{"id":"a2cd9117-a27e-43d6-8505-8e357583a8d2","arxiv_id":"1908.07428","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"This paper shows that an LSTM with a dense output layer and reverse-order sequence-to-sequence training predicts Hodgkin-Huxley CA1 neuron dynamics, with longer predictive horizons lowering the 500 ms RMSE.","lead":"An LSTM neural network was trained to forecast the spiking and bursting of a simulated hippocampal neuron model. Longer multi-step prediction windows reduced long-run errors, but voltage accuracy remained coarse in several firing regimes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Long-horizon improvement may reflect fewer closed-loop iterations rather than better learned dynamics, because the networks are trained on true-state histories but evaluated under their own predictions.","rationale":"The reader's weakest assumption correctly flags the role of true-state initialization, but the sharper issue is the ambiguity and train/evaluation mismatch around the closed-loop protocol. The paper's Fig. 3 shows an iterative chain of Deep LSTM blocks, so after the first block the conditioning history must come from the network itself if the 500-ms comparison is a genuine long-horizon test. Yet Section 2.4's loss and the dataset description in Section 2.3 describe supervised mapping from true state sequences to true state sequences, with no teacher-forcing schedule or scheduled-sampling during training. Thus the network is evaluated in a distribution (its own predicted state histories) it never saw in training. The Np=1 versus Np=200 comparison then conflates two variables: the length of the learned prediction window and the number of recursive feedback steps. This is a load-bearing confound for the abstract's 'predictions improve with increasing predictive time-horizon' claim, independent of the additional concerns about missing baselines, excluded Np=50 in regular bursting, and large absolute RMSE. The proposed reset experiment would separate the feedback-depth effect from the horizon effect. Because the paper currently provides no such separation, the conditional verdict is the right one: the core demonstration is plausible but not yet established at the level claimed.","tokens_in":17231,"tokens_out":5162,"duration_ms":55949,"concrete_test":"Re-run the regular-spiking evaluation (e.g., I=3.0 nA and the 5000-realization protocol) under two conditions: (A) run the Np=1 network closed-loop but re-supply true HH states as the conditioning window every 200 timesteps, matching Np=200's feedback depth; and (B) run the Np=200 network closed-loop with no true-state re-initialization after the first Np steps, feeding its own predicted states into the next block. Add Gaussian noise (e.g., 1 mV) to the initial conditioning states and measure RMSE sensitivity. If condition A's RMSE matches the Np=200 network, or condition B's RMSE degrades sharply, the claimed horizon advantage is an artifact of fewer error-compounding steps rather than better predictive modeling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that predictions improve with increasing predictive horizon (Abstract; Figs. 8a, 10a, 12a) rests on comparing RMSE over 500 ms of iterative prediction. The architecture in Fig. 3 is explicitly closed-loop: each Deep LSTM block receives the previous block's predicted output sequence, concatenated with inputs. However, Section 2.3 states only that 'this initial output sequence data is available to the deep LSTM neural network throughout our simulations,' and Section 2.4 trains the network to map true HH state sequences to the next true HH sequence. The manuscript never states whether the input history for each subsequent block is the network's own predicted states or a fresh true-state window, and it never trains on predicted states as inputs. If the evaluation is closed-loop, then the Np=200 network performs 25 recursive iterations over 500 ms while the Np=1 network performs 5000; its lower RMSE could be almost entirely an artifact of fewer opportunities for feedback-error accumulation. If the evaluation instead re-supplies true HH states before every block, then Fig. 8-12 do not test long-horizon closed-loop prediction at all. Either way, the headline result is protocol-dependent, and no ablation, noise injection, or error-bar analysis resolves which protocol produced the reported curves.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep LSTM architecture with a single fully connected output layer and reversed-order sequence-to-sequence mapping for multi-step prediction of Hodgkin-Huxley CA1 pyramidal neuron dynamics. The authors train networks with predictive horizons Np = 1, 50, 100, and 200, and evaluate them in closed-loop iterative prediction over 500 ms across three firing regimes: regular spiking, irregular bursting, and regular bursting. The main claim is that time-averaged RMSE decreases as the predictive horizon Np increases. The paper also reports full-state prediction results in Appendix B and discusses limitations in Section 4.","tokens_in":17319,"tokens_out":9523,"duration_ms":93029,"significance":"If the horizon-dependent improvement is genuine and not an artifact of the evaluation protocol, the result would be useful for data-driven surrogate models in control-oriented neurostimulation, because larger multi-step horizons reduce the number of recursive rollouts and may mitigate error accumulation. The paper provides the complete Hodgkin-Huxley model equations in Appendix A and evaluates across three regimes with 5000 realizations, which are strengths. However, the central claim is currently under-supported by the absence of statistical significance tests, baselines, and an ablation separating horizon length from iteration count. The claimed architectural novelty (reverse-order mapping and the single dense output layer) is not directly validated against a forward-order or single-step baseline.","major_comments":[{"comment":"The closed-loop evaluation is confounded with the number of recursive iterations. Section 2.1 states that each Deep LSTM block receives the predicted output sequence from the previous block, while Section 2.4 trains the network by Eq. (4) on true HH state sequences. Over the 500 ms evaluation window, the Np=1 network is rolled out 5000 times, whereas the Np=200 network is rolled out only 25 times. The observed decrease in RMSE with increasing Np may therefore reflect fewer opportunities for feedback-error accumulation rather than better learned dynamics. The manuscript does not report a control that fixes the number of iterations (e.g., re-supplying true states before each block or rolling out the Np=200 network for the same number of steps as Np=1). Without such a control, the abstract's claim that predictions improve with increasing predictive time-horizon is not supported as a property of the learned model.","section":"Section 2.1, Section 2.4, Figures 7-12"},{"comment":"The 'decreased consistently' claim is descriptive only. Figures 8(b), 10(b), and 12(b) plot mean RMSE over 5000 realizations without confidence bands, standard deviations, or significance tests, and Figures 8(a), 10(a), and 12(a) compare four horizon values from a single trajectory. No repeated training runs with different seeds are reported, so the reader cannot judge whether the RMSE differences between Np values are larger than training variability. Additionally, no baseline architecture is evaluated: the claimed benefits of the single dense output layer and the reversed-order mapping (Section 2.2) are not tested against a forward-order sequence-to-sequence LSTM or a standard single-step LSTM with feedback. The experimental section therefore does not currently validate either the horizon trend or the architectural novelty.","section":"Section 3, Figures 8, 10, 12"},{"comment":"The regular-bursting RMSE-vs-time panel excludes Np=50, with the text stating that the Np=50 network became unstable for some initial conditions. Since the central claim is that longer horizons uniformly improve accuracy, omitting the Np=50 curve without plotting it or reporting its RMSE prevents the reader from assessing whether the trend is monotone. If Np=50 was unstable, the authors should show the affected trajectories or quantify the failure rate; otherwise the selective exclusion weakens the cross-regime conclusion.","section":"Section 3.3, Figure 12(b)"},{"comment":"The abstract states that the network can predict multi-timescale dynamics 'with reasonable accuracy' and Section 3.1 concludes 'high accuracy over a long-time horizon,' but Sections 3.2 and 3.3 report that in the irregular-bursting regime none of the networks predicted spike timing beyond 100 ms, and in regular bursting the Np=200 network is reasonable only over the first 300 ms. Appendix B also acknowledges that the qsAHP state is not reconstructed. The summary of results should be revised to reflect these stated limitations rather than presenting the regular-spiking result as representative of all regimes.","section":"Abstract; Sections 3.2, 3.3; Appendix B"},{"comment":"The paper does not state whether the test currents and initial conditions used in Section 3 were excluded from the training set described in Section 2.3. Training currents were sampled uniformly in [0, 3] nA, and test currents are drawn from uniform distributions within subintervals of that range; if there is overlap, the evaluation is not out-of-sample. The authors should specify the splitting procedure and, ideally, ensure that the 50 test currents and the 100 initial conditions per current are disjoint from the training set.","section":"Sections 2.3 and 3"}],"minor_comments":[{"comment":"The abstract contains grammatical errors: 'are of interests' should be 'are of interest', and 'from the last several decades' should be 'over the last several decades'.","section":"Abstract"},{"comment":"The Xavier initialization bounds appear to be missing a square root around the denominator; standard Glorot initialization uses ±sqrt(6/(fan_in+fan_out)). Please correct the formula.","section":"Equation (2)"},{"comment":"The summation index k=0,...,NP and the use of 1/NP suggest the loss averages over NP+1 terms; clarify whether the horizon includes the initial condition or not.","section":"Equation (4)"},{"comment":"The text says 'we randomly and uniformly drew 50 samples (i.e., 104 data points)'; this arithmetic is unclear, as 50 samples of a 1000-ms trajectory at 0.1 ms would be much larger, and 104 appears to be a typo for 10^4.","section":"Section 2.3"},{"comment":"The state variable z_M is sometimes written as u_KM in the initial conditions list, and the parameter list duplicates θ_c with different values; use consistent notation and check parameter labels.","section":"Appendix A"},{"comment":"The text contains a typo: 'LTSM' should be 'LSTM'.","section":"Appendix B.3"},{"comment":"Consider plotting confidence intervals (e.g., percentiles or standard errors) rather than a single mean curve, since the number of realizations is large enough to make this informative.","section":"Figures 8(b), 10(b), 12(b)"},{"comment":"The paper would benefit from a statement on data and code availability; the complete HH model in Appendix A is helpful, but the training and testing pipeline is not documented at the level needed for reproduction.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision. The core experiments may be salvageable with additional controls and transparency. The main risk is that the horizon-trend is an artifact of iteration count; this should be addressed before publication. The selective exclusion of Np=50 in Figure 12(b) should also be handled transparently, and the authors should clarify the train/test split for currents and initial conditions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical demonstration, not a new principle. They train stacked LSTMs to predict the 9-state CA1 HH model in three firing regimes, and the closed-loop RMSE over 500 ms drops as the prediction horizon Np increases. The appendices show all state trajectories and 5000-realization RMSE curves, which is more than many ML-for-dynamics papers bother to do. The reverse-order seq2seq trick is borrowed from Sutskever et al., and LSTM surrogates already exist; the novelty is mostly the specific architecture configuration. That's fine, but the paper doesn't test whether this configuration beats anything else, and the central claim is underdetermined.\n\nThe most important soft spot: the evaluation protocol. Figure 3 and Section 2.1 make clear that after the first block, each Deep LSTM receives the previous block's predicted outputs, not ground truth. So the networks are compared at 500 ms with wildly different numbers of closed-loop iterations: 5000 for Np=1 versus 25 for Np=200. The Np=200 network has far fewer chances to accumulate feedback error, so the lower RMSE may reflect 'fewer steps' rather than 'better learned dynamics.' The paper trains on true-state sequences (teacher forcing) and never trains on its own predictions, so the closed-loop evaluation is mismatched with training. The stress-test concern lands. To support the abstract's claim, they need to compare networks at matched iteration counts, or report per-block error, or add noise and scheduled sampling.\n\nOther soft spots are proportionate: no baselines (vanilla LSTM, GRU, reservoir, or even persistent state), no confidence intervals or significance tests, a post-hoc drop of Np=50 in regular bursting, and RMSE values of 7-21 mV that don't justify 'high accuracy.' They also admit q_sAHP is poorly reconstructed; the loss function discussion is honest. Citation pattern is fine.\n\nWhat's genuinely good: the 5000-realization evaluation per regime, the full-state appendix, and the explicit discussion of limitations. The trend is consistent across regimes, and the paper is clearly written enough that the protocol question can be resolved by rereading.\n\nVerdict: worth sending to peer review, but it needs major revision before the claims can stand. A referee should ask for baselines, error bars, a matched-iteration experiment, and a statement of whether training is teacher-forced.","headline":"A competent LSTM surrogate for a CA1 Hodgkin-Huxley model, but the headline horizon trend is confounded by closed-loop iteration count and missing baselines.","tokens_in":17999,"tokens_out":3240,"would_cite":false,"duration_ms":35241,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LSTM networks trained for longer prediction horizons forecast neuron spiking more accurately.","keywords":["long short-term memory","brain dynamics","data-driven modeling","complex systems","Hodgkin-Huxley model","CA1 pyramidal neuron","multi-timestep prediction","sequence-to-sequence mapping"],"falsifier":"Take a trained $N_p=200$ network and start a 500 ms forecast from a conditioning history that is the true Hodgkin-Huxley trajectory shifted by 1 mV (or corrupted with Gaussian noise of comparable amplitude), then compare RMSE against the clean-history forecast; if the error grows steeply, the reported long-horizon improvement depends on perfectly known initial conditions, and if it does not, the improvement is robust to conditioning error.","tokens_in":16837,"feed_emoji":"🧠","tokens_out":8393,"duration_ms":79330,"temperature":0.7,"pith_summary":"This paper tries to establish that a long short-term memory (LSTM) neural network can serve as a purely data-driven surrogate for a biophysical neuron model, predicting the membrane potential and gating variables of a hippocampal CA1 pyramidal neuron over hundreds of milliseconds. The proposed architecture stacks three LSTM layers with a single fully connected output layer and feeds inputs in reverse order through a sequence-to-sequence mapping, then chains copies of the network to iterate forecasts. Across regular spiking, irregular bursting, and regular bursting regimes, the paper reports that longer predictive horizons produce lower time-averaged error over 500 ms than single-step prediction. The motivation is a computationally cheap model that could later be embedded in control-theoretic neurostimulation strategies.","feed_headline":"Longer LSTM horizons predict neuron spiking better","feed_subtitle":"Trained to forecast 20 ms blocks, the LSTM tracks hippocampal neuron activity for 500 ms with less error than single-step models.","key_machinery":"The load-bearing object is a deep LSTM network with three stacked hidden layers of 324 units feeding a single fully connected output layer, trained on simulated traces of the 9-dimensional Hodgkin-Huxley CA1 neuron model. It makes multi-timestep predictions through sequence-to-sequence mapping in which the input sequence is presented in reverse order, so that early inputs correspond to early outputs through short-range dependencies while later inputs map through long-range dependencies; the paper credits this temporal symmetry for improved early-trajectory accuracy. To reach long horizons, copies of the trained network are chained: the predicted output block is concatenated with the next block of inputs and fed, again in reverse order, into the next copy. The iterative chaining plus reversed-order mapping is what carries the claim that a network trained on 20 ms blocks can forecast 500 ms of spiking behavior.","core_discovery":"On the paper's own terms, the central discovery is that prediction accuracy improves with the length of the multi-timestep predictive horizon: time-averaged RMSE of the membrane potential decreases consistently as $N_p$ goes from 1 to 50 to 100 to 200 timesteps, in all three dynamical regimes, over 5000 random current-and-initial-condition realizations. Longer-horizon networks track the timing of spikes better, while shorter-horizon networks track spike amplitude better. The regular spiking regime is predicted with high accuracy over the full 500 ms; irregular bursting is captured only over roughly the first 100 ms; regular bursting is captured with reasonable accuracy over about 300 ms for $N_p = 200$. The paper also reports that one gating variable, $q_{sAHP}$, is not reconstructed because its magnitude approaches numerical precision.","pith_inferences":["Not tested in the paper: corrupting the conditioning sequence with observation noise, then checking whether the $N_p=200$ advantage survives.","Not tested in the paper: applying the same architecture to a chaotic or aperiodic system, which would show whether the reverse-order benefit generalizes beyond periodic spiking patterns.","Not tested in the paper: probing horizons beyond $N_p=200$ to see whether the RMSE trend saturates, plateaus, or reverses.","Not tested in the paper: retraining with norm-1 or weighted norm-2 losses to see if spike amplitude and timing can be recovered together, as the discussion leaves this open."],"forward_implications":["In the regular spiking regime, a network trained to predict 20 ms blocks can be iterated over 500 ms with lower RMSE than a network trained for single-step prediction, so longer-horizon training is a viable route to long forecasts.","Across all three regimes, increasing $N_p$ shifts the error trade-off: spike timing improves while spike amplitude degrades, meaning the loss function directly shapes what the surrogate captures.","Irregular bursting is not predictable beyond about 100 ms by any of the trained horizons, which sets a concrete limit on neurostimulation control in that regime.","The slow calcium-activated potassium variable $q_{sAHP}$ is not learned to reasonable accuracy, so the surrogate does not fully reconstruct the 9-dimensional state even when membrane potential traces look right.","Longer horizons require substantially more offline training cost for the same number of examples, so choosing $N_p$ is a practical trade-off between forecast accuracy and training budget."],"supporting_citations":[{"why":"Supplies the LSTM cell equations and gating structure on which the stacked architecture is built.","marker":"[25]"},{"why":"Supplies the reverse-order sequence-to-sequence mapping whose temporal symmetry the paper uses to improve early and long-horizon prediction.","marker":"[27]"},{"why":"Supplies the experimentally validated 9-dimensional Hodgkin-Huxley CA1 neuron model that generates the synthetic training data and ground-truth comparisons.","marker":"[29]"},{"why":"The RNN multistep prediction approach the paper positions against, described as suffering from poor early trajectory performance on long horizons.","marker":"[20]"},{"why":"An LSTM dynamic-system identification study whose long-horizon early-trajectory shortcoming the paper aims to overcome.","marker":"[23]"},{"why":"Earlier LSTM forecasting of high-dimensional chaotic systems, limited to single-step prediction, which the paper extends to multi-timestep forecasts.","marker":"[24]"}],"fun_headline_variants":["Longer forecast horizons reduce neuron prediction error","Neuron spike timing improves with longer LSTM horizons","Forecasting further ahead sharpens LSTM neuron models","Longer LSTM windows better predict spike timing than amplitude","For hippocampal neurons, longer LSTM forecasts win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true Hodgkin-Huxley state trajectory is available as the initial $N_p$-step input sequence for every forecast, so the paper never tests how the network behaves when the conditioning history contains errors or partial observations.","fun_headline_variants_meta":{"raw":{"variants":["Longer forecast horizons reduce neuron prediction error","Neuron spike timing improves with longer LSTM horizons","Forecasting further ahead sharpens LSTM neuron models","Longer LSTM windows better predict spike timing than amplitude","For hippocampal neurons, longer LSTM forecasts win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2515,"prompt_tokens":953,"completion_tokens":1562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":1488}},"tokens_in":569,"tokens_out":1562,"duration_ms":13023,"temperature":1.0,"reasoning_tokens":1488,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:57:43.841565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained $N_p=200$ network and start a 500 ms forecast from a conditioning history that is the true Hodgkin-Huxley trajectory shifted by 1 mV (or corrupted with Gaussian noise of comparable amplitude), then compare RMSE against the clean-history forecast; if the error grows steeply, the reported long-horizon improvement depends on perfectly known initial conditions, and if it does not, the improvement is robust to conditioning error.","supporting_citations":[{"cited_title":"Contribution of persistent na+ current and m-type k+ current to somatic bursting in ca1 pyramidal cells: combined experimental and modeling study","cited_arxiv_id":null,"evidence_quote":"Supplies the experimentally validated 9-dimensional Hodgkin-Huxley CA1 neuron model that generates the synthetic training data and ground-truth comparisons."},{"cited_title":"Multistep prediction of dynamic systems with recurrent neural networks","cited_arxiv_id":null,"evidence_quote":"The RNN multistep prediction approach the paper positions against, described as suffering from poor early trajectory performance on long horizons."},{"cited_title":"A new concept using LSTM neural networks for dynamic system identification","cited_arxiv_id":null,"evidence_quote":"An LSTM dynamic-system identification study whose long-horizon early-trajectory shortcoming the paper aims to overcome."},{"cited_title":"Data-driven forecasting of high-dimensional chaotic systems with long short-term memory networks","cited_arxiv_id":null,"evidence_quote":"Earlier LSTM forecasting of high-dimensional chaotic systems, limited to single-step prediction, which the paper extends to multi-timestep forecasts."}],"review_version":1}