{"id":"923d533c-4f7f-45fe-8cd9-2d2e77b912f8","arxiv_id":"2508.10666","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A graduate-level lecture-note review of deep learning methods and their applications in classical and quantum physics, with hands-on examples.","lead":"These are graduate-level lecture notes on deep learning for physics, covering neural network fundamentals, unsupervised learning, and planned applications to quantum systems. The notes are a teaching resource rather than a new research result, and their value depends on the accuracy and reproducibility of the worked examples.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3.24) defines the RBM partition function as Z = Σ E(z) rather than Σ e^{-E(z)}; the subsequent RBM derivation (Eqs. 3.36–3.39) is inconsistent with it, so the notes cannot be trusted as a complete, detailed guide as printed.","rationale":"The paper is explicitly a pedagogical resource: the abstract and preface promise a comprehensive, detailed, graduate-level introduction, and a student's ability to trust the equations is essential to that claim. The reader's weakest assumption identifies exactly this trust premise and notes several concrete violations. I agree with that assessment. Among the flagged issues, Eq. (3.24) is the most load-bearing because it sits at the centre of the RBM/generative-model tutorial and is internally inconsistent with the derivation that follows. Correcting it is straightforward, but the presence of such an error—along with the bias-index/dimension mismatch in Eq. (2.5) and the t-SNE KL-divergence claim—means the notes do not currently deliver on their central promise. This is a correctable defect, not a rejection of the whole enterprise; the chapters have useful conceptual framing and the authors explicitly invite errata. However, the appropriate disposition is conditional acceptance: the pedagogical claim can be satisfied once the printed equations are made internally consistent and, ideally, the numerical examples are backed by reproducible code or data. I therefore recommend CONDITIONAL rather than UNVERDICTED, since the central claim here is a pedagogical one that can be checked and corrected.","tokens_in":51014,"tokens_out":6993,"duration_ms":80738,"concrete_test":"Re-implement the RBM training loop in §3.6 exactly as printed—using Eq. (3.24) for Z—on a small binary dataset (e.g., 4-bit vectors); then replace Z with the standard Boltzmann definition and retrain. If the printed version fails to increase log-likelihood, or diverges while the corrected version learns, the error is load-bearing. Equivalently, re-derive Eq. (3.39) from Eq. (3.24) and observe that the factorized (1+exp(x)) form cannot be obtained.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The manuscript's central promise is that a student can work through the notes and come away with an accurate, hands-on command of DL for quantum physics. That promise is load-bearing on the reliability of the printed equations, and §3.6 already breaks it. Eq. (3.23) correctly states p(z) = Z^{-1} exp(-E(z)), but Eq. (3.24) then defines Z = Σ_z E(z), omitting the exponential Boltzmann weight. This is not a harmless typo: the RBM training loss, gradient (Eq. 3.42), and contrastive-divergence update (Algorithm 4) all hinge on the partition function. The subsequent derivation of P(v) (Eqs. 3.36–3.39) implicitly uses Z = Σ_{v,h} exp(-E(v,h)) and factors the hidden sum into products (1 + exp(...)); that factoring is impossible from Eq. (3.24). A student who implements the text literally will compute nonsensical gradients, so the 'complete, detailed description' is not, as printed, usable. The same trust-break is visible in the bias-index/dimension mismatch in Eq. (2.5) and the KL-divergence value after Eq. (3.8), and the numerical examples in Chs. 2–3 are not accompanied by code or data. Each of these is individually correctable, but together they show the central pedagogical claim is currently unmet.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a set of lecture notes, arXiv:2508.10666, aiming to provide a graduate-level, self-contained introduction to deep learning and its applications in classical and quantum physics. The first three chapters cover neural-network anatomy, training, classification examples (MNIST, CIFAR-10), the 2D Ising model, unsupervised methods (PCA, t-SNE, autoencoders, VAEs, GANs, restricted Boltzmann machines, normalising flows), and some physics-informed concepts. Later chapters summarize quantum mechanics and deep-learning techniques in quantum science. The paper makes no claim to new research results; its central claim is pedagogical: that a student working through the notes can acquire a correct, hands-on command of deep learning and judge when to apply it in quantum problems. The reviewed chapters contain several concrete mathematical errors, so the central claim is not met as printed.","tokens_in":51335,"tokens_out":8515,"duration_ms":92541,"significance":"If the equation-level errors were corrected, these notes would fill a useful niche: they connect deep-learning formalism to statistical physics and quantum applications at a level appropriate for graduate students. The organizational structure is clear, and the conceptual discussion—e.g., the Chinese-room thought experiment, the distinction between points and probability distributions in VAEs, and the motivation for energy-based models—is a genuine strength. The paper does not present original research, so circularity is not a concern; the concern is reliability of the reproduced material. No code, data, or machine-checked derivations are supplied, so the 'hands-on' promise depends entirely on the printed equations and figures. Because several load-bearing equations are wrong or internally inconsistent, the notes cannot currently be used as a trustworthy teaching resource.","major_comments":[{"comment":"The partition function is defined as Z = Σ_z E(z), omitting the exponential Boltzmann weight. This contradicts Eq. (3.23), which correctly has p(z) = Z^{-1} exp(-E(z)). The subsequent derivation of P(v) in Eqs. (3.36)–(3.39) factors sums of exp(...) terms, which is impossible if Z were defined as in Eq. (3.24). The gradient formula (3.42) and Box 3.4 use the correct Z = Σ_{v,h} exp(-E(v,h)), so the text is internally inconsistent. A student implementing the equations literally will obtain nonsensical normalizations and gradients. This is a load-bearing error for the RBM section and for the 'complete, detailed description' promised in the preface.","section":"§3.6, Eq. (3.24)"},{"comment":"The KL divergence is given as KL(p||q) = H + L_ce, where H = -Σ p log p and L_ce = -Σ p log q. The correct identity is KL = Σ p log(p/q) = Σ p log p - Σ p log q = L_ce - H. The sign error is not purely typographical: Box 2.8 explains the decomposition in detail and repeats it, and the text uses the one-hot special case (where H=0) to obscure the general error. This undermines the explanation of why cross-entropy is the appropriate classification loss. The error appears twice, so it is unlikely to be a one-off typo.","section":"§2.2.1, Eq. (2.19) and Box 2.8"},{"comment":"The layer transformation is dimensionally inconsistent. Eq. (2.5) states W^{(l)} is an N^{(l-1)} × N^{(l)} matrix and writes z ← W^{(l)}z + b^{(l-1)} with z ∈ R^{dim(l)}. If z has dimension N^{(l)}, then W^{(l)}z has dimension N^{(l-1)}, not N^{(l)}; if the layer output h^{(l)} should have N^{(l)} neurons, the weight matrix must have shape N^{(l)} × N^{(l-1)}. Eq. (2.6) repeats the same mixing of indices, while Eq. (2.7) uses sums over N_l and N_{l-1} that imply the correct convention is W^{(l)}: R^{N_{l-1}} → R^{N_l}. The §2.3 feed-forward shapes (e.g., X W^{(1)} with W^{(1)} of shape m × N^{(1)}) follow yet another convention. This makes it impossible for a student to verify the matrix multiplications by the stated definitions.","section":"§2.1.1, Eqs. (2.5)–(2.7)"},{"comment":"Two errors affect the t-SNE explanation. First, the conditional probability p(i|j) in Eq. (3.5) uses σ_i in the numerator but σ_j in the denominator; the standard formulation uses a single bandwidth for the reference point j. Second, the text after Eq. (3.8) says 'When these distributions are identical (L = 1), we get our closeness property perfectly!'—but the KL divergence is 0 when p = q, and L is minimized, not equal to 1. Both errors are conceptual, not just notational, and they directly concern a method presented as a core unsupervised tool.","section":"§3.1.2, Eq. (3.5) and Eq. (3.8)"},{"comment":"The numerical demonstrations are not reproducible from the text. No code, data, or random seeds are provided for the MNIST (Figs. 2.7–2.8), CIFAR-10 (Figs. 2.10, 2.13–2.15), or 2D Ising (Fig. 2.20) examples. Hyperparameters are incomplete or inconsistent: the MNIST experiment in §2.3.1 states η = 50, but §2.3.2 says 'same learning rate η = 1'; Fig. 2.7's caption says the input layer contains '282 nodes' where the text says 784. A student cannot reconstruct the claimed accuracies (56%, 95%, 82%, etc.) from the information given. This directly violates the preface's promise of 'hands-on examples' and weakens the pedagogical claim.","section":"§2.3, §2.5.3"}],"minor_comments":[{"comment":"The gradient-descent update is written as θ_{t+1} ← θ_t − η L(θ_t)/∂θ_t. The denominator should be ∂L(θ_t)/∂θ_t. As printed, the expression is not a valid derivative.","section":"§2.2.2, Eq. (2.25)"},{"comment":"The bias update line read 'b ← b − η ∂L/∂b = W − η ...'; the right-hand side should start with b, not W. This is a typo but confusing in a derivation.","section":"§2.3.1, Eq. (2.30)"},{"comment":"The Ising-model deep-learning demonstration does not cite the original works on machine-learning identification of phases (e.g., Carrasquilla and Melko, or related studies). For a set of lecture notes, this is a notable citation gap.","section":"§2.5.3"},{"comment":"The algorithm's input line says current RBM weights θ = {a, b, J}, while the text and Box 3.4 use W_{ij} for the coupling matrix. The notation should be unified.","section":"§3.6, Algorithm 4"},{"comment":"The t-SNE kernel is written as q_{ij} = (1 + ||y_i - y_j||^2)^{-1} / Σ_{k≠l}(1 + ||y_k - y_l||^2)^{-1}. The denominator should be normalized over ordered pairs (k,l) with k≠l, and the text should note that the numerator is the unnormalized Student-t kernel before division. This is minor but worth a clarifying sentence.","section":"§3.1.2, Eq. (3.7)"},{"comment":"There are many typographical and terminology issues: 'Principle Component Analysis' should be 'Principal Component Analysis'; 'Bolztmann' and 'Lecunn' are misspelled; Box 3.5 is labelled 'Box 3.4' in the text; 'mathematica' appears uncapitalized. None of these block the main points, but they should be cleaned up in a revision.","section":"Various"},{"comment":"The box calls the distance function a 'well defined metric' after listing only positivity and zero-on-equality. A metric also requires symmetry and the triangle inequality. The statement is mathematically imprecise.","section":"Box 2.6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a lecture-notes preprint rather than a research contribution. If the venue is a research journal, the lack of original results is a scope concern; if the venue publishes pedagogical reviews or lecture notes, the material is appropriate in principle. My recommendation of major_revision is based on the equation-level errors in the first three chapters, which are load-bearing for the central pedagogical claim. They are individually correctable, so rejection is not warranted, but the paper as it stands cannot be used as a reliable teaching resource. I did not audit Chapters 4–5 in detail; the errors I found in Chapters 2–3 already require a substantive revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nHere's my take on arXiv:2508.10666. It's explicitly lecture notes, not a research paper, and that's the right frame. There is nothing new here—universal approximation, backprop, CNNs, VAEs, GANs, RBMs are all textbook—but the notes are readable, the order is sensible, and the physics connections (Ising, Onsager, later quantum chapters) are what will draw people in. The Chinese Room section is a nice non-technical anchor, and the reparametrization trick / Gibbs sampling boxes are genuinely pedagogically helpful.\n\nThe soft spots are real and load-bearing for a teaching document. Eq. (3.24) defines Z = Σ E(z) instead of Z = Σ exp(−E(z)). That's not a harmless typo: the RBM derivation that follows (Eqs. 3.36–3.39) factors the hidden sum as if Z = Σ exp(−E(v,h)), which is inconsistent with the printed definition. A student who implements Eq. (3.24) will compute nonsense gradients. Eq. (2.19) gives KL = H + Lce, but with H defined as Shannon entropy, the correct relation is KL = Lce − H. Box 2.8 repeats the same mistake. Eq. (2.5) has mismatched matrix dimensions and bias index. These are all individually correctable, but together they undermine the central claim that the notes are a 'complete, detailed description' you can work through.\n\nI also could not find the quantum chapters (4–5) in the text I was given. The table of contents lists them, but the content isn't in this version. For quant-ph readers that's the entire point. If they're simply missing from the posted version, that's a serious incompleteness; if they're present but were cut off in my copy, I can't evaluate them. Either way, the version under review does not deliver the advertised scope.\n\nThe citation pattern is fine. They cite the original literature (Hinton, Kingma, Goodfellow, Onsager) and don't over-rely on self-citations.\n\nMy verdict: this deserves a serious referee, but the authors need to do a careful proofreading pass. The errors are concrete and will mislead the target reader. After corrections, I would happily point students toward it. As is, I'd hold it back.\n\nBest.","headline":"Readable, honest lecture notes, but several concrete equation errors (RBM partition function, KL decomposition) undermine the promised 'complete, detailed' pedagogy; fixable, but not yet ready for students.","tokens_in":51846,"tokens_out":2636,"would_cite":false,"duration_ms":27808,"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":"These lecture notes aim to give graduate students a complete, detailed introduction to deep learning for quantum applications.","keywords":["deep learning","quantum physics","neural networks","unsupervised learning","generative models","quantum state tomography","variational quantum eigensolver","Ising model"],"falsifier":"A reader can settle the central claim by opening the notes and trying to reproduce the examples. Eq. (3.24) defines the RBM partition function as $Z = \\sum_z E(z)$; Boltzmann normalisation requires $Z = \\sum_z e^{-E(z)}$, so the stated $p(z) = e^{-E(z)}/Z$ is not a valid probability distribution. The text after Eq. (3.8) says the KL divergence equals 1 when distributions are identical; the correct value is 0. Eq. (2.5) gives mismatched matrix dimensions and a bias index. Any one of these checks shows the notes are not yet complete and detailed as written.","tokens_in":50869,"feed_emoji":"⚛️","tokens_out":9042,"duration_ms":90582,"temperature":0.7,"pith_summary":"These lecture notes set out to be a complete, detailed introduction to deep learning for graduate students in physics, mathematics, engineering, and computer science, with quantum science as the target application area. The substantive claim is that deep learning can be taught as a transferable sequence of skills — building neural networks as tunable nonlinear functions, training them with gradient descent and automatic differentiation, and applying them to high-dimensional physical problems such as phase classification, generative modelling, and quantum state reconstruction. The notes also argue that deep learning can locate structure in data (for example, estimating the critical temperature of the 2D Ising model without using an order parameter) but does not by itself provide physical understanding, a boundary illustrated with the Chinese Room experiment. A sympathetic reader would come away with a concrete sense of when and how to use deep learning in quantum problems and what pitfalls to avoid; the text itself invites readers to report typos in its formulae.","feed_headline":"From neural networks to quantum tomography in one course","feed_subtitle":"A graduate-level text that builds deep learning from first principles and applies it to quantum systems.","key_machinery":"The load-bearing machinery is the layered neural network $f_\\theta(x) = h^{(L)} \\circ \\cdots \\circ h^{(1)}(x)$, where each layer applies a linear map (weight matrix and bias) followed by a nonlinear activation. Training is driven by gradient descent on a loss function, with gradients computed by automatic differentiation (backpropagation), and the reparametrisation trick extends differentiability to sampling from latent distributions, enabling VAEs, GANs, restricted Boltzmann machines, and normalising flows. These components carry the notes' argument because every later quantum application — phase classification, state representation, and tomography — is framed either as regression/classific","core_discovery":"The notes aim to establish a pedagogical result: a graduate student can go from no familiarity with neural networks to working knowledge of deep learning for quantum problems by following a single progressive sequence. The chain of machinery is a neural network as a tunable nonlinear function $f_\\theta$; the universal approximation theorem as the licence to use such functions; backpropagation and automatic differentiation as the training machinery; and the reparametrisation trick as the bridge to probabilistic generative models. On the quantum side, the claim is that these tools apply to concrete tasks: three-qubit entanglement classification, phase-diagram discovery with autoencoders, varia","pith_inferences":["The notes leave implicit that the confidence-dip method for estimating $T_c$ is a general recipe: any classifier trained on well-separated phases could locate an unknown transition by monitoring where its softmax confidence collapses.","Because quantum mechanics is itself a probabilistic theory, the reparametrisation trick used for VAEs is a natural bridge to quantum applications; the notes do not pursue reparametrisations tailored to quantum state manifolds.","If the promised code and datasets were supplied, the same chapter sequence could be reused as a template for other data-intensive sciences, not only quantum physics.","A testable extension suggested by the notes' structure is to benchmark neural-network quantum states against normalising-flow tomography on the same small system to compare sample efficiency and reconstruction fidelity."],"forward_implications":["A student who works through the sequence can implement standard deep-learning models (multilayer perceptrons, CNNs, autoencoders, VAEs, GANs, RBMs, normalising flows) and deploy them on physics data.","The confidence-drop method gives a model-independent way to estimate a critical temperature when no order parameter is known, benchmarked on the 2D Ising model and extendable to other transitions.","Quantum applications such as variational quantum eigensolvers, QAOA, neural-network quantum states, Hamiltonian learning, and photonic tomography are cast as concrete deep-learning tasks, so the same training principles apply across them.","The notes' Chinese Room argument sets expectations: a deep-learning model can generalise and predict, but on its own it does not supply causal understanding of the underlying physics.","The notes identify overfitting, interpretability, and noise sensitivity as the main practical constraints a quantum scientist must mitigate when applying deep learning."],"supporting_citations":[{"why":"Supplies the universal approximation theorem, the theoretical basis for using neural networks as function approximators.","marker":"[26]"},{"why":"Introduces convolutional neural networks, the architecture used for MNIST/CIFAR classification and Ising phase classification.","marker":"[27]"},{"why":"Provides t-SNE, the unsupervised nonlinear dimensionality-reduction method that motivates the later autoencoder discussion.","marker":"[33]"},{"why":"Introduces the variational autoencoder and reparametrisation trick, which makes sampling differentiable and underpins the generative models used later.","marker":"[35]"},{"why":"Defines the generative adversarial network framework used to illustrate adversarial training and mode collapse.","marker":"[39]"},{"why":"Supplies the energy-based model underlying restricted Boltzmann machines, connecting statistical physics to generative models.","marker":"[40, 41]"},{"why":"Contributes the Chinese Room thought experiment that the notes use to argue deep learning lacks semantic understanding.","marker":"[32]"}],"fun_headline_variants":["From no ML to quantum deep learning in one course","Deep learning for quantum physics: a single progressive path","Master quantum machine learning from first principles","One course: from neural network basics to quantum applications"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The promise that these notes form a complete, detailed, self-contained description rests on the assumption that the printed equations and numerical examples can be trusted and reproduced as written; as published, the RBM partition function in Eq. (3.24) ($Z = \\sum_z E(z)$), the t-SNE KL-divergence statement after Eq. (3.8), and the layer dimensions in Eq. (2.5) contain errors, and the hands-on examples include no code or data.","fun_headline_variants_meta":{"raw":{"variants":["From no ML to quantum deep learning in one course","Deep learning for quantum physics: a single progressive path","Master quantum machine learning from first principles","One course: from neural network basics to quantum applications"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000362,"raw_usage":{"total_tokens":1764,"prompt_tokens":693,"completion_tokens":1071,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":1011}},"tokens_in":437,"tokens_out":1071,"duration_ms":9752,"temperature":1.0,"reasoning_tokens":1011,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:18:25.482581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader can settle the central claim by opening the notes and trying to reproduce the examples. Eq. (3.24) defines the RBM partition function as $Z = \\sum_z E(z)$; Boltzmann normalisation requires $Z = \\sum_z e^{-E(z)}$, so the stated $p(z) = e^{-E(z)}/Z$ is not a valid probability distribution. The text after Eq. (3.8) says the KL divergence equals 1 when distributions are identical; the correct value is 0. Eq. (2.5) gives mismatched matrix dimensions and a bias index. Any one of these checks shows the notes are not yet complete and detailed as written.","supporting_citations":[{"cited_title":"Lecun, L","cited_arxiv_id":null,"evidence_quote":"Introduces convolutional neural networks, the architecture used for MNIST/CIFAR classification and Ising phase classification."},{"cited_title":"Gener- ative adversarial nets.Advances in neural information processing systems, 27, 2014","cited_arxiv_id":null,"evidence_quote":"Defines the generative adversarial network framework used to illustrate adversarial training and mode collapse."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the Chinese Room thought experiment that the notes use to argue deep learning lacks semantic understanding."}],"review_version":1}