Supervised In-Situ Learning Algorithm

• Supervised in-situ learning algorithm is a method that performs data acquisition and model updates directly within the operational environment using native system signals.
• It employs domain-specific supervision techniques—such as cross-entropy loss, teacher signals, and adaptive update rules—across diverse applications like quantum systems and additive manufacturing.
• Empirical studies validate its practical impact, showing enhanced system efficiency, improved performance metrics, and robust adaptability in complex real-world settings.

A supervised in-situ learning algorithm is a learning procedure executed directly on the target system, device, or deployment environment, with supervision supplied by measured outcomes, teacher signals, known targets, or task losses produced in that environment rather than by offline training alone. Across the cited literature, the expression is used for control-free Hamiltonian identification from native quantum evolution, bitrate selection trained on live deployment telemetry, laser additive-manufacturing monitoring with supervised cross-modality transfer, online spiking and memristive learning, ocean-state reconstruction from sparse observations, and in situ backpropagation in topological mechanical networks (Zhou et al., 17 Jun 2026, Yan et al., 2019, Xie et al., 2024, Cutolo et al., 2022, Li et al., 10 Mar 2025).

1. Terminological scope

In the literature, “in situ” consistently denotes operation on the native substrate rather than in a detached surrogate. In quantum Hamiltonian learning, the protocol acts directly on $U(t)=e^{-iHt}$, does not pause or interleave coherent control gates, uses no ancillas, and restricts each shot to preparing a product Pauli eigenstate, letting the system evolve for a randomized duration, and measuring in a product Pauli basis (Zhou et al., 17 Jun 2026). In adaptive bitrate control, “in situ” means training the Transmission Time Predictor directly on telemetry collected from the live Puffer deployment environment, explicitly contrasting that regime with training in emulation, simulation, or on traces collected elsewhere (Yan et al., 2019). In ocean reconstruction, sparse in-situ measurements enter the Optimal Interpolation update through the observation operator $H$ in $y=Hx+\epsilon$ (Cutolo et al., 2022).

The adjective “supervised” is used in several distinct but technically precise senses. In Puffer, it refers to direct supervision by realized chunk transmission times and cross-entropy training of a predictor over 21 transmission-time bins (Yan et al., 2019). In laser additive manufacturing, it refers to defect labels driving classification, semantic alignment, or cross-modality mapping objectives (Xie et al., 2024). In spiking and memristive systems, supervision is implemented as an explicit teacher signal: output-layer nudging in the STDP-based SNN and a label-dependent teacher bias current in the memristive 1T1M crossbar (Hu et al., 2022, Prajapati et al., 28 Jul 2025). In structured quantum learning, supervision is tied to prior knowledge of the operator basis and device structure, with $H=\sum_j \theta_j H_j$ estimated inside a known model class (Liu et al., 9 Oct 2025). In the ansatz-free quantum case, the paper defines the method as “supervised-by-data”: support and coefficients are inferred from labeled outcomes collected under randomized, band-limited time sampling (Zhou et al., 17 Jun 2026).

Taken together, these usages suggest that “supervised in-situ learning algorithm” is not a single canonical update rule but an umbrella designation for methods that combine deployment-time or substrate-native data acquisition with a supervised estimation mechanism.

2. Recurring operational structure

A recurring operational pattern is direct interaction with the physical or deployed system, followed by immediate statistical or algorithmic updates. In the ansatz-free Hamiltonian setting, Stage I uses same-Pauli experiments and a displacement sieve to identify heavy support projections, while Stage II uses cross-basis Pauli experiments, signed parity signals, and visibility correction to estimate coefficients on the candidate set $U_{\text{proj}}$ (Zhou et al., 17 Jun 2026). In Puffer, the learning loop collects “video_sent,” “video_acked,” and “client_buffer” telemetry, retrains daily on the previous 14 days of live data, and serves the resulting predictor online inside a per-session daemon (Yan et al., 2019). In topological mechanical neural networks, each training iteration consists of a forward experiment, an adjoint experiment, and a local parameter update based on forward and adjoint measurements (Li et al., 10 Mar 2025).

The supervision channel is equally system-specific. In laser additive manufacturing, paired audio and visual observations are synchronized, encoded, and trained with binary cross-entropy and cross-modality losses so that one modality can be removed at prediction time (Xie et al., 2024). In object navigation, the paper itself is self-supervised, but it explicitly formulates a supervised in-situ variant in which sparse segmentation labels are combined with LocCon via $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$ (Min et al., 2022). In Boundary Forests, supervision enters through a boundary-addition criterion applied immediately to each arriving labeled example, so that training is naturally online (Mathy et al., 2015).

A plausible implication is that the defining architectural property is not any particular optimizer, but the co-location of sensing, supervision, and update within the operational environment.

3. Objective functions, estimators, and update rules

The mathematical form of supervision varies sharply by domain. In Puffer, the learned component is the Transmission Time Predictor, trained with stochastic gradient descent on the cross-entropy loss

$$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$

while decision-making is handled by stochastic MPC with

$$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$

and deployment parameters $\lambda=1$, $\mu=100$ (Yan et al., 2019).

In supervised cross-modality transfer for additive manufacturing, semantic alignment combines binary classification with a cross-domain alignment term,

$H$0

with $H$1 and margin $H$2, while fully supervised and semi-supervised mapping variants use mean-squared error to regress features across modalities before target-side classification (Xie et al., 2024). In the ISH-image setting, the encoder is trained unsupervised by a denoising reconstruction objective and the supervised stage is an $H$3-regularized logistic regression,

$H$4

applied independently to 2,081 GO terms (Cohen et al., 2019).

In the quantum ansatz-free setting, supervision is not an empirical risk over labels in the conventional machine-learning sense, but a measurement-driven derivative-estimation pipeline. Same-Pauli auto-correlation satisfies

$H$5

and kernel-weighted samples are constructed so that $H$6. In Stage II, odd-kernel sampling and visibility correction give unbiased coefficient estimators with $H$7 (Zhou et al., 17 Jun 2026). In topological mechanical neural networks, the supervised objective is binary cross-entropy on the output ratio $H$8, and the in situ adjoint rule is local: $H$9 The paper explicitly presents this as a two-step, local-rule method that updates TMNNs using only local information (Li et al., 10 Mar 2025).

Teacher-signal formulations are prominent in spiking and memristive hardware. The STDP-based supervised SNN adds the nudging term $y=Hx+\epsilon$0 to output neurons during the learning phase and updates weights by

$y=Hx+\epsilon$1

inside a controlled time window (Hu et al., 2022). The memristive 1T1M system injects a label-dependent bias current $y=Hx+\epsilon$2 into the correct-class neuron, then applies $y=Hx+\epsilon$3 when $y=Hx+\epsilon$4 and $y=Hx+\epsilon$5 when $y=Hx+\epsilon$6, thereby implementing supervised in-situ STDP directly in hardware (Prajapati et al., 28 Jul 2025).

4. Substrates and system architectures

The range of substrates is unusually broad. The quantum Hamiltonian learner is explicitly control-free, ancilla-free, and restricted to Pauli product state preparation and measurement; it does not assume locality, although locality enables stronger SPAM-robustness guarantees after calibration (Zhou et al., 17 Jun 2026). The parallelized structured Hamiltonian learner targets fully connected Ising-type models with one locally driven qubit, exploits $y=Hx+\epsilon$7 invariant subspaces, and estimates all couplings without decoupling non-learnable interactions during the same experiment (Liu et al., 9 Oct 2025).

In networking, the Puffer deployment uses a “dumb” HTML5 <video> + MediaSource Extensions browser client, a server-side ABR controller, and a learned Transmission Time Predictor implemented as a fully connected DNN with two hidden layers of 64 neurons each. Training is done in PyTorch, while inference runs in C++ on the serving daemons (Yan et al., 2019). In laser additive manufacturing, the semantic-alignment model uses a shared convolutional encoder $y=Hx+\epsilon$8 in a Siamese architecture and a shared classifier $y=Hx+\epsilon$9, while fully supervised and semi-supervised mapping methods decompose training into phases with CNNs, FC networks, and autoencoders (Xie et al., 2024).

In representation learning for neural ISH images, the substrate is a deep convolutional denoising autoencoder that maps downsampled $H=\sum_j \theta_j H_j$0 images to a $H=\sum_j \theta_j H_j$1 bottleneck, yielding a 1,800-dimensional representation for downstream one-vs-rest logistic regression (Cohen et al., 2019). In ocean-state reconstruction, CLOINet integrates OINet, CluNet, and RefiNet around the Optimal Interpolation update

$H=\sum_j \theta_j H_j$2

with $H=\sum_j \theta_j H_j$3 parameterized in cluster space to encode non-local correlations (Cutolo et al., 2022). In physical learning hardware, the substrate may itself embody both state and update. The OxRAM Bayesian learner uses a fabricated array of 16,384 devices whose cycle-to-cycle conductance variability implements the proposal distribution of a Metropolis–Hastings sampler (Dalgaty et al., 2020). The topological mechanical classifier is a mass–spring metamaterial with QSHE-inspired pseudospin edge modes, while the memristive SNN uses a single-layer 1T1M crossbar with analog winner-take-all, lateral inhibition, and refractory circuitry implemented without external microcontrollers or ancillary control hardware (Li et al., 10 Mar 2025, Prajapati et al., 28 Jul 2025).

5. Guarantees, scaling laws, and empirical performance

The strongest formal guarantee in the set appears in ansatz-free quantum Hamiltonian learning. The proposed algorithm learns an ansatz-free Hamiltonian $H=\sum_j \theta_j H_j$4 with $H=\sum_j \theta_j H_j$5 in total evolution time

$H=\sum_j \theta_j H_j$6

and the paper proves a lower bound of

$H=\sum_j \theta_j H_j$7

for any control-free protocol. It also states that the characteristic probe time resolution depends only on $H=\sum_j \theta_j H_j$8 instead of $H=\sum_j \theta_j H_j$9, and that average probe time per shot is $U_{\text{proj}}$0 (Zhou et al., 17 Jun 2026). The structured parallelized variant reports CRLB-saturating precision, $U_{\text{proj}}$1 rounds for $U_{\text{proj}}$2, and robustness to depolarizing noise and time-dependent coherent errors in the targeted model class (Liu et al., 9 Oct 2025).

Several application papers emphasize deployment-scale empirical validation. Puffer reports 14.2 years of video streamed to 56,000 users, with a primary randomized experiment spanning 458,801 streams from 44,907 client IPs. In that study, Fugu achieved stall ratio $U_{\text{proj}}$3, mean SSIM $U_{\text{proj}}$4 dB, and SSIM variation $U_{\text{proj}}$5 dB, while RobustMPC-HM achieved the lowest stall ratio $U_{\text{proj}}$6 at lower mean SSIM $U_{\text{proj}}$7 dB (Yan et al., 2019). The paper also reports that the emulation-trained version of Fugu performed “catastrophically” when deployed on Puffer, whereas the in-situ-trained version robustly improved QoE (Yan et al., 2019).

In laser additive manufacturing, semantic alignment CMKT achieved Accuracy $U_{\text{proj}}$8 $U_{\text{proj}}$9, AUC-ROC $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$0, and Balanced Acc $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$1, while removing the audio modality at prediction; feature-level multimodal fusion reached Accuracy $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$2 and AUC-ROC $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$3 (Xie et al., 2024). In neural ISH image analysis, CDAE-based representations improved the previous state of the art from an average AUC of $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$4 to $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$5 on the top-15 GO categories, and the average AUC across all 2,081 categories was $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$6 (Cohen et al., 2019). In ocean reconstruction, CLOINet reduced RMSE by up to $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$7 relative to OI baselines and resolved scales about $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$8 smaller, described equivalently as about $$L=\lambda_{\mathrm{LocCon}}L_{\mathrm{LocCon}}+\lambda_{\mathrm{sup}}L_{\mathrm{sup\_seg}}+\lambda_{\mathrm{reg}}L_{\mathrm{reg}}$$9 smaller wavelengths at the $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$0 threshold (Cutolo et al., 2022).

Hardware-oriented studies also report concrete performance. The OxRAM Bayesian sampler achieved $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$1 test accuracy in one breast-cancer run and a median test accuracy of $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$2 across 100 train/test iterations, outperforming a software neural-network baseline with the same number of “memory elements” (Dalgaty et al., 2020). The memristive supervised in-situ STDP architecture achieved perfect pattern recognition on $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$3 binary images and classification accuracies of $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$4 on Iris and $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$5 on BCW, while maintaining an average recognition rate of $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$6 under $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$7 input noise (Prajapati et al., 28 Jul 2025). The STDP-based supervised SNN reached $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$8 on MNIST with a $$L(\theta) = - \sum_{i\in D} \sum_{k=1}^{21} y_{i,k}\log p_{\theta,k}(x_i),$$9-$$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$0-$$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$1 architecture (Hu et al., 2022). Boundary Forests report empirical training-time scaling as $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$2 and testing as $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$3, with online updates performed one example at a time (Mathy et al., 2015).

6. Limitations, controversies, and open directions

A recurrent issue is that “in situ” does not eliminate distributional complexity; it only changes where learning occurs. Puffer explicitly shows that Internet paths and user behavior produce skewed, heavy-tailed distributions, that rebuffering is rare but impactful, and that even with roughly $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$4 years of data per scheme, a $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$5 improvement in rebuffering ratio would be statistically indistinguishable at $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$6 confidence (Yan et al., 2019). The real-robot navigation study argues that supervised training with 3D mesh annotations causes models to learn simulation artifacts, and positions self-supervised in-situ finetuning as a corrective; its supervised extension is presented only as a hybrid objective when labels are available on-robot (Min et al., 2022).

Several methods have explicit structural limits. The ansatz-free Hamiltonian learner assumes a known norm bound $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$7, faces a candidate-set bound $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$8, and in the grid setting requires time resolution $$QoE(K_i^s,K_{i-1}) = Q(K_i^s) - \lambda |Q(K_i^s)-Q(K_{i-1})| - \mu \max\{T(K_i^s)-B_i,0\}$$9 for uniform support identifiability in ansatz-free families (Zhou et al., 17 Jun 2026). CLOINet relies on OGCM outputs for supervision, uses the synoptic assumption, and states that uncertainty quantification is not explicitly estimated (Cutolo et al., 2022). Boundary Forests are online and fast, but the baseline formulation does not include forgetting, pruning, or rebalancing, which becomes relevant under concept drift or strict memory budgets (Mathy et al., 2015).

Cross-modality and hardware methods introduce their own assumptions. CMKT assumes overlapping task-relevant information across modalities, and supervised mapping requires paired, synchronized samples across modalities (Xie et al., 2024). The memristive 1T1M SNN is robust to stuck-at-conductance faults and moderate $\lambda=1$0 variations, but the paper reports a stronger sensitivity to threshold variations, with Iris accuracy falling to $\lambda=1$1 at $\lambda=1$2 threshold variation (Prajapati et al., 28 Jul 2025). The topological mechanical classifier preserves a nontrivial spin Bott index under moderate disorder, yet its current formulation remains linear and relies on bounded parameter changes to avoid Anderson localization (Li et al., 10 Mar 2025).

These limitations clarify a common misconception. Supervised in-situ learning is not a guarantee of universal robustness or a synonym for a single training doctrine. In the literature, it names a set of strategies for aligning supervision with native dynamics, deployment telemetry, or physical state variables. The central technical question is therefore not whether learning is “in situ” in the abstract, but how the supervision channel, update rule, and substrate constraints are matched to one another in a given system.