{"id":"868f3915-8854-4f01-baa6-741939bf26ca","arxiv_id":"2507.07907","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Optimal control on high-dimensional learning equations produces nontrivial training schedules, such as easy-hard-easy curricula, that outperform standard baselines in toy and real-data tests.","lead":"The authors build a framework that combines statistical physics with optimal control theory to design optimal training schedules for simple neural networks. They derive schedules for curriculum learning, dropout, and denoising autoencoders that beat standard baselines, offering a principled route to meta-learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed 'typical' DAE noise schedule (Discussion) is contradicted by the paper's own MNIST-based optimum (Fig. 4a), whose shape is inverted; the transferable-principle claim is internally inconsistent.","rationale":"The central claim has three load-bearing parts: exactness of the order-parameter ODEs (Eq. 10), global optimality of the computed schedules within the model class, and transfer of qualitative features to real architectures and data. The reader's weakest assumption (concentration) targets part one; it is real but is the standard statistical-physics-of-learning approximation, it is cited to [13] as provable under additional assumptions, and S5 shows excellent single-trajectory agreement for a one-switch curriculum, constant dropout, and constant DAE noise at N=10,000-30,000. A catastrophic failure is therefore unlikely; the honest status of that concern is 'unproved but standard and numerically supported in closely related regimes.' Part three, by contrast, already fails on the paper's own evidence. Section 3.3 reports two optimal DAE noise schedules with opposite shapes for the same model class: Fig. 3a (Gaussian mixture) initially decreases then moderately increases, while Fig. 4a (MNIST-fitted clusters) initially increases then gradually decreases. Section 4 then asserts the toy shape as 'typical' for DAEs, with no caveat for the MNIST configuration and no attempt to explain the flip. Either the two computations are not actually the same problem (a parameter or objective mismatch visible only on rerun), or the optimal schedule is so sensitive to cluster statistics and initial conditions that the claimed 'interpretable strategies' are not transferable principles. Both readings damage the abstract's promise and the 'qualitative features transfer' message, and the discrepancy is checkable with the public code. This is why I elevate the DAE inconsistency over the proof gap as the single most load-bearing concern: the proof gap is absence of guarantee, while the DAE contradiction is counter-evidence in hand. The reader's additional issues are sound (the CIFAR schedule parameter is selected on the test set, so the easy-hard-easy advantage is a fitted minimum, although with 30 seeds the U-shape is credible). The curriculum and dropout results are independent of the DAE issue, the code is public, and the framework is coherent, so the paper should remain a CONDITIONAL accept, with conditions: reconcile or qualify the DAE schedule claim, move the CIFAR schedule selection off the test set, and add finite-N checks of the headline time-varying schedules to S5.","tokens_in":26970,"tokens_out":21831,"duration_ms":229545,"concrete_test":"Using the authors' public CasADi code, recompute the optimal DAE noise schedule at matched settings: (i) toy parameters of Fig. 3 (initial conditions S71, sigma=0.1 for both clusters, p_c=0.5, N=1000) and (ii) MNIST-fitted parameters of Fig. 4 (S72, sigma=(0.05,0.033), p_c=(0.47,0.53), N=784), each at the same Delta_F (e.g., 0.25 and 0.4) and alpha_F (1.0 and 1.2), plus a continuous interpolation in initial overlap R to locate the regime boundary where the optimal shape flips from decrease-then-increase to increase-then-decrease. If both shapes appear within one fixed parameter configuration, one of the two computations is erroneous; if they occupy disjoint regimes, Section 4 must state that regime and drop the unqualified 'typically' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 (Discussion) states that 'for denoising autoencoders, the optimal noise schedule typically features an initial decrease followed by a moderate increase toward the end of training.' This is directly contradicted by the paper's own Fig. 4a (MNIST-fitted DAE), which shows 'an initial increase in noise followed by a gradual decrease toward the end,' a pattern the text calls characteristic of all tested noise levels. Both are presented as optimal schedules for the same model class (K=C1=2, g(z)=z, trained skip connection), differing only in fitted cluster parameters, initial conditions (S71 vs. S72), N, and alpha_F. No passage reconciles the sign flip or bounds the regime in which each shape holds. This internal inconsistency hits two components of the central claim directly: the abstract's 'nontrivial yet interpretable strategies' (the identified principle does not transfer between two instances of the same model class) and 'qualitative features transfer to real architectures and datasets' (the toy Gaussian-mixture shape does not transfer to the authors' own MNIST instance). The mechanism offered for the toy case (skip-connection convergence, Fig. 3d) is not discussed for MNIST. Separately, the concentration premise (Eq. 10) is admitted unproved in S1 and the finite-N validation in S5 covers a one-switch anti-curriculum, a constant dropout, and a constant DAE noise schedule, but not the headline two-switch easy-hard-easy, the sharp-eta-drop joint schedule, or the optimal DAE noise schedule itself; that gap is real but is a missing check, whereas the DAE sign contradiction is counter-evidence already present in the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework that combines the high-dimensional asymptotics of online SGD with optimal control, writing low-dimensional order-parameter ODEs (Eq. 10) for the sequence multi-index model and minimizing the final generalization error over a protocol u(α). The framework is applied to three scenarios: curriculum learning (reporting an easy-hard-easy schedule at fixed learning rate and easy-to-hard when the learning rate is jointly optimized), dropout regularization (delayed-onset dropout), and denoising autoencoders (nontrivial noise schedules and batch-size augmentation). The theoretical schedules are compared with finite-N simulations for constant protocols and with CIFAR-10/MNIST experiments for the curriculum and DAE cases. The central claim is that these schedules are globally optimal within the model class and that their qualitative features transfer to real architectures and datasets.","tokens_in":27266,"tokens_out":6809,"duration_ms":73961,"significance":"The framework is a natural and potentially valuable contribution: it packages the established teacher-student ODE machinery into an optimal-control setting, and the reported schedules are genuine results of a minimization rather than fits of the reported improvements. The code is public, the ODEs for the three scenarios are written out in the supplement, and the finite-N checks for constant protocols are reassuring. If the concentration assumption were proved and the DAE schedule-shape contradiction resolved, the paper would provide a principled testbed for meta-learning grounded in statistical physics. As it stands, the strength of the empirical and transfer claims exceeds the evidence, so the manuscript needs substantial revision before the central claims can be accepted.","major_comments":[{"comment":"The claimed DAE noise schedule is internally inconsistent. Section 3.3 (paragraph beginning \"We observe that the optimal schedule typically...\") and Section 4 both state that for denoising autoencoders the optimal noise schedule typically has an initial decrease followed by a moderate increase toward the end of training. The MNIST-fitted case in Fig. 4a is described two paragraphs later as following the opposite \"characteristic pattern: an initial increase in noise followed by a gradual decrease toward the end of the training trajectory.\" Both cases use the same model class (linear activation, K=C1=2, trained skip connection) and differ only in fitted cluster parameters, initial conditions (S71 vs S72), N, and alpha_F. The text never reconciles the sign flip or states the parameter regime in which each shape holds. Since the abstract and Discussion advertise \"nontrivial yet interpretable strategies\" whose \"qualitative features transfer to real architectures and datasets,\" this direct contradiction between two instances of the same model class undermines a central claim. Please either identify the controlling parameter for the shape change, or restrict the transferability claim to the tested regime.","section":"3.3, 4"},{"comment":"The paper's optimality statements rest on the premise that Eq. (10) exactly describes the online SGD dynamics in the high-dimensional limit. Supplementary S1 explicitly concedes \"we do not provide a rigorous proof of this result here\" and cites numerical verification. That verification, in S5, covers only a one-switch anti-curriculum (Fig. S4), a constant dropout rate (Fig. S5), and a constant DAE noise level with constant batch size (Fig. S6). None of the headline schedules—the easy-hard-easy curriculum, the jointly optimized learning-rate/curriculum protocol, the optimal time-varying dropout schedule, or the optimal noise schedules of Fig. 3a and Fig. 4a—is validated against finite-N SGD. Without either a proof of concentration for time-dependent controls or numerical validation on the actual optimal schedules, the guarantee that these protocols are optimal for the simulated systems is not established. At minimum, please add finite-N comparisons for the computed optimal protocols and discuss the N-dependence of the claimed optimality.","section":"S1, S5, Eq. (10)"},{"comment":"The paper describes the obtained schedules as \"globally optimal\" (Section 2.3). The implemented methods, however, provide only local or necessary conditions: the forward-backward sweep enforces Pontryagin's necessary conditions, and the direct CasADi formulation is a nonlinear program for a nonconvex dynamical system. No convexity result or systematic multi-start/global-optimization check is given. Since the word \"optimal\" is central to the paper's claims, either demonstrate that the reported solutions are global (e.g., by exhaustive comparison for the low-dimensional controls or by establishing convexity of the reduced problem) or rephrase the claims as locally optimal within the model class.","section":"2.3, S2"}],"minor_comments":[{"comment":"The caption reads \"Optimal noise schedule Δ as a function of the training.\" and ends mid-sentence; it should be \"as a function of the training time α.\"","section":"Fig. 4a caption"},{"comment":"The CIFAR-10 experiment uses a one-parameter family (initial easy fraction a, with the hard phase fixed at 50% and the rest easy). This supports non-monotonicity but does not test the specific two-switch schedule or the predicted phase timing. Please soften \"consistent with the optimal control solution\" or add a comparison with the theoretically predicted phase lengths.","section":"3.1, S3"},{"comment":"The notation μ is used both for the sample index (x^μ) and for the cluster means μ_{l,c_l}; the footnote acknowledges this, but the double use still makes several equations in S1 harder to read.","section":"S1"},{"comment":"The expression x̃σ(·) is missing a space and the definition of the output dimension is terse; adding a sentence clarifying the tensor shape would help.","section":"Eq. (11)"},{"comment":"The statement that stochastic fluctuations \"vanish\" is stated without a forward reference to S1, where the lack of a rigorous proof is immediately conceded. Please add the pointer and state the numerical verification explicitly at this location.","section":"2.2"}],"recommendation":"major_revision","confidential_remarks":"The fit to the journal's scope is good. The self-citation to [45] is legitimate and not a concern. The internal DAE contradiction is the most serious issue and should be resolved before publication; the concentration gap can be addressed by adding simulations of the actual optimal schedules or by clearly marking the optimality claim as conditional on the ODE model. I do not share the skeptic's circularity concern: the schedules are genuine outputs of an optimization problem on known ODEs, not fitted to the reported gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one. The core idea is clean: take the low-dimensional order-parameter ODEs for online SGD in the sequence multi-index model, treat the metaparameters as controls, and solve an optimal-control problem that minimizes final generalization error. That is a real unification of earlier optimal-learning-rate work from the 1990s and the authors' own ICLR 2025 continual-learning paper, and it produces genuinely new outputs: the easy-hard-easy curriculum at fixed learning rate, the easy-to-hard schedule when the learning rate is jointly optimized, delayed-onset dropout, and nontrivial DAE noise and batch-size schedules. The OC solutions are genuine minimizations, not fits, and the code is public. The ODEs are standard high-dimensional asymptotics and are checked against finite-N simulations for constant or simple protocols. Credit where due: this is a useful framework and the curriculum results alone are worth a paper.\n\nThe soft spots are real but not evenly distributed. The biggest one is an internal contradiction in the DAE section. The Discussion says the optimal noise schedule 'typically features an initial decrease followed by a moderate increase toward the end.' That matches the toy Gaussian-mixture result in Fig. 3a. But the MNIST-fitted model in Fig. 4a shows the opposite shape: an initial increase in noise followed by a gradual decrease. Both are presented as optimal schedules for the same model class, differing only in fitted cluster statistics and initial conditions. No passage reconciles the sign flip or bounds the regimes. This directly undercuts the abstract's 'nontrivial yet interpretable strategies' and the claim that qualitative features transfer to real data. The authors need to fix this, either by showing which data properties select which shape or by dropping the universal-principle language.\n\nTwo smaller issues. First, the CIFAR-10 validation in S3 selects the easy-hard-easy fraction a by picking the best test error (a=35%), so the quantitative comparison is a fit, not an independent prediction. The qualitative non-monotonic shape is still confirmed, but the claim 'transfers to realistic architectures' should be softened to 'is consistent with' unless they prespecify a from the theory. Second, the concentration assumption behind the ODEs is admitted unproved in S1, and the finite-N checks in S5 cover a one-switch anti-curriculum, constant dropout, and constant DAE noise, not the headline two-switch easy-hard-easy, the sharp-eta-drop schedule, or the optimal DAE noise schedule itself. Missing checks are not counter-evidence, but for a paper whose selling point is optimal schedules, those are exactly the trajectories that should be simulated.\n\nThis deserves a serious referee. I would accept it with major revision: resolve the DAE shape contradiction, reword the transfer claims, and add finite-N validation of at least the easy-hard-easy and optimal DAE noise schedules. Given the framework's breadth and the code, the field gets value from this even if the qualitative-transfer slogan now needs qualification.","headline":"Solid framework and genuinely new schedules, but the paper's own MNIST DAE results contradict its headline 'transferable principle,' so the Discussion needs the sharpest fix.","tokens_in":27891,"tokens_out":2063,"would_cite":true,"duration_ms":24311,"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":"This paper claims that the design of learning protocols can be reduced to an optimal control problem on the low-dimensional order-parameter dynamics of online SGD, and that solving it yields nontrivial, interpretable schedules that…","keywords":["optimal control","statistical physics","curriculum learning","dropout","denoising autoencoder","generalization error","order parameters","stochastic gradient descent"],"falsifier":"Run 100 independent online-SGD trajectories in the curriculum teacher-student model with N = 30,000 using the ODE-derived easy-hard-easy schedule, and compare the mean final misclassification error to both the ODE prediction and to the best error found by a brute-force grid search over piecewise-constant schedules; if the finite-N mean differs from the ODE prediction beyond sampling error, or if the grid search finds a schedule with lower mean error, the paper's concentration premise is falsified.","tokens_in":26696,"feed_emoji":"🧭","tokens_out":10656,"duration_ms":108881,"temperature":0.7,"pith_summary":"This paper claims that the search over training protocols—curricula, learning-rate schedules, dropout probabilities, noise levels—can be reduced to an optimal control problem on a low-dimensional set of order parameters, instead of the full high-dimensional parameter space. In the high-dimensional limit, online stochastic gradient descent is described by closed-form ODEs for these order parameters, and the paper solves the resulting control problem to find protocols that minimize final generalization error. For the examples studied, the optimal protocols are nontrivial and interpretable: an easy-hard-easy curriculum at fixed learning rate, easy-to-hard when learning rate is jointly optimized, delayed onset of dropout, and noise and batch-augmentation schedules for denoising autoencoders. A sympathetic reader would care because these schedules are derived from first principles rather than heuristic trial-and-error, and the paper reports that their qualitative structure transfers to convolutional networks on CIFAR-10 and to MNIST denoising.","feed_headline":"Physics finds optimal curricula and dropout schedules","feed_subtitle":"They beat constant and monotonic baselines and transfer to real datasets.","key_machinery":"The machinery is the set of order-parameter ODEs in Eq. (10), which replace the high-dimensional SGD dynamics with a low-dimensional deterministic system for the overlaps Q, M, R and the readout v; the control vector u(alpha) enters these ODEs as a time-dependent parameter. The paper applies optimal control to this reduced system rather than the original network, using both Pontryagin's maximum principle (adjoint equations) and direct discretization, so the curse of dimensionality is lifted and the resulting schedules can be read off from the behavior of interpretable order parameters.","core_discovery":"The central discovery is that the design of optimal learning protocols can be formulated as an optimal control problem directly on the dynamics of order parameters, with the final generalization error as the objective. The paper derives closed-form ODEs for online SGD in the sequence multi-index model, then solves the control problem with Pontryagin's maximum principle and direct multiple-shooting methods. The solutions it finds—easy-hard-easy curricula at fixed learning rate, easy-to-hard with jointly optimized learning rate, delayed dropout, and noise/batch schedules for denoising autoencoders—outperform constant and monotonic baselines, and the qualitative structure of the schedules transfers to Cluttered CIFAR-10 and MNIST.","pith_inferences":["A natural extension is to apply the same order-parameter control formulation to batch learning, where the dynamics are described by dynamical mean-field theory rather than ODEs; the paper lists this as an open direction but does not derive schedules for it.","The contrast between easy-hard-easy at fixed learning rate and easy-to-hard with a tuned learning rate suggests a principle left implicit in the paper: the optimal ordering of examples depends on whether other controls can compensate for the noise amplification of hard examples.","If the concentration property were proven rigorously rather than only verified numerically, the framework would constitute a general foundation for meta-learning, since minimizing final generalization error over protocol choices is exactly a meta-objective.","A testable extension is to check whether delayed-onset dropout, which is validated only in the model, transfers to real networks, just as the curriculum schedule was shown to transfer to Cluttered CIFAR-10."],"forward_implications":["Within the sequence multi-index model, the schedules obtained from the optimal-control solution are globally optimal among all protocols in the allowed control set, not merely local improvements.","At fixed learning rate, the optimal curriculum is non-monotonic easy-hard-easy; when the learning rate is optimized jointly, the optimal curriculum becomes easy-to-hard, with a warm-up phase and a sharp drop at the transition to hard samples.","The optimal dropout schedule keeps all nodes active early and gradually reduces the activation probability after the teacher-student alignment saturates, meaning that dropout should be delayed rather than applied at a constant rate.","For denoising autoencoders, the optimal noise schedule decreases then increases toward the end of training and improves the MSE over the constant schedule by roughly 10-30% in the model and up to 40% on MNIST.","The qualitative structure of the optimal schedules transfers to real architectures and datasets, as demonstrated on Cluttered CIFAR-10 and MNIST."],"supporting_citations":[{"why":"Supplies the high-dimensional SGD dynamics that concentrate into the closed ODEs used throughout the paper.","marker":"[13]"},{"why":"Provides the analytically tractable teacher-student curriculum model that Section 3.1 extends to optimal schedules.","marker":"[16]"},{"why":"Defines the sequence multi-index model that gives the framework its closed-form dynamics.","marker":"[19]"},{"why":"Provides the dropout model with constant activation probability that the optimal dropout schedule generalizes.","marker":"[45]"},{"why":"Supplies the high-dimensional asymptotics of denoising autoencoders used for the noise-schedule optimization.","marker":"[69]"},{"why":"Gives Pontryagin's maximum principle used as the indirect optimal-control method.","marker":"[78]"},{"why":"Provides the software framework used for the direct multiple-shooting optimal-control solutions.","marker":"[79]"}],"fun_headline_variants":["Statistical physics derives optimal learning schedules","Physics control optimizes curricula and dropout","Optimal training protocols from physics control","Control theory on order parameters yields optimal learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the order-parameter ODEs exactly describe online SGD in the high-dimensional limit; the paper itself states in the supplementary material that this concentration property is verified numerically rather than proven rigorously, so if it fails at finite N or for real networks, the derived schedules are not guaranteed to be optimal for the actual systems.","fun_headline_variants_meta":{"raw":{"variants":["Statistical physics derives optimal learning schedules","Physics control optimizes curricula and dropout","Optimal training protocols from physics control","Control theory on order parameters yields optimal learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2612,"prompt_tokens":900,"completion_tokens":1712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":1661}},"tokens_in":516,"tokens_out":1712,"duration_ms":17360,"temperature":1.0,"reasoning_tokens":1661,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:30:38.884287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run 100 independent online-SGD trajectories in the curriculum teacher-student model with N = 30,000 using the ODE-derived easy-hard-easy schedule, and compare the mean final misclassification error to both the ODE prediction and to the best error found by a brute-force grid search over piecewise-constant schedules; if the finite-N mean differs from the ODE prediction beyond sampling error, or if the grid search finds a schedule with lower mean error, the paper's concentration premise is falsified.","supporting_citations":[{"cited_title":"Dy- namics of stochastic gradient descent for two-layer neural networks in the teacher-student setup","cited_arxiv_id":null,"evidence_quote":"Supplies the high-dimensional SGD dynamics that concentrate into the closed ODEs used throughout the paper."},{"cited_title":"An analytical theory of curriculum learning in teacher-student networks","cited_arxiv_id":null,"evidence_quote":"Provides the analytically tractable teacher-student curriculum model that Section 3.1 extends to optimal schedules."},{"cited_title":"High-dimensional learning of narrow neural networks.Journal of Statistical Mechanics: Theory and Experiment, 2025(2):023402, 2025","cited_arxiv_id":null,"evidence_quote":"Defines the sequence multi-index model that gives the framework its closed-form dynamics."},{"cited_title":"Analytic theory of dropout regularization.Phys","cited_arxiv_id":null,"evidence_quote":"Provides the dropout model with constant activation probability that the optimal dropout schedule generalizes."},{"cited_title":"High-dimensional asymptotics of denoising autoencoders.Ad- vances in Neural Information Processing Systems, 36:11850–11890, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the high-dimensional asymptotics of denoising autoencoders used for the noise-schedule optimization."},{"cited_title":"Some mathematical problems arising in connection with the theory of optimal au- tomatic control systems","cited_arxiv_id":null,"evidence_quote":"Gives Pontryagin's maximum principle used as the indirect optimal-control method."},{"cited_title":"Casadi: a software framework for nonlinear optimization and optimal control.Mathematical Programming Computation, 11:1–36, 2019","cited_arxiv_id":null,"evidence_quote":"Provides the software framework used for the direct multiple-shooting optimal-control solutions."}],"review_version":1}