QuantumDeepLearningBlog

https://medium.com/@davidjeffreymerwin

---

Key sources in context

Background for the quantum deep-learning writing linked above and for the notebook in this repository, written as a compact annotated bibliography. None is used to claim a quantum advantage. ^E1

How to read each entry. The caption, metadata, descriptors, tags and labels form the catalog card. The abstract, gloss, commentary (with its elucidation and exposition), exegesis and critique are the reading. Marginalia point to files in this repository, scholia compare the sources with each other, and each entry has one footnote. The anchors (#b11, #b11-fn) are the markup. Code comments, docstrings, directives and decorators in this repository are listed after the entries.

Sources: B19 · B26 · B24 · B39 · B38 · B27 · B23 ^E2

B19 · TensorFlow Quantum: A Software Framework for Quantum Machine Learning

Caption. Cirq circuits as Keras layers: hybrid quantum-classical ML.

Metadata. Michael Broughton, Guillaume Verdon, Trevor McCourt, et al. (2020). arXiv:2003.02989. Preprint. arxiv.org/abs/2003.02989. Verified 10 Nov 2026.
Descriptors. Software toolchain · Quantum machine learning · Tags. tfq keras pqc hello-many-worlds · Labels. evidence:preprint role:tool type:preprint status:verified

• Abstract. Describes TensorFlow Quantum's design for representing circuits as tensors, differentiating parameterised circuits, and combining them with classical Keras models, with example applications.
• Gloss. PQC: Parameterised quantum circuit, the trainable layer. Parameter shift: A rule for exact gradients of circuit expectation values.
• Commentary. TensorFlow Quantum, the QML layer built on TensorFlow. Elucidation: lang_c.ipynb is TensorFlow's 'Install TensorFlow for C' guide. It is classical TensorFlow, not TFQ, and it sits under the same framework TFQ extends. Exposition: Encode data into circuits, apply a PQC, measure expectation values, and back-propagate through them alongside classical layers.
• Exegesis. A software paper: it shows what can be built, not that quantum layers improve accuracy.¹
• Critique. TFQ releases can lag current TensorFlow versions; check compatibility before running a copied notebook.
• Marginalia.
  • lang_c.ipynb: Classical C API setup (libtensorflow 2.15.0 CPU build). No quantum code in this file.
  • README.md: The Medium writing linked above is where TFQ ideas are discussed.
• Scholia. B24: The model family TFQ tutorials implement. B26: The field TFQ serves.
• Footnote. ¹ TFQ documentation: tensorflow.org/quantum.
• In context. The software link between this repository's TensorFlow notebook and quantum deep learning.

B26 · Quantum machine learning

Caption. The landmark QML review, caveats included.

Metadata. Jacob Biamonte, Peter Wittek, Nicola Pancotti, et al. (2017). Nature 549(7671): 195–202. Journal article. DOI: 10.1038/nature23474. Verified 10 Nov 2026.
Descriptors. Quantum machine learning · Tags. review qram linear-algebra-speedups · Labels. evidence:peer-reviewed role:foundation type:journal-article status:verified

• Abstract. Reviews quantum machine learning: quantum linear-algebra subroutines, quantum-enhanced models, annealing-based learning, and practical obstacles such as data loading.
• Gloss. QRAM: Hypothetical quantum random-access memory for loading classical data in superposition.
• Commentary. The landmark map of quantum machine learning. Elucidation: It sorts claims into three kinds: speedups from linear algebra, quantum-enhanced models, and annealing-based learning. Each comes with assumptions, such as fast data loading, that a blog post should state. Exposition: Speedups often come from fast linear algebra on quantum states; the cost of preparing those states can erase them.
• Exegesis. Lists potential speedups with their assumptions. It does not report demonstrated practical advantage.¹
• Critique. Several linear-algebra speedups it surveys were later weakened by 'dequantized' classical algorithms (from 2018), and most rely on qRAM-style data loading that does not yet exist. Read it as a map of hopes with stated assumptions.
• Marginalia.
  • README.md: Use as the survey frame for any post that promises a QML speedup.
• Scholia. B39: A concrete experiment in the kernel branch. B38: A broader, later review.
• Footnote. ¹ Later 'dequantization' work weakened several of the speedups surveyed.
• In context. Survey background for the blog.

B24 · Classification with Quantum Neural Networks on Near Term Processors

Caption. Parameterised circuits as binary classifiers.

Metadata. Edward Farhi, Hartmut Neven (2018). arXiv:1802.06002. Preprint. arxiv.org/abs/1802.06002. Verified 10 Nov 2026.
Descriptors. Quantum machine learning · Variational algorithms · Tags. qnn binary-classification mnist · Labels. evidence:preprint role:foundation type:preprint status:verified

• Abstract. Proposes a quantum neural network in which a parameterised circuit acts on an input state and a measured readout qubit gives the class label; tested in simulation, including downsampled MNIST.
• Gloss. Readout qubit: The qubit measured to produce the prediction.
• Commentary. The quantum neural network idea at the center of 'quantum deep learning'. Elucidation: It defines a QNN concretely: a parameterised circuit acting on a quantum state, with one readout qubit measured to give the label. Exposition: Encode input as a quantum state, apply trainable unitaries, measure a Pauli on the readout qubit, minimise a loss over parameters.
• Exegesis. Shows the model can learn small tasks in simulation. It does not show an advantage over classical networks.¹
• Critique. Simulated, small-scale; trainability caveats arrived soon after (B27).
• Marginalia.
  • README.md: Cite when a post explains what a QNN is.
• Scholia. B27: Why QNNs can be hard to train. B39: The kernel alternative.
• Footnote. ¹ Tested in simulation, including on downsampled MNIST.
• In context. The definition the blog's QNN posts build on.

B39 · Supervised learning with quantum-enhanced feature spaces

Caption. Quantum kernels on real hardware, with synthetic data.

Metadata. Vojtěch Havlíček, Antonio D. Córcoles, Kristan Temme, et al. (2019). Nature 567(7747): 209–212. Journal article. DOI: 10.1038/s41586-019-0980-2. Verified 10 Nov 2026.
Descriptors. Quantum machine learning · Gate-model computing · Tags. quantum-kernel feature-map svm · Labels. evidence:peer-reviewed role:foundation type:journal-article status:verified

• Abstract. Proposes two classifiers that use a quantum circuit as a feature map, a variational classifier and a quantum kernel estimator, and demonstrates both on a superconducting processor using two qubits and artificially constructed data.
• Gloss. Feature map: A circuit that encodes a data point into a quantum state. Kernel: Inner product of two encoded states, used by an SVM.
• Commentary. Quantum kernels demonstrated on real hardware. Elucidation: It gives the blog a second route besides variational QNNs: encode the data with a circuit, estimate overlaps between encoded points, then train a classical SVM. Exposition: Map data into a large Hilbert space with a hard-to-simulate circuit, estimate pairwise overlaps on hardware, then train a classical SVM.
• Exegesis. Shows the methods work on a small device with data designed to suit the feature map. It does not show advantage on real-world datasets.¹
• Critique. Two-qubit demo; later work found kernel concentration problems at scale analogous to barren plateaus.
• Marginalia.
  • README.md: A candidate topic for a post. Label it as literature, not as an experiment run here.
• Scholia. B24: The variational sibling. B27: Kernels face a similar concentration problem.
• Footnote. ¹ The experiment used two qubits and data built to suit the feature map.
• In context. Kernel-method background for the blog.

B38 · Variational quantum algorithms

Caption. The review of VQAs: QAOA, VQE, QNNs and their obstacles.

Metadata. M. Cerezo, Andrew Arrasmith, Ryan Babbush, et al. (2021). Nature Reviews Physics 3(9): 625–644. Journal article. DOI: 10.1038/s42254-021-00348-9. Verified 10 Nov 2026.
Descriptors. Variational algorithms · Trainability · Quantum machine learning · Tags. vqe qaoa ansatz error-mitigation · Labels. evidence:peer-reviewed role:foundation type:journal-article status:verified

• Abstract. Reviews variational quantum algorithms: cost functions, ansatz design, optimisers, applications from chemistry to optimization and ML, and challenges including barren plateaus, noise and error mitigation.
• Gloss. Ansatz: The parameterised circuit template. Error mitigation: Post-processing that reduces noise bias without full error correction.
• Commentary. The review that unifies variational quantum algorithms. Elucidation: QNNs, QAOA and VQE share one loop (circuit cost, classical update) and one set of obstacles. Posts can explain all three in one frame. Exposition: A quantum circuit evaluates a cost; a classical optimiser updates parameters; the loop repeats. Everything hinges on cost design and trainability.
• Exegesis. The review presents VQAs as leading candidates for near-term advantage. It does not report a demonstrated practical advantage.¹
• Critique. A 2021 snapshot; trainability theory has advanced since.
• Marginalia.
  • README.md: The survey frame for posts on variational methods.
• Scholia. B27: The obstacle it reviews. B23: The NISQ context.
• Footnote. ¹ Crossref lists 11 authors.
• In context. The map for variational topics.

B27 · Barren plateaus in quantum neural network training landscapes

Caption. Gradients that vanish exponentially: the QML counterweight.

Metadata. Jarrod R. McClean, Sergio Boixo, Vadim N. Smelyanskiy, et al. (2018). Nature Communications 9: 4812. Journal article. DOI: 10.1038/s41467-018-07090-4. Verified 10 Nov 2026.
Descriptors. Trainability · Quantum machine learning · Tags. barren-plateau counterevidence random-circuits · Labels. evidence:peer-reviewed role:caution type:journal-article status:verified claim:counterweight

• Abstract. Shows that for sufficiently random parameterised circuits, gradient variance decays exponentially in qubit number, making gradient-based training impractical at scale.
• Gloss. Barren plateau: A flat loss landscape where gradients are exponentially small. 2-design: A circuit ensemble that mimics random unitaries up to second moments.
• Commentary. The key counterweight to optimistic QNN claims. Elucidation: It explains why a training curve from a small simulation should not be extrapolated in a post. Exposition: Random deep circuits concentrate expectation values; derivatives concentrate around zero, so the optimiser sees noise.
• Exegesis. Applies to circuits that form approximate 2-designs. It does not say all QNNs are untrainable; structured or shallow circuits may escape.¹
• Critique. Follow-ups (e.g., B38's summary) refine conditions: cost locality, depth and noise all matter.
• Marginalia.
  • README.md: Pair it with any post that shows a QNN training curve.
• Scholia. B24: The model it constrains. B38: Its follow-ups, summarised.
• Footnote. ¹ The DOI was verified on Crossref after a wrong first draft was rejected.
• In context. A caution for the blog.

B23 · Quantum Computing in the NISQ era and beyond

Caption. The essay that named NISQ and set expectations.

Metadata. John Preskill (2018). Quantum 2: 79. Journal article. DOI: 10.22331/q-2018-08-06-79. Verified 10 Nov 2026.
Descriptors. Gate-model computing · Benchmarking · Tags. nisq perspective expectations · Labels. evidence:peer-reviewed role:caution type:journal-article status:verified claim:counterweight

• Abstract. An influential perspective defining noisy intermediate-scale quantum devices and assessing what they might achieve before full error correction, including optimization and ML, with deliberate caution.
• Gloss. NISQ: Noisy intermediate-scale quantum: tens to hundreds (now thousands) of qubits without full error correction.
• Commentary. The framing for what today's noisy devices can and cannot do. Elucidation: It keeps the blog's claims measured: near-term usefulness is something to test, not to assume. Exposition: Noise limits circuit depth; useful near-term advantage is possible but must be demonstrated, and error correction remains the long-run route.
• Exegesis. Preskill says NISQ devices may be useful and are worth exploring; he explicitly does not promise near-term transformative applications.¹
• Critique. A perspective, not data; the qubit-count range has since shifted.
• Marginalia.
  • README.md: Link from any post that predicts near-term impact.
• Scholia. B26: The field it tempers. B38: The algorithms NISQ invites.
• Footnote. ¹ Based on a keynote Preskill gave in December 2017.
• In context. The tone-setter for the blog.

Code annotations in this repository

Where the four code-annotation elements appear in this repository's own files.^E3

Element	Example in this repository
Code comment	lang_c.ipynb: the Apache license header (# Unless required by applicable law…) is the only code comment; the rest is explained in Markdown cells
Docstring	None; the guide explains code in Markdown cells
Directive	%%bash and %%writefile cell magics in lang_c.ipynb tell Jupyter how to run a cell; #include <stdio.h> in its C example is a preprocessor directive; #@title is a Colab form directive
Decorator	None in this repository

Excursus: What 'quantum deep learning' can mean today

In the literature, 'quantum deep learning' covers three different things. The first is variational QNNs, trainable circuits used like neural-network layers (B24, B38). The second is quantum kernels, circuits that embed data for a classical learner (B39). The third is speedups from quantum linear algebra that assume fast data loading (B26). All three are real research programmes, and none has shown a practical advantage on real-world data. The obstacles are specific and well studied: barren plateaus and concentration (B27) and NISQ noise (B23). TensorFlow Quantum (B19) is how one would experiment with the first two. This repository's only notebook is classical TensorFlow setup, which is an honest starting point for writing about all of this.

Endnotes

1. Verification. Every source was checked on 10 November 2026 (America/New_York). DOIs were resolved at doi.org and matched against Crossref; preprints were checked on arXiv, the patent on Google Patents, and documentation and course pages on the live page.
2. Numbering. B-numbers are stable identifiers shared across David Merwin's annotated bibliographies, so a source keeps the same number in every repository. Gaps in the numbering are expected.
3. Scope. Sources are cited for background and method. None is used to claim a quantum advantage. The code-annotation table describes files as they are; no code was changed.

P.S.

P.S. When a post cites one of these sources, the exegesis line says what that source does not claim. It is a good sentence to quote.