{"id":"f99aee01-e8d7-4e4e-9be4-45b5e518e6ee","arxiv_id":"2411.18272","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A proposed ReRAM synapse stores the eligibility trace of the e-prop learning rule as local temperature, with weight updates driven by temperature-dependent conductance change, but the stated thermal time constant is too short to accumulate the trace across the milli-second training frames.","lead":"This paper proposes a ReRAM-based thermal synapse where a tiny heater heats the memory cell and the temperature acts as the memory of recent activity, controlling how the connection strength updates during training. The authors simulate two learning tasks to argue the design is fast and energy-efficient, but the key physics, a heat-based memory that persists across a full training frame, conflicts with the devices' own stated nanosecond-to-microsecond cooling time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Thermal accumulation is internally inconsistent: Eq. (23) with δt=1 ms and τTH≈1 µs resets the trace each step, yet Fig. 11(d) reports accuracy saturating at that τTH; the central mechanism is not established.","rationale":"Section 3 and Appendix 7.3 define the eligibility trace as the accumulated temperature rise in the ReRAM, with Eq. (23) as the governing recurrence. The reported system-level simulations (Sec. 5.2) adopt δt=1 ms and claim τTH≈1 µs is sufficient (Fig. 11(d)). Substituting these numbers into Eq. (23) gives tPW/τTH ~ 10^3 (if tPW is the e-update duration) or complete decay between steps (if tPW<τTH, as Sec. 4 instructs). Either way the temperature state does not persist across the ~700 steps of a TIMIT dataframe, so the end-of-frame cooling makes the programming pulse encode only the most recent e(t). This is not a discrepancy with external consensus but an internal contradiction between the model equations and a headline simulation result. The reader's weakest_assumption correctly identifies this. The independent test is to reproduce Fig. 11(d) from Eq. (23) with the stated parameters; without code or a detailed algorithmic description, the reported saturation cannot be validated. Thus the central claim—that the device computes and stores eΣ locally—is not established, and the REJECT verdict is appropriate.","tokens_in":22194,"tokens_out":6263,"duration_ms":54632,"concrete_test":"Re-implement the TIMIT benchmark with the temperature state updated exactly per Eq. (23) using δt=1 ms, the τTH values of Fig. 11(d), and the e-update pulse width stated in Sec. 3.2; set the final weight update Δw = η·[T(U)−Tamb] (the temperature just before ϕW), rather than any hidden accumulator. Compare the resulting accuracy curve to Fig. 11(d). If the curve is flat and high at τTH=1 µs, the reported figure cannot have been produced from Eq. (23), and the authors should provide the actual update rule; if the curve collapses, the central claim is refuted by their own model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the ReRAM local temperature accumulates the eligibility trace over the U-step dataframe (Eq. 8). The paper's own thermal model, Eq. (23), T'(t)=T'(t−1)+(tPW/τTH)[PH(t)RTH−T'(t−1)], is a first-order relaxation with time constant τTH. With the stated simulation step δt=1 ms (Appendix 7.2) and the claimed sufficient τTH≈1 µs (Sec. 5.2, Fig. 11(d)), the relaxation ratio tPW/τTH is huge if tPW≈δt, so T' jumps to the instantaneous steady-state PH RTH; conversely, if tPW<τTH as the paper recommends in Sec. 4, then because δt/τTH=1000, the temperature decays to ambient before the next e-update. In either reading, the accumulated eΣ is lost: the weight-update pulse sees only the last step's eligibility. Fig. 11(d) shows no such collapse, so either the simulation does not implement Eq. (23) faithfully, or the reported τTH is not the physical thermal time constant used in the device model. The paper does not provide code or data, and the integrated heater+ReRAM was not fabricated (Sec. 3.1 measured only standalone devices at elevated ambient temperature). The concept is creative and the ΔG(T) measurements are useful, but the core mechanism as parameterized cannot perform the stated accumulation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'thermal neoHebbian synapse' in which a ReRAM device stores the synaptic weight in its conductance and encodes the eligibility trace e(t) in the local temperature of the ReRAM. A resistive heater is 3D-integrated with the ReRAM; during the e-update phase, voltage signals representing the pre-synaptic activity f(t) and post-synaptic pseudo-gradient ψ(t) are applied across the heater, producing Joule heating and a local temperature rise proportional to f(t)×ψ(t). The temperature is claimed to accumulate over the presentation of a U-step dataframe, and at the end of the frame a fixed-amplitude programming pulse produces a conductance change proportional to the accumulated temperature, implementing Δw = η·eΣ. The paper presents measurements of temperature-dependent ReRAM switching (Sec. 3.1), COMSOL electrothermal simulations of the coupled heater–ReRAM structure (Sec. 4), and system-level simulations of e-prop training on TIMIT phoneme classification and of reinforcement learning in a maze (Sec. 5). The authors report that the proposed synapse trains close to ideal software synapses despite thermal decay, crosstalk, and device variability.","tokens_in":22433,"tokens_out":6673,"duration_ms":63037,"significance":"If the thermal accumulation mechanism worked as described, the proposed synapse would be an elegant and compact way to implement online three-factor learning (e-prop) in neuromorphic hardware, with a small per-synapse footprint and low write energy. The experimental data on temperature-dependent SET/RESET switching of ReRAM (Fig. 3) and the COMSOL thermal-coupling analysis (Fig. 6) are useful contributions in their own right. However, the central claim of the paper—that the ReRAM local temperature can accumulate the eligibility trace over the dataframe and retain it until the weight update—is undermined by the paper's own model equations and parameter choices, and the integrated device was not fabricated. The creativity of the concept and the quality of the temperature-dependent switching measurements do not compensate for the lack of a demonstration that the core mechanism actually works as claimed.","major_comments":[{"comment":"The central accumulation premise is inconsistent with the paper's own thermal model. Equation (23) is a first-order relaxation with time constant τTH; with the stated simulation step δt = 1 ms (Appendix 7.2) and the claimed sufficient τTH ≈ 1 µs (Sec. 5.2, Fig. 11d), the ratio tPW/τTH is either 1000 (if tPW ≈ δt) or, if tPW < τTH as recommended in Sec. 4, the temperature decays to ambient between successive e-updates because δt/τTH = 1000. In either reading, the accumulated temperature rise eΣ is lost, and the weight-update pulse would see only the last step's eligibility. The authors' statement that 'test accuracy increases with an increase in τTH and saturates for τTH values exceeding 1 µs' (Sec. 5.2) is therefore in direct conflict with Eq. (23). The explicit Euler form of Eq. (23) is numerically unstable for tPW/τTH > 1, so the model as written cannot support the reported saturation. The paper does not specify tPW or how the thermal decay between pulses is handled in the benchmark simulations, and no code or data are provided to resolve this ambiguity. This is a load-bearing inconsistency that invalidates the claim that the synapse can accumulate the eligibility trace over a dataframe.","section":"Sec. 5.2, Appendix 7.2, 7.3, Eq. (23)"},{"comment":"The integrated heater+ReRAM synapse was not fabricated. The experiments in Sec. 3.1 measure only a standalone ReRAM at different ambient temperatures, which demonstrates that the switching magnitude ΔG/G0 depends on temperature but does not validate the proposed synapse's core functionality: local temperature encoding of e(t), thermal accumulation over the dataframe, and the proportionality of the weight update to the accumulated temperature. The abstract and Sec. 6 claim the synapse has been 'experimentally validated,' but the actual device that would realize the proposed mechanism exists only in COMSOL simulations and in network-level models. The absence of a fabricated integrated cell, or at minimum of an experimental surrogate that demonstrates accumulation over the relevant timescale, leaves the central mechanism unverified.","section":"Sec. 3.1, Sec. 6"},{"comment":"The network-level evaluation is largely circular with respect to the central claim. The hardware-aware simulations implement the eligibility accumulation through Eq. (23), which was constructed specifically to produce a temperature rise proportional to f(t)×ψ(t) and to accumulate over time, and the weight update uses Eq. (27) fitted to the measured ΔG(T) data. The finding that the network trains close to ideal (Fig. 11b) therefore partly reflects the model's design rather than an independently validated physical mechanism. The only independent anchors are the ΔG(T) measurements and the COMSOL thermal response of a single heating pulse (Fig. 6d); neither demonstrates that the eligibility trace can be retained for the duration of a dataframe (hundreds of milliseconds to seconds) with the proposed device parameters. The paper should provide either an experimental demonstration of the accumulation or a clearly separate justification of the thermal model's fidelity beyond its construction.","section":"Sec. 5.2, Appendix 7.5"}],"minor_comments":[{"comment":"The heading and text repeatedly use 'Phenome' instead of 'Phoneme'; this should be corrected throughout the section and the abstract.","section":"Sec. 5.2"},{"comment":"The word 'noeHebbian' appears in the second paragraph of Sec. 3.2; it should be 'neoHebbian.'","section":"Sec. 3.2"},{"comment":"The section heading 'Eligibiiility-based Learning' contains a typo; it should be 'Eligibility-based Learning.'","section":"Sec. 2"},{"comment":"The abstract and Sec. 6 state the synapse has been 'experimentally validated,' but Sec. 3.1 only reports measurements on a standalone ReRAM at elevated ambient temperature. I recommend softening this wording to 'experimentally characterized' or explicitly stating that the integrated device is a proposal supported by simulations.","section":"Abstract and Sec. 6"},{"comment":"The mapping between the discount factor γ used in the reinforcement-learning simulations and the thermal time constant τTH is not specified. The text states that temperature decay realizes γ, but the figures use γ as a free parameter without giving the conversion, which makes it difficult to assess whether the reported optimal γ values correspond to realistic device parameters.","section":"Sec. 5.1"},{"comment":"The data availability statement says data are available upon reasonable request, but no code or data repository is provided. Given that the central simulation results cannot be reproduced without the code and fitted parameters, I recommend providing a public repository with the simulation scripts and fitting routines.","section":"Data Availability"}],"recommendation":"reject","confidential_remarks":"This manuscript presents a creative concept and useful temperature-dependent ReRAM switching data, but the central mechanism is internally inconsistent with the paper's own thermal model and parameter choices. The absence of a fabricated integrated device and the lack of code/data make it impossible to verify the network-level results. In my view, the issues are load-bearing and cannot be fixed by a minor revision; the authors would need to either correct the thermal accumulation model, demonstrate experimentally or via a more faithful simulation that the trace can be retained over the dataframe, or substantially revise the central claim. I recommend rejection, though the authors may consider a future resubmission after a fundamental rework of the accumulation mechanism or a clear demonstration of its feasibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper has a genuinely novel device idea but a load-bearing inconsistency in its own equations. The thermal neoHebbian synapse concept—encoding the eligibility trace in local ReRAM temperature via a co-located heater—is creative and directly addresses a real need in e-prop hardware. The measured ΔG/G0 vs temperature data (Fig. 3) are real and potentially useful for the community, and the COMSOL thermal coupling analysis is a reasonable first pass at quantifying crosstalk.\n\nThe problem is that the central accumulation premise fails under the paper's own numbers. Equation (23) gives a first-order thermal relaxation with time constant τTH. The simulations use δt = 1 ms (Appendix 7.2), and the paper claims accuracy saturates for τTH ≈ 1 μs (Sec. 5.2, Fig. 11(d)). If the heater pulse width is comparable to δt, the temperature jumps to steady state each step; if tPW < τTH as recommended, the temperature decays to ambient before the next e-update because δt/τTH = 1000. Either way, the eligibility trace eΣ is lost, and the weight update sees only the last step's contribution. Fig. 11(d) shows no such collapse, which means the simulation either does not implement Eq. (23) faithfully or the reported τTH is not the physical time constant used. That is not a minor typo; it is the core mechanism of the paper.\n\nSecond, the paper's \"experimentally validated\" claim overreaches. The experiments in Sec. 3.1 measured standalone ReRAM devices at different ambient temperatures, not the integrated heater+ReRAM cell. The heater integration is simulated, not fabricated. That is an honest limitation, but the abstract and conclusion treat it as validated.\n\nThird, the benchmark simulations rely on a four-parameter fitted ΔG model per SET/RESET (Eq. 27), with no code or data provided. Given the internal inconsistency, the accuracy results are not trustworthy without seeing the implementation.\n\nWhere the paper does well: the concept is clearly explained, the prior-work survey (PCM drift, optical, CMOS eligibility traces) is useful, and the differential array scheme for signed updates is sensible. The ΔG(T) measurements, if made available, would be a standalone contribution.\n\nBottom line: the paper is not publishable in its current form. The central mechanism is internally inconsistent, and the validation does not match the claims. That said, the idea is worth pursuing. A serious referee should see it, not desk-reject it, because the flaw is specific and might be fixable if the thermal time constant is actually in the millisecond range and the simulations are corrected. I'd send it to review, but I'd expect major revision or eventual rejection if the inconsistency persists.\n\nRecommendation: send to peer review, but flag the thermal time constant issue clearly.","headline":"Novel thermal eligibility-trace concept undermined by an internal inconsistency: the paper's own numbers predict no trace accumulation, so the central claim is not established.","tokens_in":23125,"tokens_out":3833,"would_cite":false,"duration_ms":31377,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A ReRAM synapse stores its eligibility trace as local heat, enabling online e-prop training without backpropagation through time.","keywords":["NeoHebbian synapse","eligibility trace","e-prop","ReRAM","three-factor learning rule","spiking neural network","thermal coupling","online learning"],"falsifier":"Measure the actual ReRAM temperature in a fabricated 1T-1H-1M cell while applying e-update pulses spaced 1 ms apart with a thermal time constant near 1 microsecond; if the temperature returns to ambient before each next pulse, the accumulated eligibility $e_\\Sigma$ is erased and the weight update is no longer proportional to the summed trace.","tokens_in":21892,"feed_emoji":"🔥","tokens_out":6763,"duration_ms":56921,"temperature":0.7,"pith_summary":"The paper proposes a synaptic device that can run the e-prop learning rule for recurrent spiking neural networks directly in hardware, without backpropagation through time. The central idea is a thermal neoHebbian synapse: a ReRAM cell whose conductance is the coupling weight, while its local temperature is the eligibility trace. A resistive heater receives signals from the pre- and post-synaptic neurons, and the resulting Joule heating makes the ReRAM temperature rise in proportion to their product; at the end of a data frame, a fixed programming pulse changes the conductance by an amount proportional to the accumulated temperature rise. The authors experimentally characterize the temperature-dependent switching of fabricated ReRAM devices and use hardware-aware simulations on a maze-navigation reinforcement task and the TIMIT phoneme classification task to argue that such synapses are fast, compact, and energy-efficient. If correct, this would give neuromorphic chips a physically local, low-energy way to implement three-factor learning rules and train recurrent spiking networks online.","feed_headline":"This synapse stores its learning signal as heat","feed_subtitle":"A heater-coupled ReRAM cell encodes eligibility traces as temperature, enabling online e-prop training in hardware.","key_machinery":"The central object is the thermal neoHebbian synapse: a 1T-1H-1M unit cell in which a resistive heater is 3D-integrated with a ReRAM cell, sharing an electrode so that heat flows directly into the switching oxide. The eligibility trace $e(t)$ is computed physically as Joule heating proportional to $f(t) \\times \\psi(t)$, accumulated as the local temperature rise $e_\\Sigma$ over the dataframe, and read out during the weight-update phase as a temperature-enhanced conductance change under a fixed programming pulse. The argument runs on Eq. (8), $\\Delta w = \\eta e_\\Sigma$, together with the first-order thermal model Eq. (23), which sets the required relationship between pulse width, thermal time constant, and accumulated temperature.","core_discovery":"The paper's central claim is that the local temperature of a ReRAM device can serve as the eligibility trace in the e-prop three-factor learning rule, making each synapse a self-contained online-learning element. In the proposed 1T-1H-1M cell (one transistor, one heater, one ReRAM), the transistor and heater encode the product $f(t) \\times \\psi(t)$ as dissipated power; the resulting temperature rise of the ReRAM accumulates over a dataframe, and a fixed-amplitude programming pulse at the end of the frame turns that accumulated temperature into a conductance change $\\Delta w = \\eta e_\\Sigma$. The authors report experimental data from standalone ReRAM devices showing that normalized conductance change under fixed SET/RESET pulses grows with ambient temperature, and they fit this data into system-level simulations. In those simulations the synapse matches ideal software synapses on TIMIT phoneme classification when given roughly 8-bit precision, loses only about 3% accuracy under scaled thermal crosstalk, and its natural temperature decay implements the discount factor in a reinforcement-learning maze task.","pith_inferences":["Because temperature is a hidden variable, a practical chip would need either a calibration scheme or a secondary readout to know the accumulated eligibility before the weight update; the paper does not demonstrate that readout.","The proportionality between temperature and eligibility assumes the product $f(t) \\times \\psi(t)$ maps to heater power through transistor biasing; a fabricated 1T-1H-1M cell, rather than a standalone ReRAM at controlled ambient temperature, would be the direct test of that mapping.","The same thermal mechanism should generalize to any three-factor rule whose third factor can be encoded as a positive/negative signal pair, so other rules such as reward-modulated STDP could reuse the device unchanged.","Thermal crosstalk, treated here mainly as a source of error, could in principle be exploited as a built-in lateral coupling for neighborhood-based learning rules, though that would require deliberate engineering rather than suppression."],"forward_implications":["Online training of recurrent spiking networks with e-prop becomes implementable in crossbar arrays without unfolding the network or propagating errors backward in time.","Both state variables live locally in the synapse, so no separate high-precision digital unit is needed for eligibility traces, shrinking per-synapse area to 450F^2 with no added footprint for the heater.","Per-timestep learning energy is estimated around 5 pJ, orders of magnitude below prior PCM- or CMOS-based eligibility implementations.","Natural temperature decay is not only tolerable but useful: it provides the discount factor in reinforcement learning, letting the agent weight recent actions more heavily.","With roughly 8-bit ReRAM precision, degradation on TIMIT phoneme classification stays under 3%, and scaling to denser crossbars costs about 3% accuracy from thermal crosstalk."],"supporting_citations":[{"why":"Defines e-prop and the eligibility-trace equations for recurrent spiking neural networks; the learning rule the synapse is built to implement.","marker":"[36]"},{"why":"Supplies the fabricated ReRAM stack and device characteristics whose temperature-dependent switching is measured.","marker":"[10]"},{"why":"Introduces the hardware-aware training techniques used in the benchmark simulations of the neoHebbian array.","marker":"[45]"},{"why":"Prior PCM-based eligibility-trace synapse used as a comparison point for area, energy, and maturity.","marker":"[56]"},{"why":"Prior optical-memristive three-factor synapse used as a comparison point for energy and integration approach.","marker":"[59]"},{"why":"Provides evidence that thermal time constants near 1 microsecond are achievable in practical ReRAM crossbars, supporting the saturation claim.","marker":"[52]"},{"why":"Supplies the temperature-coefficient model used to simulate the effect of local heating on ReRAM conductance.","marker":"[50]"}],"fun_headline_variants":["Heat-based synapses accelerate online neuromorphic training","ReRAM synapse uses temperature as eligibility trace","Synapse encodes learning signal as heat for fast training","Heater-coupled ReRAM stores eligibility trace in heat","Online learning via heat: a new synapse design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the ReRAM's local temperature building up over a whole dataframe and staying there until the weight update, yet the paper's own thermal model has that temperature relaxing in about a microsecond between one-millisecond simulation steps.","fun_headline_variants_meta":{"raw":{"variants":["Heat-based synapses accelerate online neuromorphic training","ReRAM synapse uses temperature as eligibility trace","Synapse encodes learning signal as heat for fast training","Heater-coupled ReRAM stores eligibility trace in heat","Online learning via heat: a new synapse design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1371,"prompt_tokens":954,"completion_tokens":417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":346}},"tokens_in":570,"tokens_out":417,"duration_ms":4316,"temperature":1.0,"reasoning_tokens":346,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:23:08.416110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual ReRAM temperature in a fabricated 1T-1H-1M cell while applying e-update pulses spaced 1 ms apart with a thermal time constant near 1 microsecond; if the temperature returns to ambient before each next pulse, the accumulated eligibility $e_\\Sigma$ is erased and the weight update is no longer proportional to the summed trace.","supporting_citations":[{"cited_title":"Bellec, F","cited_arxiv_id":null,"evidence_quote":"Defines e-prop and the eligibility-trace equations for recurrent spiking neural networks; the learning rule the synapse is built to implement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fabricated ReRAM stack and device characteristics whose temperature-dependent switching is measured."},{"cited_title":"Bhattacharya, S","cited_arxiv_id":null,"evidence_quote":"Introduces the hardware-aware training techniques used in the benchmark simulations of the neoHebbian array."},{"cited_title":"Demira˘ g, F","cited_arxiv_id":null,"evidence_quote":"Prior PCM-based eligibility-trace synapse used as a comparison point for area, energy, and maturity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior optical-memristive three-factor synapse used as a comparison point for energy and integration approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides evidence that thermal time constants near 1 microsecond are achievable in practical ReRAM crossbars, supporting the saturation claim."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the temperature-coefficient model used to simulate the effect of local heating on ReRAM conductance."}],"review_version":1}