From hyperplanes to hyperellipsoids: characterizing the inherent interpretability of linear and single-qubit mixed-state binary classification models

Kaitlin Gili

REVIEW 3 major objections 5 minor 23 references

A single-qubit mixed-state classifier is the ellipsoid version of standard linear classification: it learns a hyperellipsoid rather than a hyperplane.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 23:22 UTC pith:XA3QSYSK

load-bearing objection A correct but narrowly scoped teaching note: the ellipsoid claim holds only for the Pauli-Z measurement, and the model is a constrained linear classifier on squared features — worth fixing the abstract, then publishable as pedagogy. the 3 major comments →

arxiv 2607.15433 v1 pith:XA3QSYSK submitted 2026-07-16 cs.LG quant-ph

From hyperplanes to hyperellipsoids: characterizing the inherent interpretability of linear and single-qubit mixed-state binary classification models

Kaitlin Gili This is my paper

classification cs.LG quant-ph MSC 68T0581P6862H30

keywords interpretabilitylinear classificationsingle-qubit modelmixed statehyperellipsoidbinary classificationquantum machine learningamplitude encoding

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proves an equivalence: a binary classifier built from a single qubit mixed state, with real amplitude encoding and a standard Z measurement, reduces to f_p(x)=2Σ p_d x_d² −1. The decision boundary is therefore a hyperellipsoid centered at the origin, with semi-axis lengths a_d=1/√(2p_d). This makes the qubit model the 'ellipsoid version' of a linear model, and it lets instructors introduce quantum ML as a one-step geometric extension of linear classification. The two models also differ in interpretability: linear weights carry absolute feature importance (sign and magnitude), while the qubit's probability weights carry only relative importance because they are nonnegative and sum to one.

Core claim

The central claim is that a single-qubit mixed-state classification model, under real amplitude encoding and the diagonal computational-basis measurement Z, is not a mysterious quantum object but a hyperellipsoid learner. Substituting the diagonal measurement into the trace rule collapses the state's off-diagonal terms, leaving only the squared feature contributions, so the model output is 2Σ p_d x_d² −1. Setting this to zero yields the hyperellipsoid Σ x_d²/a_d² =1 with a_d=1/√(2p_d). Thus training the probabilities p_d is geometrically equivalent to stretching or shrinking the ellipsoid along each feature direction. The author argues this gives the model a distinct inductive bias — curved,

What carries the argument

The central object is the mixed-state density matrix v_mixed=Σ p_d v_pure(x_d)v_pure^T(x_d) together with the diagonal measurement operator Z=diag(1,−1). The trace rule Tr(v_mixed Z) produces the output; because Z is diagonal, only the diagonal entries of the density matrix survive, removing all off-diagonal coherence terms. The normalization Tr(v_mixed)=1 forces the trainable weights p_d onto the probability simplex, and setting the output to zero gives the hyperellipsoid equation whose semi-axes are determined by the p_d.

Load-bearing premise

The hyperellipsoid conclusion depends on fixing the measurement to the diagonal Pauli-Z operator; with any other measurement the output would also include off-diagonal (coherence) terms and the decision boundary would no longer be a centered ellipsoid.

What would settle it

Compute Tr(v_mixed X) for X = [[0,1],[1,0]] with the same mixed state; the output is 2β Σ p_d x_d, a linear decision boundary. If instead the model uses a measurement with off-diagonal entries, the 'ellipsoid version' characterization fails, showing the claim rests on the Z readout.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A single-qubit mixed-state classifier is classically simulatable and equivalent to a linear model on squared features with nonnegative weights summing to one; no quantum hardware is needed to implement it.
The model's decision boundary can curve around data (e.g., concentric groups) where a linear classifier fails, but it is always centered at the origin and cannot translate, so offset clusters are linearly separable but not ellipsoid-separable.
Feature importance in the qubit model is inherently relative: weights are positive and sum to one, so a feature's influence is always compared to the others, in contrast to the absolute sign-and-magnitude importance of linear weights.
Practically, probabilities should be lower-bounded (e.g., at 10⁻⁹) to prevent the ellipsoid from becoming unbounded along a feature direction when p_d →0.
For ML pedagogy, the derivation offers a route to quantum ML that requires no quantum background, only linear classification and the geometry of ellipsoids.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The equivalence is specific to the Z measurement: substituting a different Hermitian observable, such as X, brings off-diagonal density-matrix terms back into the output (e.g., Tr(v_mixed X)=2βΣp_d x_d), producing a linear rather than ellipsoidal boundary; so the hyperellipsoid characterization is a property of the Z readout, not of the single-qubit state alone.
The squared-feature form suggests immediate testable extensions: the same proof style could characterize multi-qubit or non-linear encodings, and the model's generalization behavior on datasets whose Bayes boundary is an ellipsoid versus a translated ellipsoid could be measured empirically.
The relative-importance bias is similar to the softmax normalization in attention mechanisms, which the paper notes; one could test whether this bias explains attention weights' known sensitivity to input scaling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A correct but narrowly scoped teaching note: the ellipsoid claim holds only for the Pauli-Z measurement, and the model is a constrained linear classifier on squared features — worth fixing the abstract, then publishable as pedagogy.

read the letter

Short version: the paper derives, correctly, that a specific single-qubit mixed-state classifier — amplitude-encoded real features, probabilities as trainable weights, and a Pauli-Z readout — gives f(x) = 2∑p_d x_d^2 − 1, so the decision surface is a hyperellipsoid centered at the origin with semi-axes 1/√(2p_d). That math checks out. What the paper does well is present it cleanly and without quantum jargon, with a nice geometric story about stretching and shrinking the ellipsoid. The relative-importance framing (probabilities sum to one, so weights trade off) is a genuinely useful way to talk about this model in a classroom.

The soft spots are all about scope and framing. First, the abstract and title claim this is 'the ellipsoid version' of a single-qubit mixed-state model, but the result depends entirely on the choice of measurement Z. Choose X or any off-diagonal observable and the boundary isn't a quadratic surface at all — off-diagonal density-matrix terms become visible and the simple ellipsoid picture evaporates. The body does say 'if we choose Z', so a careful reader won't be misled, but the headline is broader than the proof. Second, calling it the 'ellipsoid version' of a standard linear model is a stretch. A standard linear model has a trainable bias and unconstrained weights; this model has fixed bias −1 and weights constrained to the probability simplex. So it's a proper subset of linear classifiers on the feature map x→(x_1^2,…,x_D^2). That is a special case, not a parallel version. The paper even notes the lack of translation, so it's not hidden, but the abstract's framing overstates it. Third, the 'inherent interpretability' discussion is qualitative; the relative vs absolute importance distinction is asserted, not formalized or tested. That's fine for a teaching note, but the paper should say 'we conjecture' or 'we suggest' rather than implying a demonstrated property.

The novelty is small: Eq. 3 is a known polynomial-feature construction. What's arguably new is the specific model (trainable probabilities as weights) and the geometric teaching angle. The derivation is self-contained, no circularity, and the author is honest that the model is classically simulatable and that empirical work is a future direction. Citation pattern looks fine, apart from the self-citation to the companion perspective piece, which is used only for motivation.

Net: this is a solid, short pedagogical note, not a research breakthrough. It deserves a serious referee at an education-focused venue, with the expectation that the abstract gets qualified and the equivalence claim gets toned down. I'd also ask the author to add a sentence acknowledging that other measurements break the ellipsoid picture, and to define precisely what 'relative feature importance' means. Would I cite it? Probably not in my own research, but I'd recommend it to instructors. Bring to reading group? Maybe, as a prompt to talk about framing and scope.

Referee Report

3 major / 5 minor

Summary. The paper presents a short pedagogical characterization comparing a standard linear classifier and a single-qubit mixed-state classifier. With amplitude encoding of each feature into the α coefficient of a real pure state, a convex mixture with trainable probabilities p_d, and a Pauli-Z measurement, the model output is shown in Eq. 3 to be f_p(x) = 2 Σ_d p_d x_d^2 − 1, so the decision boundary is the axis-aligned hyperellipsoid Σ_d x_d^2 / a_d^2 = 1 with a_d = 1/√(2p_d) (Eq. 4). The paper argues that the qubit model is consequently the 'ellipsoid version' of linear classification, has relative rather than absolute feature importance, and offers a gentle entry point for teaching quantum ML. The authors explicitly note that the model is classically simulatable and frame the work as conceptual/pedagogical rather than empirical.

Significance. If the characterization is accepted, the paper provides a clean and genuinely accessible bridge between classical linear classification and a simple quantum-inspired model, with an explicit geometric picture and a transparent derivation. The comparison of absolute vs. relative feature importance is a useful conceptual contribution for interpretability discussions. The paper is honest about the model being classically simulatable and does not oversell practical quantum advantage. The derivation itself is short, self-contained, and, under the stated assumptions, correct. The paper's value lies in pedagogy and conceptual clarity, not in novel mathematical content; this is appropriate given the stated goals.

major comments (3)

[Abstract and §2.2, Eq. (3)] The headline claim that the single-qubit mixed-state model is 'just the ellipsoid version' of a linear model is only established for the fixed measurement operator O = Z. Because Z is diagonal, Tr(v_mixed Z) depends only on the diagonal entries of v_mixed; the off-diagonal coherence terms are invisible. As the skeptic notes, with a generic observable such as Pauli-X the output would include terms p_d x_d √(1−x_d²), whose zero set is not a quadratic hypersurface and hence not an ellipsoid. The full text does specify the choice of Z, but the abstract and title state the conclusion without this qualification, so the central claim is broader than what Eq. (3)–(4) prove. I recommend either restricting the abstract/title to 'for a Pauli-Z measurement' or expanding the derivation to characterize the full set of single-qubit measurements and show which yield ellipsoidal boundaries.
[§2.2, Eq. (4) and following paragraph] The equivalence to a linear model on squared features is itself a constrained equivalence: the qubit model has no bias term (the constant −1 is fixed, not learned), nonnegative weights p_d, and the simplex constraint Σ p_d = 1. Thus the set of representable decision boundaries is a strict subset of centered, axis-aligned ellipsoids that contain the origin on the boundary. The paper mentions the restriction imposed by normalization, but it does not spell out that this means the model cannot represent translated or rotated ellipsoids, nor a general linear classifier on the raw features. This is not an error, but the 'just the ellipsoid version' phrasing understates the gap in expressivity. Please qualify the comparison to say 'a constrained linear model on squared features with fixed bias −1'.
[§2.2, paragraph after Eq. (4)] The statement 'the hyperellipsoid is most compressed along a feature direction when p_d = 1' is imprecise when D > 1. For fixed Σ p_d = 1, setting p_d = 1 forces all other probabilities to zero, so the boundary becomes a cylinder in all other directions. The sentence should be phrased in terms of varying one probability while holding the others fixed. Similarly, the 'hyper-sphere when p_d = 1/D' statement is correct only because all probabilities are equal; the radius r = √(D/2) is correct. These are minor clarifications, but they affect the geometric intuition the paper aims to convey.

minor comments (5)

[§2.2, Eq. (2)] The notation β is used both as the second amplitude coefficient and as the bias term in Eq. (1). Since the bias is β and the qubit amplitude is also β, this is confusing. Please rename the amplitude, e.g., γ or β_qubit.
[§2.2, Eq. (3)] There is a missing absolute value in the intermediate step: the diagonal entries should be |x_d|² and 1 − |x_d|². Since the text restricts to real data, x_d² and 1 − x_d² are fine, but the absolute value notation is used in Eq. (2) and then dropped in Eq. (3). Please keep the notation consistent.
[§2.2, last sentence of the hyperellipsoid paragraph] Minor typographical: 'For the reader who is interesting in learning more' should be 'interested.'
[References] Reference [1] is a joint paper with Z. P. Bradshaw, but the main text says 'in a recent perspective piece [1], we argue...' The use of 'we' for a single-author manuscript is fine, but the reference is not obviously the author's own prior piece. Consider clarifying the citation or adding a sentence that the perspective piece is by the author and coauthor, if that is the case.
[Figure 1] The figure is not described in enough detail in the text. It would help to state explicit toy data parameters and the values of p_d or w that are plotted, so readers can reproduce the illustration.

Circularity Check

0 steps flagged

No significant circularity: the hyperellipsoid result follows algebraically from explicitly stated model definitions; the only self-citation is motivational.

full rationale

The derivation is self-contained. The single-qubit mixed-state output f_p(x) = 2 sum_d p_d x_d^2 - 1 is obtained in Eq. (3) by a direct trace computation with the explicitly chosen observable Z, and the decision boundary in Eq. (4) is just the rearrangement sum_d x_d^2 / a_d^2 = 1 with a_d = 1/sqrt(2p_d). No parameter is fitted to any subset of data and then renamed as a prediction; no uniqueness theorem or prior work is invoked to force the result; and no ansatz is smuggled in through citation. The choice of the Pauli-Z measurement is stated explicitly in the text ('If we choose a standard binary computational measurement Z=...'), and while the abstract's phrasing is broader than the proven conditional, that is a scope/correctness limitation rather than a circular reduction. The only self-citation, Ref. [1], is used in the Introduction as motivation for the research direction, not as a premise of the mathematical derivation. The central claim therefore has independent content.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The derivation is self-contained linear algebra given the model definition. The only non-standard input is the specific model architecture (amplitude encoding, Pauli-Z measurement, trainable probabilities), which is chosen by the author and is the source of the ellipsoid inductive bias.

axioms (5)

domain assumption Data features are min-max normalized to [0,1] before encoding as α = x_d.
Required for α to be a valid probability amplitude with |α|^2 ≤ 1; stated in §2.1 and §2.2.
standard math Pure single-qubit states are unit vectors: |α|^2 + |β|^2 = 1, with α, β real.
Standard quantum state normalization; used to set |β|^2 = 1 - x_d^2 in Eq. 2.
domain assumption The mixed state is a convex combination of pure-state outer products with probabilities p_d, so Tr(v_mixed) = 1 and ∑p_d = 1.
Standard density-matrix formalism; carries the normalization constraint central to the relative-importance claim.
domain assumption The measurement observable is fixed as Pauli-Z, so f_p(x) = Tr(v_mixed Z).
This choice makes off-diagonal elements inert and yields the quadratic form in Eq. 3; the ellipsoid conclusion depends on it.
ad hoc to paper Each feature x_d is encoded exclusively in the α coefficient of its own pure state (amplitude encoding).
This encoding, combined with the Z measurement, produces the squared-feature dependence; other encodings would give different boundaries.

pith-pipeline@v1.3.0-alltime-deepseek · 5309 in / 16596 out tokens · 154356 ms · 2026-08-01T23:22:45.536707+00:00 · methodology

0 comments

read the original abstract

We characterize and compare the inherent interpretability offerings of a standard linear model with a single qubit mixed state model for the task of supervised binary classification. A side by side comparison reveals that a single qubit mixed state model for binary classification is just the ``ellipsoid version" of standard linear model classification. More precisely, rather than learning a hyperplane to classify data, we learn a hyperellipsoid. We discuss the consequences of the geometric inductive biases of both models, as well as how each model contains a different feature importance inductive bias. This short characterization offers an accessible route to quantum machine learning (ML) ideas for readers who have zero background in quantum and are only familiar with linear classification in ML. In support of ML pedagogy, we encourage instructors to utilize this piece to smoothly introduce quantum ML ideas into the undergraduate ML classroom.

Figures

Figures reproduced from arXiv: 2607.15433 by Kaitlin Gili.

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 4 canonical work pages

[1]

Inherent Interpretability Provides Inherent Value in Quantum Machine Learning

Kaitlin Gili and Zachary P. Bradshaw. “Inherent Interpretability Provides Inherent Value in Quantum Machine Learning”. In:arXiv preprint arXiv:2607.13827(2026). arXiv:2607.13827

Pith/arXiv arXiv 2026
[2]

Inherently Interpretable Machine Learning: A Contrasting Paradigm to Post-hoc Explainable AI

Patrick Zschech, Sven Weinzierl, and Mathias Kraus. “Inherently Interpretable Machine Learning: A Contrasting Paradigm to Post-hoc Explainable AI”. In:Business & Information Systems Engineering (2025). Received 17 Dec 2024; accepted 24 Jul 2025; published 15 Sep 2025, pp. 1–19.DOI:10.1007/ s12599-025-00964-0.URL:https://doi.org/10.1007/s12599-025-00964-0

doi:10.1007/s12599-025-00964-0 2025
[3]

Recognizing mechanistic reasoning in student scientific inquiry: A frame- work for discourse analysis developed from philosophy of science

Rosemary S. Russ et al. “Recognizing mechanistic reasoning in student scientific inquiry: A frame- work for discourse analysis developed from philosophy of science”. In:Science Education92.3 (2008), pp. 499–525.URL:https : / / onlinelibrary . wiley . com / doi / abs / 10 . 1002 / sce . 20264

2008
[4]

Bishop.Pattern Recognition and Machine Learning

Christopher M. Bishop.Pattern Recognition and Machine Learning. Information Science and Statistics. New York, NY: Springer, 2006.ISBN: 978-0-387-31073-2.URL:https://link.springer.com/ book/9780387310732

arXiv 2006
[5]

Trevor Hastie, Robert Tibshirani, and Jerome Friedman.The Elements of Statistical Learning: Data Min- ing, Inference, and Prediction. 2nd ed. Springer Series in Statistics. New York, NY: Springer, 2009.ISBN: 978-0-387-84857-0.URL:https://doi.org/10.1007/978-0-387-84858-7. 4

doi:10.1007/978-0-387-84858-7 2009
[6]

Gareth James et al.An Introduction to Statistical Learning: with Applications in R. 2nd ed. Springer Texts in Statistics. New York: Springer, 2021.DOI:10.1007/978-1-0716-1418-1

doi:10.1007/978-1-0716-1418-1 2021
[7]

Nielsen and Isaac L

Michael A. Nielsen and Isaac L. Chuang.Quantum Computation and Quantum Information. 10th An- niversary. Cambridge: Cambridge University Press, 2010.ISBN: 978-1-107-00217-3.DOI:10.1017/ CBO9780511976667

2010
[8]

LIBLINEAR: A Library for Large Linear Classification

Rong-En Fan et al. “LIBLINEAR: A Library for Large Linear Classification”. In:Journal of Machine Learning Research9.61 (2008), pp. 1871–1874.URL:https://www.jmlr.org/papers/v9/ fan08a.html

2008
[9]

Deep Residual Learning for Image Recognition

Kaiming He et al. “Deep Residual Learning for Image Recognition”. In:Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition. 2016, pp. 770–778.URL:https://doi.org/ 10.1109/CVPR.2016.90

doi:10.1109/cvpr.2016.90 2016
[10]

Jacob Devlin et al.BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
[11]

Understanding Intermediate Layers Using Linear Classifier Probes

Guillaume Alain and Yoshua Bengio. “Understanding Intermediate Layers Using Linear Classifier Probes”. In:arXiv preprint arXiv:1610.01644(2016). arXiv:1610.01644 [stat.ML]

Pith/arXiv arXiv 2016
[12]

Head2Toe: Utilizing Intermediate Representations for Better Transfer Learning

Utku Evci et al. “Head2Toe: Utilizing Intermediate Representations for Better Transfer Learning”. In: Proceedings of the 39th International Conference on Machine Learning. Ed. by Kamalika Chaudhuri et al. V ol. 162. Proceedings of Machine Learning Research. PMLR, 2022, pp. 6009–6033.URL:https: //proceedings.mlr.press/v162/evci22a.html

2022
[13]

Different Scaling of Linear Models and Deep Learning in UK Biobank Brain Images versus Machine-Learning Datasets

Marc-Andr ´e Schulz et al. “Different Scaling of Linear Models and Deep Learning in UK Biobank Brain Images versus Machine-Learning Datasets”. In:Nature Communications11.1 (Aug. 2020), p. 4238.URL: https://doi.org/10.1038/s41467-020-18037-z

work page doi:10.1038/s41467-020-18037-z 2020
[14]

Muhammad Arbab Arshad, Sakib Shahriar, and Khizar Anjum.The Power Of Simplicity: Why Simple Linear Models Outperform Complex Machine Learning Techniques – Case Of Breast Cancer Diagnosis
[15]

Creating Superpositions That Correspond to Efficiently Integrable Prob- ability Distributions

Lov Grover and Terry Rudolph. “Creating Superpositions That Correspond to Efficiently Integrable Prob- ability Distributions”. In: (2002). arXiv:quant-ph/0208112 [quant-ph]

Pith/arXiv arXiv 2002
[16]

Quantum Support Vector Machine for Big Data Classification

Patrick Rebentrost, Masoud Mohseni, and Seth Lloyd. “Quantum Support Vector Machine for Big Data Classification”. In:Physical Review Letters113.13 (Sept. 2014), p. 130503.URL:https : / / doi . org/10.1103/PhysRevLett.113.130503

doi:10.1103/physrevlett.113.130503 2014
[17]

A Quantum-Inspired Version of the Nearest Mean Classifier

Giuseppe Sergioli et al. “A Quantum-Inspired Version of the Nearest Mean Classifier”. In:Soft Comput- ing22.3 (2018), pp. 691–705.URL:https : / / link . springer . com / article / 10 . 1007 / s00500-016-2478-2

2018
[18]

An efficient geometric approach to quantum-inspired classifi- cations

Roberto Leporini and Davide Pastorello. “An efficient geometric approach to quantum-inspired classifi- cations”. In:Scientific Reports12.1 (2022), p. 8781.DOI:10.1038/s41598-022-12392-1

work page doi:10.1038/s41598-022-12392-1 2022
[19]

Quantum-Inspired Applications for Classification Problems

Cesarino Bertini and Roberto Leporini. “Quantum-Inspired Applications for Classification Problems”. In:Entropy25.3 (2023), p. 404.DOI:10.3390/e25030404

work page doi:10.3390/e25030404 2023
[20]

Data Re-Uploading for a Universal Quantum Classifier

Adri ´an P ´erez-Salinas et al. “Data Re-Uploading for a Universal Quantum Classifier”. In:Quantum4 (Feb. 2020), p. 226.URL:https://doi.org/10.22331/q-2020-02-06-226

doi:10.22331/q-2020-02-06-226 2020
[21]

What We Can Do with One Qubit in Quantum Machine Learning: Ten Classical Machine Learning Problems That Can Be Solved with a Single Qubit

Manuel P. Cu ´ellar. “What We Can Do with One Qubit in Quantum Machine Learning: Ten Classical Machine Learning Problems That Can Be Solved with a Single Qubit”. In:Quantum Machine Intelligence 6.2 (Nov. 2024), p. 76.URL:https://doi.org/10.1007/s42484-024-00210-y. 5

work page doi:10.1007/s42484-024-00210-y 2024
[2019]

arXiv:1810.04805 [cs.CL]

Pith/arXiv arXiv
[2023]

arXiv:2306.02449 [cs.LG]

Pith/arXiv arXiv

[1] [1]

Inherent Interpretability Provides Inherent Value in Quantum Machine Learning

Kaitlin Gili and Zachary P. Bradshaw. “Inherent Interpretability Provides Inherent Value in Quantum Machine Learning”. In:arXiv preprint arXiv:2607.13827(2026). arXiv:2607.13827

Pith/arXiv arXiv 2026

[2] [2]

Inherently Interpretable Machine Learning: A Contrasting Paradigm to Post-hoc Explainable AI

Patrick Zschech, Sven Weinzierl, and Mathias Kraus. “Inherently Interpretable Machine Learning: A Contrasting Paradigm to Post-hoc Explainable AI”. In:Business & Information Systems Engineering (2025). Received 17 Dec 2024; accepted 24 Jul 2025; published 15 Sep 2025, pp. 1–19.DOI:10.1007/ s12599-025-00964-0.URL:https://doi.org/10.1007/s12599-025-00964-0

doi:10.1007/s12599-025-00964-0 2025

[3] [3]

Recognizing mechanistic reasoning in student scientific inquiry: A frame- work for discourse analysis developed from philosophy of science

Rosemary S. Russ et al. “Recognizing mechanistic reasoning in student scientific inquiry: A frame- work for discourse analysis developed from philosophy of science”. In:Science Education92.3 (2008), pp. 499–525.URL:https : / / onlinelibrary . wiley . com / doi / abs / 10 . 1002 / sce . 20264

2008

[4] [4]

Bishop.Pattern Recognition and Machine Learning

Christopher M. Bishop.Pattern Recognition and Machine Learning. Information Science and Statistics. New York, NY: Springer, 2006.ISBN: 978-0-387-31073-2.URL:https://link.springer.com/ book/9780387310732

arXiv 2006

[5] [5]

Trevor Hastie, Robert Tibshirani, and Jerome Friedman.The Elements of Statistical Learning: Data Min- ing, Inference, and Prediction. 2nd ed. Springer Series in Statistics. New York, NY: Springer, 2009.ISBN: 978-0-387-84857-0.URL:https://doi.org/10.1007/978-0-387-84858-7. 4

doi:10.1007/978-0-387-84858-7 2009

[6] [6]

Gareth James et al.An Introduction to Statistical Learning: with Applications in R. 2nd ed. Springer Texts in Statistics. New York: Springer, 2021.DOI:10.1007/978-1-0716-1418-1

doi:10.1007/978-1-0716-1418-1 2021

[7] [7]

Nielsen and Isaac L

Michael A. Nielsen and Isaac L. Chuang.Quantum Computation and Quantum Information. 10th An- niversary. Cambridge: Cambridge University Press, 2010.ISBN: 978-1-107-00217-3.DOI:10.1017/ CBO9780511976667

2010

[8] [8]

LIBLINEAR: A Library for Large Linear Classification

Rong-En Fan et al. “LIBLINEAR: A Library for Large Linear Classification”. In:Journal of Machine Learning Research9.61 (2008), pp. 1871–1874.URL:https://www.jmlr.org/papers/v9/ fan08a.html

2008

[9] [9]

Deep Residual Learning for Image Recognition

Kaiming He et al. “Deep Residual Learning for Image Recognition”. In:Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition. 2016, pp. 770–778.URL:https://doi.org/ 10.1109/CVPR.2016.90

doi:10.1109/cvpr.2016.90 2016

[10] [10]

Jacob Devlin et al.BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

[11] [11]

Understanding Intermediate Layers Using Linear Classifier Probes

Guillaume Alain and Yoshua Bengio. “Understanding Intermediate Layers Using Linear Classifier Probes”. In:arXiv preprint arXiv:1610.01644(2016). arXiv:1610.01644 [stat.ML]

Pith/arXiv arXiv 2016

[12] [12]

Head2Toe: Utilizing Intermediate Representations for Better Transfer Learning

Utku Evci et al. “Head2Toe: Utilizing Intermediate Representations for Better Transfer Learning”. In: Proceedings of the 39th International Conference on Machine Learning. Ed. by Kamalika Chaudhuri et al. V ol. 162. Proceedings of Machine Learning Research. PMLR, 2022, pp. 6009–6033.URL:https: //proceedings.mlr.press/v162/evci22a.html

2022

[13] [13]

Different Scaling of Linear Models and Deep Learning in UK Biobank Brain Images versus Machine-Learning Datasets

Marc-Andr ´e Schulz et al. “Different Scaling of Linear Models and Deep Learning in UK Biobank Brain Images versus Machine-Learning Datasets”. In:Nature Communications11.1 (Aug. 2020), p. 4238.URL: https://doi.org/10.1038/s41467-020-18037-z

work page doi:10.1038/s41467-020-18037-z 2020

[14] [14]

Muhammad Arbab Arshad, Sakib Shahriar, and Khizar Anjum.The Power Of Simplicity: Why Simple Linear Models Outperform Complex Machine Learning Techniques – Case Of Breast Cancer Diagnosis

[15] [15]

Creating Superpositions That Correspond to Efficiently Integrable Prob- ability Distributions

Lov Grover and Terry Rudolph. “Creating Superpositions That Correspond to Efficiently Integrable Prob- ability Distributions”. In: (2002). arXiv:quant-ph/0208112 [quant-ph]

Pith/arXiv arXiv 2002

[16] [16]

Quantum Support Vector Machine for Big Data Classification

Patrick Rebentrost, Masoud Mohseni, and Seth Lloyd. “Quantum Support Vector Machine for Big Data Classification”. In:Physical Review Letters113.13 (Sept. 2014), p. 130503.URL:https : / / doi . org/10.1103/PhysRevLett.113.130503

doi:10.1103/physrevlett.113.130503 2014

[17] [17]

A Quantum-Inspired Version of the Nearest Mean Classifier

Giuseppe Sergioli et al. “A Quantum-Inspired Version of the Nearest Mean Classifier”. In:Soft Comput- ing22.3 (2018), pp. 691–705.URL:https : / / link . springer . com / article / 10 . 1007 / s00500-016-2478-2

2018

[18] [18]

An efficient geometric approach to quantum-inspired classifi- cations

Roberto Leporini and Davide Pastorello. “An efficient geometric approach to quantum-inspired classifi- cations”. In:Scientific Reports12.1 (2022), p. 8781.DOI:10.1038/s41598-022-12392-1

work page doi:10.1038/s41598-022-12392-1 2022

[19] [19]

Quantum-Inspired Applications for Classification Problems

Cesarino Bertini and Roberto Leporini. “Quantum-Inspired Applications for Classification Problems”. In:Entropy25.3 (2023), p. 404.DOI:10.3390/e25030404

work page doi:10.3390/e25030404 2023

[20] [20]

Data Re-Uploading for a Universal Quantum Classifier

Adri ´an P ´erez-Salinas et al. “Data Re-Uploading for a Universal Quantum Classifier”. In:Quantum4 (Feb. 2020), p. 226.URL:https://doi.org/10.22331/q-2020-02-06-226

doi:10.22331/q-2020-02-06-226 2020

[21] [21]

What We Can Do with One Qubit in Quantum Machine Learning: Ten Classical Machine Learning Problems That Can Be Solved with a Single Qubit

Manuel P. Cu ´ellar. “What We Can Do with One Qubit in Quantum Machine Learning: Ten Classical Machine Learning Problems That Can Be Solved with a Single Qubit”. In:Quantum Machine Intelligence 6.2 (Nov. 2024), p. 76.URL:https://doi.org/10.1007/s42484-024-00210-y. 5

work page doi:10.1007/s42484-024-00210-y 2024

[22] [2019]

arXiv:1810.04805 [cs.CL]

Pith/arXiv arXiv

[23] [2023]

arXiv:2306.02449 [cs.LG]

Pith/arXiv arXiv