---
title: 'QA-Net: Quantum-Assisted Networks'
url: https://www.emergentmind.com/topics/quantum-assisted-network-qa-net
type: topic
---

# QA-Net: Quantum-Assisted Networks

The designation **Quantum-Assisted Network (QA-Net)** refers, in current arXiv literature, to a class of hybrid quantum-classical architectures in which quantum devices, channels, or circuits are coupled to classical networking, control, or inference components to allocate quantum resources only where they are operationally useful. In its originating communication-oriented formulation, QA-Net augments classical HTTP with quantum metadata and quantum-capable payload handling so that privacy-sensitive traffic can be selectively routed through quantum mechanisms rather than forcing all traffic onto a fully quantum substrate [2505.18282]. Subsequent work uses the same label for structurally related hybrids in satellite-assisted entanglement routing, ultra-lightweight inverse reconstruction for electrical impedance tomography, and entanglement-assisted non-local interferometry [2604.09306], [2507.14031], [2509.09464]. This suggests that QA-Net is presently best understood as a design pattern for quantum-classical augmentation rather than as a single standardized protocol stack.

## 1. Terminological Scope and General Design Pattern

In the communication setting, QA-Net is introduced as a response to a practical asymmetry: several small-scale quantum key distribution networks exist, but large-scale quantum networking still depends on quantum repeaters, quantum channels, quantum memories, and quantum network protocols that are not yet broadly deployable. The proposed remedy is not wholesale replacement of classical infrastructure, but augmentation of existing networks with quantum devices, properties, and phenomena [2505.18282].

Later uses of the term preserve that hybrid character while changing the domain. In SatQNet, the quantum object is a time-varying entanglement-distribution network whose routing policy is learned with a decentralized directed line graph neural network and reinforcement learning [2604.09306]. In QuantEIT, QA-Net denotes a compact model in which parallel 2-qubit circuits provide an implicit nonlinear prior for EIT image reconstruction, followed by a single linear decoding layer [2507.14031]. In non-local optical interferometry, the network consists of spatially separated stations with entangled quantum memories, photon mode erasure, and non-destructive heralding for weak-light phase sensing [2509.09464].

Across these instances, the common structure is a division of labor between quantum and classical subsystems. Quantum resources are reserved for tasks such as confidentiality, entanglement establishment, latent-feature generation, or non-local sensing, while classical components perform metadata binding, routing control, optimization, or downstream decoding. This suggests that the unifying concept is selective quantum augmentation under resource constraints.

## 2. Packet-Level Quantum-Classical Augmentation

The foundational QA-Net communication architecture modifies HTTP at Layer 7 into **Q-HTTP**, allowing a packet or frame to carry both a classical payload and, in the same logical packet, a quantum payload or pointers to it [2505.18282]. The Q-HTTP header retains conventional HTTP fields such as **Host**, **Content-Type**, **Content-Length**, **Cookie**, and **User-Agent**, and adds quantum metadata fields: **Q-Encryption-Mode** with values \(\{\text{“quantum”} \mid \text{“classical”}\}\), **X-Quantum-Channel-ID**, **Q-Protocol** with values \(\{\text{BB84} \mid \text{E91} \mid \text{SDC}\}\), and **Q-Payload-Length**. The payload is split into a classical payload, a quantum payload header containing encoding instructions, and a quantum payload carried via a separate quantum channel.

A crucial clarification is that the quantum payload is not physically embedded in the same electrical bits as the classical packet. It rides in parallel on a quantum channel, while the Q-HTTP header binds the classical and quantum components. This addresses a common misconception: the proposal is not a claim that ordinary packet bits can directly contain qubits, but a logical binding of dual-plane transport within a unified application-layer abstraction [2505.18282].

The write-up models the division of a packet through a privacy-dependent coefficient:
\[
P(\text{packet}) = \alpha \cdot C + (1-\alpha)\cdot Q,
\]
with \(\alpha\in[0,1]\), \(C\) denoting classical bits, and \(Q\) denoting quantum qubits. In discrete form, if \(L\) is the total payload length, the system sends \(\lfloor \alpha L\rfloor\) classical bits and \(\lceil (1-\alpha)L\rceil\) qubits. The same description also states that each outgoing packet is tagged with a privacy label \(\ell\in[0,1]\) and that, for the prototype, a privacy coefficient \(\alpha=\ell\) is defined, while simultaneously specifying that \(\ell=0\) implies \(\alpha\approx1\) and \(\ell=1\) implies \(\alpha\approx0\) [2505.18282]. This suggests that the prototype description blends a privacy-score convention with a classical-fraction convention; the operational intent is nevertheless clear: higher privacy drives a larger quantum share of the payload.

The end-to-end workflow is application-layer driven. Alice composes an e-mail, a Layer 7 privacy classifier assigns \(\ell\in\{0,1\}\), and the Q-HTTP client constructs a packet accordingly. If \(\ell=1\), the packet sets **Q-Encryption-Mode=quantum**; through Layers 6–5, a QKD session negotiates or retrieves a quantum key; as the packet traverses TCP/IP/Ethernet, switches and routers inspect the **Q-Encryption-Mode** flag and route traffic to either a quantum gateway using single-photon qubit transmission on \(\lambda_q\) or a classical gateway on \(\lambda_c\). WDM couplers multiplex \(\lambda_c\) and \(\lambda_q\) on one fiber, and Bob’s endpoint demultiplexes, decrypts, reassembles the fragments according to the Q-HTTP header, and delivers plaintext to Layer 7 [2505.18282].

## 3. Privacy Classification, Security Semantics, and Resource Reduction

The classifier that drives packet splitting in QA-Net is trained on a privacy-labeling task constructed from the Enron Email corpus. The base corpus contains approximately \(587\) K messages; \(20\,000\) messages are sampled, with “private” content injected via regular expressions for SSN, credit-card numbers, passwords, account balances, and transaction IDs. Labels are balanced, with \(10\,000\) non-private messages and \(10\,000\) private messages, and the split is \(80\%\) train/validation and \(20\%\) test. PCA and t-SNE visualizations on TF-IDF embeddings show overlap, indicating no trivial separability [2505.18282].

Four models are implemented: Logistic Regression, CNN, LSTM, and BiLSTM. For the neural models, the common objective is binary cross-entropy,
\[
L(\theta) = -\frac{1}{N}\sum_{i=1}^{N}\left[y_i\log p_i + (1-y_i)\log(1-p_i)\right].
\]
The Logistic Regression baseline uses TF-IDF vectors with `max_features=5000` over unigrams, bigrams, and trigrams, and outputs \(\sigma(w\cdot x+b)\). The CNN uses an embedding layer with `vocab_size=5000` and `dim=128`, 1D convolutions with filter sizes \(\{3,5,7\}\) and `num_filters=100` each, followed by MaxPooling, `Dropout(0.4)`, `Dense(64)`, and Sigmoid. The LSTM uses an embedding layer of the same vocabulary size and dimension, `LSTM(units=128, return_sequences=False)`, BatchNorm, `Dropout(0.5)`, `Dense(1)`, and Sigmoid. The BiLSTM uses an embedding layer, `Bidirectional(LSTM(units=64))`, `Dropout(0.5)`, `Dense(1)`, and Sigmoid [2505.18282].

The reported test results are: Logistic Regression with **90.0%** accuracy, precision\(_1=0.95\), recall\(_1=0.84\), and \(F1_1=0.91\); CNN with **94.6%** accuracy, precision\(_1=0.97\), recall\(_1=0.93\), and \(F1_1=0.95\); LSTM with **95.4%** accuracy, precision\(_1=0.92\), recall\(_1=0.99\), and \(F1_1=0.96\); and BiLSTM with **94.2%** accuracy, precision\(_1=0.95\), recall\(_1=0.93\), and \(F1_1=0.94\) [2505.18282]. The LSTM therefore yields the highest reported accuracy and \(F1\) in this prototype.

The resource argument is central to the QA-Net proposal. In a purely quantum network, every packet uses QKD-derived keys and single photons, so the resource cost scales with \(N_{\text{total}}\). In QA-Net, only
\[
N_{\text{private}}=\sum \ell_i \cdot 1
\]
packets use quantum, and the savings are
\[
S=1-\left(N_{\text{private}}/N_{\text{total}}\right).
\]
Sampling \(500\)–\(800\) e-mails five times yields approximately \(50\)–\(55\%\) non-private traffic; accordingly, only about \(45\)–\(50\%\) of packets require quantum encoding, corresponding to approximately **50% reduction in quantum resource consumption**—specifically qubits and single-photon pulses—relative to a fully quantum network [2505.18282].

The security semantics are layered rather than uniform. Private packets retain provable quantum security through protocols such as BB84 or three-stage QSDC, whereas non-private packets continue to use classical encryption, specifically TLS 1.3, against active adversaries. The stated conclusion is that overall confidentiality is at least classical and approximately quantum for sensitive traffic [2505.18282]. A second classifier pass at the receiver is proposed as an integrity check; unexpected \(\ell=1\) packets can trigger key-revocation or replay defenses.

## 4. Edge-Centric QA-Net for Satellite-Assisted Entanglement Routing

In SatQNet, QA-Net appears not as packet augmentation but as a decentralized control architecture for a satellite-assisted quantum network modeled as a time-varying graph
\[
G(t)=(V,E(t)),
\]
where \(V\) contains terrestrial stations and satellites, and \(E(t)\subseteq V\times V\) contains physical edges capable of Bell-pair generation [2604.09306]. Each edge \(e=(u,v)\) carries a single-photon success probability \(P_{\text{Link}}(u,v,t)\), a fidelity \(F_e(t)\) that decays under memory decoherence as
\[
F_e(t+\Delta t)=(F_e(t)-F_B)e^{-(\Delta t)/T_2}+F_B,
\]
and a Bell-pair generation rate
\[
\lambda_e(t)=r_{\text{attempt}}\cdot P_{\text{Link}}(u,v,t).
\]

End-to-end entanglement along a path \(p=[v_0\to v_1\to \dots \to v_k]\) is constructed through successive entanglement swaps. The path fidelity is written as a nested swap composition and, in the ideal depolarizing model, is approximated by
\[
F_{\text{path}}\approx \prod_{e\in p}F_e.
\]
This product form also defines the terminal reward in the reinforcement-learning formulation: if the agent reaches the destination and performs swaps along \([e_0,\dots,e_{T-1}]\), then
\[
R_T = F_{\text{path}}=\prod_{i=0}^{T-1}F_{e_i},
\]
and otherwise \(R_T=0\) [2604.09306].

The distinctive architectural device is the **directed line graph** \(L(G)=(E_{\text{dir}},A_{\text{dir}})\), whose nodes are directed edges \((u\to v)\) of the underlying undirected graph, and where a directed line-graph edge exists from \(e=(u\to v)\) to \(e'=(v\to w)\) when \(u\neq w\). A separate GNN runs on \(L(G)\). Each directed edge carries an input feature vector \(x_e\) encoding \(\{F_e,\lambda_e,\text{swapProb},\dots\}\) and an initial embedding \(h_e^{(0)}=\text{Enc}(x_e)\). After \(K\) rounds of message passing with learnable \(\phi_m\) and \(\phi_u\), the edge embedding \(h_e^{(K)}\) incorporates neighboring-link information up to \(K\) hops away. Candidate outgoing edges at repeater \(v\) are then scored by
\[
Q(s,h_e^{(K)})=\text{MLP}_Q([\,s_{\text{local}}\parallel h_e^{(K)}]),
\]
and actions are selected \(\epsilon\)-greedily [2604.09306].

A key feature is decentralized runtime execution. Each repeater or satellite runs the same GNN and Q-policy, there is no global controller, and each node exchanges current embeddings only with immediate neighbors once per GNN round. This directly addresses a standard difficulty in satellite-assisted quantum networking: global topology information becomes stale because topology is highly dynamic and the classical control plane is delayed. The runtime cost is reported as \(O(K\cdot d_{\text{avg}})\) GNN steps per routing request, with each exchange consisting of a small vector of approximately \(32\) floats [2604.09306].

The evaluation metrics are entanglement distribution rate (EDR), average end-to-end fidelity \(\langle F_{\text{path}}\rangle\), and routing latency. On random graphs with \(|V|=100\)–\(300\) and on a real-world European backbone with approximately \(500\) PoPs plus a Starlink orbital constellation, SatQNet reports **20–50% higher EDR** than heuristics such as MGER, LBER, and AER and global-view methods such as Q-PATH and Q-LEAP once \(|V|>30\); **\(\sim0.02\)–\(0.05\)** higher \(\langle F_{\text{path}}\rangle\) under high dynamics; and success-rate gains on Europe, routing at least one high-fidelity entanglement to **97% of cities** versus **65%** for Q-PATH with \(p<10^{-3}\). It also outperforms the strongest local-learning baseline, RELiQ\(^*\), by **10–30% in EDR** without collecting or storing the global network graph [2604.09306].

## 5. QA-Net as a Quantum-Assisted Inference Architecture

In QuantEIT, QA-Net denotes an ultra-lightweight reconstruction model for chest electrical impedance tomography in which a bank of parallel 2-qubit quantum circuits replaces a deep multilayer network as the source of nonlinear expressivity [2507.14031]. The architecture employs \(n_c\) independent circuits in parallel, with \(n_c=2\), each acting on \(n_q=2\) qubits initialized in the tensor-product ground state \(\lvert 0\rangle^{\otimes n_q}\). Each circuit applies parameterized single-qubit \(R_Y(\alpha)\) rotations to each qubit, followed by a chain of CNOT gates that entangle the two qubits. Measurement in the Pauli-\(Z\) basis produces real expectation values \(f_k^{(i)}\), and the outputs from all circuits are concatenated into a latent vector
\[
\mathbf F=\bigl[f_0^{(1)},f_1^{(1)},\dots,f_0^{(n_c)},f_1^{(n_c)}\bigr]^\top\in\mathbb R^{n_q n_c}.
\]

The circuit-level parameterization is explicit. The single-qubit rotation is
\[
R_Y(\alpha)=\exp\!\Bigl(-i\,\frac{\alpha}{2}Y\Bigr),
\]
the entangling gate is CNOT, and the measurement observable for circuit \(i\) and qubit \(k\) is
\[
f_k^{(i)}(\pmb{\phi}^{(i)})=
\big\langle\psi^{(i)}(\pmb{\phi}^{(i)})\big|\,Z_k\,\big|\psi^{(i)}(\pmb{\phi}^{(i)})\big\rangle.
\]
There is no classical input beyond the fixed reference state; all nonlinearity is induced by the gate sequence. The resulting mapping \(\Psi_{\text{quantum}}\!:\pmb{\phi}\mapsto \mathbf F\) is described as an implicit regularizer or prior in the inverse problem [2507.14031].

The classical part of the model is deliberately minimal:
\[
\Delta\pmb{\sigma} = \mathrm{Sigmoid}(\pmb W\,\mathbf F+\pmb b),
\]
with \(\pmb W\in\mathbb R^{n\times 4}\) and \(\pmb b\in\mathbb R^n\). The Sigmoid enforces \(\Delta\sigma_j\in[0,1]\). The total learnable parameters are therefore four quantum parameters and \(4n+n\) classical parameters. Reconstruction is unsupervised and training-data-free: with the linearized inverse model \(\Delta\mathbf v=J\,\Delta\pmb{\sigma}\), QA-Net optimizes
\[
\Delta\pmb{\sigma}^* = \arg\min_{\pmb\theta}
\left\lVert \Delta\mathbf v - J\,\Psi(\pmb\theta)\right\rVert_2^2
+\pmb\lambda^\top \pmb{\mathcal R}(\Psi(\pmb\theta)),
\]
where \(\pmb\theta=(\pmb\phi,\pmb W,\pmb b)\), \(\pmb\lambda\) weights Laplacian smoothing, Total Variation, and \(\ell_1\) regularization, gradients are computed with the parameter-shift rule in PennyLane, and Adam performs the updates [2507.14031]. No pre-training or external dataset is used; measured voltages \(\Delta\mathbf v\) are the only data.

Quantitatively, the 2D simulation with 16 electrodes on a \(64\times64\) grid reports **CC = 0.87**, **PSNR = 18.56 dB**, **MSSIM = 0.93**, and **ERR = 0.54**, compared with R-SIP at approximately **CC \(\approx 0.80\)**, **PSNR \(\approx 16\) dB**, and **ERR \(\approx 0.75\)**, as well as Noser [2507.14031]. In 3D simulation on a \(32\times32\times40\) voxel grid, the reported values are **CC = 0.80**, **PSNR = 29.84 dB**, **MSSIM = 0.81**, and **ERR = 0.56**, outperforming baselines by at least **10% on most metrics**. Under additive Gaussian noise from **SNR = 10 to 60 dB**, all metrics degrade by less than **5%** for **SNR \(\ge 30\) dB**, while at **10–20 dB** performance drops by at most **10%**. Real-world phantom and clinical data are reported to localize unilateral ventilation defects and perfusion dynamics in 3D [2507.14031].

The computational argument is equally important. Total learnable parameters are **20,484** for QuantEIT 2D versus **8.85 M** for R-SIP, and **204,804** for QuantEIT 3D versus **82.62 M** for R-SIP. Classical FLOPs are described as negligible for QuantEIT because only a single linear layer is used, in contrast to **17.7 M** FLOPs for R-SIP in 2D and **165 M** in 3D. CPU-simulated inference on Intel i9 takes approximately **18.1 s** in 2D and **19.0 s** in 3D for QuantEIT, versus **5.0 s** and **32.3 s** for R-SIP, respectively. The paper characterizes this as achieving comparable or superior reconstruction using only **0.2% of the parameters** of leading classical unsupervised methods [2507.14031].

## 6. Entanglement-Assisted QA-Net for Non-Local Optical Interferometry

A distinct physical realization of QA-Net appears in non-local optical interferometry with two separate stations, \(L\) and \(R\), each in a dilution refrigerator at approximately \(100\) mK and each containing a single silicon-vacancy center in a diamond nano-cavity [2509.09464]. The SiV provides a fast electron-spin communication qubit coupled to a long-lived \(^{29}\)Si nuclear-spin memory qubit. The station outputs form the arms of a phase-stabilized Mach–Zehnder interferometer. To demonstrate a long baseline, up to **1.55 km** of standard telecom fiber is inserted in each arm, while the stations remain synchronized through classical timing signals on auxiliary fibers.

The entanglement protocol has two stages. First, electron–electron entanglement is produced by a single-mode spin–photon gate (SMSPG): both electrons are prepared in \(|+_e\rangle\), a weak coherent pulse is split over the two arms, and detection of exactly one photon after recombination projects the electrons, under interferometer phase lock to \(\pi\), into the Bell state
\[
|\Psi^-\rangle_{e_L,e_R}=\frac{1}{\sqrt2}(|\uparrow\downarrow\rangle-|\downarrow\uparrow\rangle).
\]
Second, the electron state is transferred to the nuclear memories via SMPHONE, with a mid-circuit electron readout post-selecting successful gates; the resulting Bell state between nuclear spins is reported with fidelity \(F\approx0.73\) [2509.09464].

After memory entanglement, a weak signal pulse with average photon number \(\mu_{\rm sig}\ll1\) is sent to both stations in the approximate joint state
\[
|\psi_{\rm sig}\rangle\approx|0_L0_R\rangle + \sqrt{\mu_{\rm sig}/2}\bigl(|0_L1_R\rangle + e^{i\phi}|1_L0_R\rangle\bigr),
\]
where \(\phi\) is the differential phase. Which-path information is erased by mixing each reflected signal mode with a weak coherent local oscillator on a 50/50 beamsplitter and conditioning on the click criterion \((i\neq i')\wedge(j\neq j')\). A Pauli-\(Z\) correction is applied to the nuclear memory when the click-sign patterns differ between stations. This realizes an effective \(X\)-basis measurement on each photonic mode while preserving only the encoded phase \(\phi\) in the nuclear state [2509.09464].

Photon presence is then heralded non-destructively. Local C\(_n\)NOT\(_e\) gates couple nuclear and electron spins, the electrons are rotated by \(\pi/2\) around \(Y\), and only runs with even two-electron parity are retained. Because vacuum inputs lead to odd parity due to the initial nuclear Bell state, this step heralds a photon without revealing which station received it. The heralding probability is
\[
P_h=\eta_{\rm erasure}\,\eta_{\rm herald}\,P(n_{\rm photon}\ge1)\approx
\eta_{\rm erasure}\,\eta_{\rm herald}\,\mu_{\rm sig}
\]
for \(\mu_{\rm sig}\ll1\), with \(\eta_{\rm herald}\approx \frac12\) [2509.09464].

The figure of merit is Fisher information \(\mathcal F_I\). With heralding,
\[
\mathcal F_I\approx p_{\rm succ}V^2 \propto
\eta_{\rm erasure}\,\eta_{\rm herald}\,\overline V^2\,\mu_{\rm sig},
\]
whereas without heralding,
\[
\mathcal F_I\propto \mu_{\rm sig}^2.
\]
Thus, in the weak-signal regime, the entanglement-assisted architecture converts quadratic into linear scaling in \(\mu_{\rm sig}\), and the signal-to-noise ratio scales as \(\sqrt{\mu_{\rm sig}}\) rather than \(\mu_{\rm sig}\) [2509.09464].

Experimentally, the electron–electron Bell fidelity is approximately **0.83** at **50 m** and drops to about **0.63** at **1.55 km**; the nuclear–nuclear Bell fidelity is approximately **0.73** at **50 m** and approximately **0.63** at **1.55 km**. The entanglement rate reaches **13 Hz** for electrons and **1.9 Hz** at \(F\ge0.79\), while the nuclear rate is approximately **0.28–2.8 Hz**. The heralding probability \(P_h\approx \eta\,\mu_{\rm sig}\) is observed to be linear in \(\mu_{\rm sig}\), nuclear parity visibility without heralding is approximately **0.03** at \(\mu=0.25\) and rises to approximately **0.31** at \(\mu=2\), while with heralding it is approximately **0.09** averaged over \(\mu\). Over **1.55 km**, a differential-phase oscillation visibility of **0.11(4)** is observed, establishing non-local interferometry at five times the previous record baseline [2509.09464]. Reported error sources include fiber-link phase noise, approximately **50%** cavity-photon coupling losses, **50%** herald efficiency from amplitude-based gates, and approximately **30%** mis-heralding from imperfect Bell states.

## 7. Common Themes, Misconceptions, and Prospective Directions

Several recurrent themes unify these otherwise heterogeneous QA-Net realizations. First, QA-Net is consistently **hybrid** rather than purely quantum. Q-HTTP binds classical headers to a quantum side channel rather than replacing classical networking; SatQNet relies on classical GNN inference and message exchange to improve quantum entanglement routing; QuantEIT uses a quantum latent generator with a classical linear decoder and classical optimization; and non-local interferometry depends on classical synchronization, active phase stabilization, and conditioning logic in addition to remote entanglement [2505.18282], [2604.09306], [2507.14031], [2509.09464].

Second, QA-Net is consistently **resource-selective**. In the networking prototype, only privacy-sensitive packets consume qubits and single-photon pulses. In SatQNet, local message passing avoids the overhead of collecting the global network graph. In QuantEIT, two 2-qubit circuits are used to replace much larger deep models. In interferometry, heralding and mode erasure are used to recover useful phase information in the low-photon regime. This suggests that QA-Net is less a doctrine of maximal quantization than a doctrine of targeted quantum deployment.

Third, the literature corrects several natural misconceptions. Q-HTTP does not mean that qubits are literally inserted into the same classical packet bits; the quantum payload rides on a separate channel logically bound by metadata [2505.18282]. “Training-data-free” in QuantEIT does not mean parameter-free; the method still optimizes \((\pmb\phi,\pmb W,\pmb b)\) with Adam using parameter-shift gradients [2507.14031]. “Decentralized” in SatQNet does not imply isolation; nodes still exchange edge embeddings with immediate neighbors every GNN round [2604.09306]. In interferometry, heralding is non-destructive with respect to which-path information, not an elimination of loss or gate imperfections [2509.09464].

Prospective directions follow directly from the cited works. The communication-oriented QA-Net is motivated by currently feasible quantum technologies and by the longer-term development of quantum repeaters, channels, memories, and network protocols [2505.18282]. SatQNet points toward scalable entanglement routing in highly dynamic terrestrial-satellite constellations without global state collection [2604.09306]. QuantEIT identifies GPU-accelerated quantum simulation such as cuQuantum and execution on near-term quantum hardware as routes to additional speedups [2507.14031]. The interferometric QA-Net explicitly proposes time-bin multiplexing, frequency multiplexing, repeaters, multiplexed entanglement generation, low-loss fiber packaging, wavelength-division multiplexers, and larger memory registers as scaling mechanisms [2509.09464]. Taken together, these directions indicate that QA-Net research is evolving toward architectures in which quantum resources are not monolithically deployed, but are inserted at the points where entanglement, measurement, or nonlinear representation can most effectively alter the system-level operating point.

Source: https://www.emergentmind.com/topics/quantum-assisted-network-qa-net