\\Phi$; a promising untested direction is learning which neuron-model or mutation combination suits a given problem class, turning the NHU configuration itself into a search problem."],"forward_implications":["If NeurOptimiser is right, general-purpose optimisation can run on event-driven neuromorphic hardware without a CPU orchestrator, making the same algorithm deployable in low-power embedded settings.","The spike-triggered rule is pluggable: both simple resets and differential-evolution mutations are expressible, so established metaheuristic operators can be ported onto spiking substrates rather than re-designed from scratch.","Coordination by spike propagation means communication cost scales with actual spiking activity rather than with a fixed synchronisation schedule, so sparse firing directly lowers energy consumption.","Heterogeneous populations that mix linear and Izhikevich neuron dynamics proved more robust than homogeneous ones on multimodal and ill-conditioned BBOB problems, pointing to neuron-model diversity as a search resource.","Runtime and resource usage scale linearly with population size and dimension, supporting the feasibility of large asynchronous populations on neuromorphic hardware."],"supporting_citations":[{"why":"Defines the Nheuristics class that NeurOptimiser instantiates as its formal basis.","marker":"[7]"},{"why":"Supplies the neuromorphic software framework in which NeurOptimiser is implemented and that targets the Loihi 2 chip.","marker":"[8]"},{"why":"Provides the BBOB/COCO benchmark platform and evaluation protocol used for all experiments.","marker":"[9]"},{"why":"Provides the differential-evolution mutation taxonomy that inspires the current-to-rand/1 perturbation used as $h_s$.","marker":"[28]"},{"why":"Supplies the spiking neuron model background, including LIF and Izhikevich models, used for $h_d$.","marker":"[29]"},{"why":"Defines the Izhikevich model parameters and dynamical regimes reproduced in the experiments.","marker":"[31]"},{"why":"Supplies the tensor contraction operation used to convert spike matrices into activation signals.","marker":"[36]"},{"why":"Provides the pure random-search baseline that NeurOptimiser must beat in the benchmark comparisons.","marker":"[40]"},{"why":"Provides the Loihi per-event energy costs and spikes-per-event assumptions underlying the power estimates.","marker":"[41]"}],"fun_headline_variants":["Spikes alone run evolutionary search","Spike-based evolution without a central brain","Milliwatt evolution on Loihi 2, all spike-driven","Neurons replace the optimisation loop entirely"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that counting events in a CPU simulation and pricing each one with the Loihi chip's published per-event energies gives a realistic picture of real neuromorphic hardware; if real spike counts or per-event costs are substantially different, the milliwatt feasibility claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Spikes alone run evolutionary search","Spike-based evolution without a central brain","Milliwatt evolution on Loihi 2, all spike-driven","Neurons replace the optimisation loop entirely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001203,"raw_usage":{"total_tokens":5001,"prompt_tokens":1035,"completion_tokens":3966,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":3908}},"tokens_in":651,"tokens_out":3966,"duration_ms":28279,"temperature":1.0,"reasoning_tokens":3908,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:23:25.819724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same NeurOptimiser configurations on a Loihi 2 chip and measure energy per optimisation step; if measured power for the n=90, d=40, m=89 worst case is not in the roughly watt-level range of the estimate, the power-feasibility claim fails. A cheaper check is to instrument the CPU simulation to record total emitted spikes and compare that count with the n(n-1)md synaptic-event formula used in the estimate.","supporting_citations":[{"cited_title":"Neuromorphic-based metaheuristics: A new generation of low power, low latency and small footprint optimization algorithms","cited_arxiv_id":"2505.16362","evidence_quote":"Defines the Nheuristics class that NeurOptimiser instantiates as its formal basis."},{"cited_title":"Lava: A software framework for neuromorphic computing,","cited_arxiv_id":null,"evidence_quote":"Supplies the neuromorphic software framework in which NeurOptimiser is implemented and that targets the Loihi 2 chip."},{"cited_title":"COCO: A platform for comparing continuous optimizers in a black-box setting,","cited_arxiv_id":null,"evidence_quote":"Provides the BBOB/COCO benchmark platform and evaluation protocol used for all experiments."},{"cited_title":"Differential evolution mutations: Taxonomy, comparison and convergence analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the differential-evolution mutation taxonomy that inspires the current-to-rand/1 perturbation used as $h_s$."},{"cited_title":"Spiking neural networks and their applications: A review,","cited_arxiv_id":null,"evidence_quote":"Supplies the spiking neuron model background, including LIF and Izhikevich models, used for $h_d$."},{"cited_title":"Simple model of spiking neurons,","cited_arxiv_id":null,"evidence_quote":"Defines the Izhikevich model parameters and dynamical regimes reproduced in the experiments."},{"cited_title":"Tensor Contraction Layers for Parsimo- nious Deep Nets,","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor contraction operation used to convert spike matrices into activation signals."},{"cited_title":"Benchmarking the pure random search on the bbob-2009 testbed,","cited_arxiv_id":null,"evidence_quote":"Provides the pure random-search baseline that NeurOptimiser must beat in the benchmark comparisons."},{"cited_title":"Loihi: A neuromorphic manycore processor with on-chip learning,","cited_arxiv_id":null,"evidence_quote":"Provides the Loihi per-event energy costs and spikes-per-event assumptions underlying the power estimates."}],"review_version":1}