Precision Neural Networks Overview

• Precision Neural Networks (PNNs) are models defined by deliberately structured parameterizations, such as physical controls, inverse-covariance operators, or quantization modes.
• They encompass diverse architectures including photonic networks with analog weights, graph neural networks using precision matrices, and ReLU models with geometric constraints.
• By embedding device physics and hardware-aware training into the optimization process, PNNs improve accuracy, robustness, and energy efficiency across various application domains.

Searching arXiv for the cited papers to ground the article in current literature. arxiv_search(query="(Zhang et al., 2024) Online training and pruning of multi-wavelength photonic neural networks", max_results=5) arxiv_search(query="(Xu et al., 2024) Perfecting Imperfect Physical Neural Networks with Transferable Robustness using Sharpness-Aware Training", max_results=5) arxiv_search(query="(Cavallo et al., 18 Sep 2025) Precision Neural Networks Joint Graph And Relational Learning", max_results=5) arxiv_search(query="(Feizi et al., 2017) Porcupine Neural Networks Almost All Local Optima are Global", max_results=5) arxiv_search(query="(Guerra et al., 2020) Switchable Precision Neural Networks", max_results=5) Precision Neural Networks (PNNs) do not denote a single standardized architecture in the recent literature. The label is used for physical neural networks of the form $y=F(x;\Theta)$, for graph neural networks whose shift operator is the precision matrix $\Theta=\mathbf{C}^{-1}$, for two-layer ReLU networks whose weight vectors are constrained to a finite set of lines, for shared quantized models with switchable bitwidths, for polynomial networks with multiplicative interactions, and for proximal or Parseval-proximal networks obtained from operator-splitting algorithms (Xu et al., 2024, Cavallo et al., 18 Sep 2025, Feizi et al., 2017, Guerra et al., 2020, Choraria et al., 2022, Hasannasab et al., 2019). A common thread is the use of a deliberately structured parameterization—physical controls, inverse-covariance operators, geometric constraints, proximal maps, or quantization modes—to shape accuracy, robustness, and efficiency.

1. Terminological scope and canonical formulations

In physical-neural-network work, a PNN replaces digital linear-algebra hardware by a physical system parameterized by tunable controls $\Theta$, with training posed as

$$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$

Here $\Theta$ may consist of heater currents, phases, voltages, or misalignment parameters, and the central problem is that the mapping from desired weights to physical controls is distorted by fabrication variance, drift, and limited control precision (Xu et al., 2024).

In precision-matrix graph learning, PNN means a graph convolutional architecture whose graph shift operator is the precision matrix $\mathbf{\Theta}=\mathbf{C}^{-1}$. A layer is

$$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$

so filtering is performed directly on inverse-covariance structure rather than on covariance itself (Cavallo et al., 18 Sep 2025).

In "Porcupine Neural Networks," the model is a two-layer ReLU network

$h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$

with the additional constraint that each $w_i$ lies on one of a finite set of lines. If $u_{g(j)}$ is the unit direction of the assigned line, then

$\Theta=\mathbf{C}^{-1}$0

This geometric restriction is used to analyze the nonconvex landscape and to show that most local optima are global (Feizi et al., 2017).

In switchable-precision work, PNN refers to a single shared quantized model capable of operating at multiple bitwidths. The forward pass uses quantized activations and weights,

$\Theta=\mathbf{C}^{-1}$1

and the same backbone can be switched at runtime across several $\Theta=\mathbf{C}^{-1}$2 modes according to memory, latency, power, or accuracy constraints (Guerra et al., 2020).

Other works use PNN for polynomial neural networks, where multiplicative interactions create explicit polynomial structure, and for proximal neural networks, where layers are proximity operators or unfolded proximal steps (Choraria et al., 2022, Hasannasab et al., 2019, Le et al., 2023). The term is therefore intrinsically polysemous.

2. Physical and photonic realizations

Microring-resonator photonic PNNs implement analog weights by resonance tuning. In a multi-wavelength silicon photonic circuit, three lasers at $\Theta=\mathbf{C}^{-1}$3, $\Theta=\mathbf{C}^{-1}$4, and $\Theta=\mathbf{C}^{-1}$5 are intensity-modulated, wavelength-multiplexed, split into two MRR weight banks, and combined by a balanced photodetector so that each bank realizes

$\Theta=\mathbf{C}^{-1}$6

with trainable parameters given by tuning currents $\Theta=\mathbf{C}^{-1}$7 through measured mappings $\Theta=\mathbf{C}^{-1}$8 rather than by abstract software weights (Zhang et al., 2024).

Integrated photonic backpropagation has also been demonstrated directly on chip. A 2-input, 8-hidden, 1-output photonic network with PIN-based weight blocks and IM-based nonlinearities implements forward and backward linear maps optically, while the nonlinear activation and its gradient are realized by the same kind of integrated modulator hardware operated in different modes. In this architecture, on-chip training on a nonlinear 2D point-separation task reaches $\Theta=\mathbf{C}^{-1}$9 accuracy across 200 points after 40 epochs, and the resulting performance matches the ideal digital model in both accuracy and robustness for the demonstrated tasks (Ashtiani et al., 17 Jun 2025).

Analog in-memory-computing PNNs model circuit dynamics directly as ODEs. For neuron $\Theta$0 in layer $\Theta$1,

$\Theta$2

so the PNN is the mathematical avatar of the physical IMC circuit itself. To make large-scale training practical, differentiable spike-time discretization reduces computational cost by up to 20 times in speed and 100 times in memory, while post-layout SPICE validation shows that the learned physical model reduces model–hardware discrepancy by at least an order of magnitude (Sakemi et al., 2024).

Mechanical PNNs instantiate the same idea in a different substrate. A network of bistable liquid-filled chambers connected by tubes evolves according to

$\Theta$3

where $\Theta$4 is a weighted graph Laplacian determined by hydraulic conductances, each node has a bistable pressure–volume curve, and the stable binary states provide embedded mechanical memory and direct actuation (Ben-Haim et al., 2024).

At the device level, photonic precision is also shaped by component design. A PCM-based silicon photonic MZI optimized for PNNs achieves $\Theta$5 loss, $\Theta$6 crosstalk, and length $\Theta$7, and this reduction of loss and crosstalk substantially improves system-level scalability and accuracy in large interferometric meshes (Shafiee et al., 2023). Even the input encoding stage matters: combining multiple data features into a single photonic input fixes relative feature importance, and on the Iris dataset an optimal encoding improves accuracy by up to $\Theta$8 relative to other encoding methods, surpassing the performance of networks where features are not combined (Queiroz et al., 2024).

3. Training, robustness, and control under hardware nonidealities

A central line of work treats precision as a training problem over physical control variables. For MRR-based multi-wavelength photonic networks, the trainable variables are tuning currents, updated in situ by perturbation-based gradient descent,

$\Theta$9

with the task loss augmented by a power-aware pruning term,

$$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$0

This directly trades classification accuracy against tuning power. In on-chip Iris experiments, the network maintains $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$1 classification accuracy while reducing tuning power from $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$2 to $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$3, a $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$4 reduction. The same method preserves $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$5 accuracy at both $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$6 and $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$7, lifts RF fingerprinting hardware accuracy from $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$8 under conventional offline mapping to $$\min_{\Theta}\;L\big(F(x;\Theta),y_{\text{target}}\big).$$9 after online fine-tuning, and reduces tuning power by $\Theta$0 on that task. In large simulated CNNs, a CIFAR-100 model with 642,912 weights reduces average tuning power per MRR from $\Theta$1 to $\Theta$2 with accuracy changing from $\Theta$3 to $\Theta$4 (Zhang et al., 2024).

Sharpness-Aware Training addresses the same precision problem from a loss-landscape perspective. It augments standard optimization by searching for flat minima through the min–max objective

$\Theta$5

On an MRR PNN for MNIST, standard backprop and SAT both reach about $\Theta$6 training accuracy in the model, but after deployment on real hardware at $\Theta$7, standard BP drops to $\Theta$8 while SAT remains at $\Theta$9. Under temperature drift from $\mathbf{\Theta}=\mathbf{C}^{-1}$0 to $\mathbf{\Theta}=\mathbf{C}^{-1}$1, BP falls as low as $\mathbf{\Theta}=\mathbf{C}^{-1}$2, whereas SAT stays near the $\mathbf{\Theta}=\mathbf{C}^{-1}$3 level. In MZI-mesh simulations, offline SAT reaches $\mathbf{\Theta}=\mathbf{C}^{-1}$4 accuracy with $\mathbf{\Theta}=\mathbf{C}^{-1}$5, online SAT reaches $\mathbf{\Theta}=\mathbf{C}^{-1}$6 with $\mathbf{\Theta}=\mathbf{C}^{-1}$7, and cross-device transfer at $\mathbf{\Theta}=\mathbf{C}^{-1}$8 remains about $\mathbf{\Theta}=\mathbf{C}^{-1}$9, compared with about $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$0 for DAT (Xu et al., 2024).

Hardware-aware training and pruning can also be cast directly in terms of hardware sensitivity. In integrated photonic networks, the loss is augmented by penalties such as $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$1, $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$2, or $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$3 so that weights are driven into noise-robust and energy-efficient operating regions. On experimental MRR hardware, this improves handwritten-digit accuracy from $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$4 to $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$5 without a thermoelectric controller and yields a 4-bit improvement in computing precision; the same framework reports a tenfold energy reduction and extends to PCM-based PNNs as well (Xu et al., 2024).

Not all physical training methods rely on exact backpropagation through hardware dynamics. In an optoelectronic delay system governed by

$$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$6

the DFA–adjoint method replaces the exact adjoint by a random feedback field

$$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$7

thereby avoiding backward integration of the delayed adjoint equation. Numerically it attains $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$8 accuracy on MNIST, $$\mathbf{x}_f^{(\ell)}= \sigma\!\left(\sum_{j=1}^{F_{\text{in}}^{(\ell)}}\mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})\,\mathbf{x}_j^{(\ell-1)}\right), \qquad \mathbf{H}_{fj}^{(\ell)}(\mathbf{\Theta})=\sum_{k=0}^{K}h_{kfj}^{(\ell)}\,\mathbf{\Theta}^k,$$9 on Fashion-MNIST, and $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$0 on CIFAR-10, versus $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$1, $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$2, and $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$3 for full adjoint training; experimentally it reaches a test accuracy of $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$4 after 20 epochs on a physical optoelectronic delay system, while training time is less than one-fifth of the adjoint method over the investigated delays (Sunada et al., 26 Feb 2025).

These results establish that, in physical PNNs, precision is often achieved less by post hoc calibration than by embedding device physics, drift sensitivity, and control energy directly into the training objective.

4. Precision-matrix and relational PNNs

In graph-relational learning, Precision Neural Networks replace covariance-based graph shifts by the precision matrix $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$5. Because $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$6 encodes conditional independence in the Gaussian case, the resulting graph is typically sparser and more structurally informative than the covariance graph. A PNN layer is

$h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$7

so message passing is polynomial filtering on the precision-induced graph. Spectrally, $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$8 and $h(x;W)=\sum_{i=1}^{k}\phi(w_i^\top x),$9 share eigenvectors, with precision eigenvalues $w_i$0, so PNNs operate in the same eigenspace as PCA while changing the notion of locality from marginal correlation to conditional dependence (Cavallo et al., 18 Sep 2025).

The distinctive contribution of this line is task-aware joint estimation of the graph and the predictor. Rather than first estimating $w_i$1 by graphical lasso and then training a network, the method solves

$w_i$2

subject to $w_i$3 and $w_i$4, with

$w_i$5

Alternating optimization updates the precision matrix, the task-aligned auxiliary precision, and the network parameters in separate steps.

The theoretical analysis gives a finite-sample error bound. If $w_i$6 is the minimizer of the precision-update step and $w_i$7 is the true precision matrix, then with probability tending to 1,

$w_i$8

which separates estimation error into the usual graphical-lasso rate and a task-induced bias term.

Empirically, the method outperforms covariance-based and two-step precision baselines. On ADNI2 cortical-thickness age prediction, MAE improves from $w_i$9 for VNN, $u_{g(j)}$0 for GL-based PNN, and $u_{g(j)}$1 for sample-precision PNN to $u_{g(j)}$2 for joint PNN. On ABIDE, MAE improves from $u_{g(j)}$3 for VNN and $u_{g(j)}$4 for GL-based PNN to $u_{g(j)}$5 for joint PNN, with the learned precision matrix remaining sparse, yielding $u_{g(j)}$6 zeros (Cavallo et al., 18 Sep 2025).

5. Geometric, polynomial, and proximal PNNs

Porcupine Neural Networks impose a geometric constraint on a two-layer ReLU model by forcing each first-layer weight vector to lie on one of finitely many lines. In the matched case, the population loss takes the form

$u_{g(j)}$7

where $u_{g(j)}$8 and $u_{g(j)}$9 are line-wise magnitude sums and $\Theta=\mathbf{C}^{-1}$00 is a kernel applied to the line inner-product matrix $\Theta=\mathbf{C}^{-1}$01. For scalar PNNs and degree-one PNNs, the loss is piecewise convex on sign-pattern regions, and in most regions every local minimum is global. The bad regions are explicitly characterized by all-aligned sign patterns, and empirical as well as theoretical results suggest that an unconstrained network can be approximated using a polynomially-large PNN (Feizi et al., 2017).

Polynomial Neural Networks use multiplicative interactions to alter the spectral bias of learning. For a two-layer $\Theta=\mathbf{C}^{-1}$02-Net,

$\Theta=\mathbf{C}^{-1}$03

the infinite-width NTK yields a $\Theta=\mathbf{C}^{-1}$04-kernel whose high-frequency eigenvalues decay more slowly than those of the standard two-layer ReLU NTK. The paper proves

$\Theta=\mathbf{C}^{-1}$05

whereas the standard NTK has decay $\Theta=\mathbf{C}^{-1}$06. This slower decay increases the RKHS support for high-frequency functions and empirically accelerates the learning of high-frequency spherical harmonics, sinusoids, and image power-spectrum components (Choraria et al., 2022).

Proximal Neural Networks proceed from convex analysis rather than from geometry or spectral theory. A foundational result shows that for a bounded linear operator $\Theta=\mathbf{C}^{-1}$07 with closed range and any proximity operator $\Theta=\mathbf{C}^{-1}$08,

$\Theta=\mathbf{C}^{-1}$09

is again a proximity operator on a suitably renormed Hilbert space. When $\Theta=\mathbf{C}^{-1}$10 is a frame analysis operator and $\Theta=\mathbf{C}^{-1}$11 is soft shrinkage, the frame-shrinkage operator $\Theta=\mathbf{C}^{-1}$12 is therefore a proximity operator. If the $\Theta=\mathbf{C}^{-1}$13 are Parseval frame operators, the resulting Parseval proximal neural network is averaged, has Lipschitz constant 1, and can be trained on Stiefel manifolds by Riemannian optimization (Hasannasab et al., 2019).

A closely related unfolded-proximal line uses fixed numbers of proximal iterations as network depth. For Gaussian denoising, dual forward–backward, dual inertial forward–backward, Chambolle–Pock, and strongly convex Chambolle–Pock are unrolled into DDFB, DDiFB, DCP, and DScCP architectures. With $\Theta=\mathbf{C}^{-1}$14 and $\Theta=\mathbf{C}^{-1}$15, PNN-LNO and PNN-LFO have about $\Theta=\mathbf{C}^{-1}$16 and $\Theta=\mathbf{C}^{-1}$17 parameters, whereas DRUnet has about $\Theta=\mathbf{C}^{-1}$18. DScCP-LNO is reported as the strongest denoiser among the PNNs, and the learned-normalized-operator strategy improves both PSNR and robustness relative to the more flexible but less constrained learned-flexible-operator strategy (Le et al., 2023).

6. Runtime precision, parameter reuse, and scaling

Switchable Precision Neural Networks target runtime control of numerical bitwidth. A single shared backbone is trained to support several quantization modes, with switchable batch normalization and a self-distillation scheme in which the full-precision branch teaches lower-precision branches through a KL loss on outputs and an $\Theta=\mathbf{C}^{-1}$19 loss on features. The base quantizer is

$\Theta=\mathbf{C}^{-1}$20

and the same model can span BinaryConnect-like and BNN-like operating points. On ImageNet with ResNet-18, SP-Net reports $\Theta=\mathbf{C}^{-1}$21 top-1 at $\Theta=\mathbf{C}^{-1}$22-bit weights and $\Theta=\mathbf{C}^{-1}$23-bit activations, $\Theta=\mathbf{C}^{-1}$24 at $\Theta=\mathbf{C}^{-1}$25, $\Theta=\mathbf{C}^{-1}$26 at $\Theta=\mathbf{C}^{-1}$27, and $\Theta=\mathbf{C}^{-1}$28 at $\Theta=\mathbf{C}^{-1}$29. On Tiny ImageNet with $\Theta=\mathbf{C}^{-1}$30-bit weights, the switchable model reaches $\Theta=\mathbf{C}^{-1}$31 at $\Theta=\mathbf{C}^{-1}$32-bit activations versus $\Theta=\mathbf{C}^{-1}$33 for an independently trained model, and $\Theta=\mathbf{C}^{-1}$34 at $\Theta=\mathbf{C}^{-1}$35-bit activations versus $\Theta=\mathbf{C}^{-1}$36 independently (Guerra et al., 2020).

Parameter-efficient scaling in physical PNNs motivates a different kind of precision control: reuse of scarce trainable hardware blocks. ReLaX-Net assumes a time-scale separation between fast forward dynamics and slowly trainable hardware weights, and replaces a single recurrent matrix with a small periodic set $\Theta=\mathbf{C}^{-1}$37 reused over temporal depth $\Theta=\mathbf{C}^{-1}$38: $\Theta=\mathbf{C}^{-1}$39 with $\Theta=\mathbf{C}^{-1}$40 chosen periodically from the $\Theta=\mathbf{C}^{-1}$41 parameter sets. This architecture requires only the addition of fast switches to conventional PNNs. In SVHN experiments at $\Theta=\mathbf{C}^{-1}$42, the stateless-RNN limit $\Theta=\mathbf{C}^{-1}$43 achieves about $\Theta=\mathbf{C}^{-1}$44 accuracy, whereas the no-reuse limit $\Theta=\mathbf{C}^{-1}$45 reaches about $\Theta=\mathbf{C}^{-1}$46, and intermediate values provide a favorable parameter–accuracy trade-off. Similar improvements are reported on Shakespeare next-character prediction (Tsuchiyama et al., 28 Oct 2025).

These two lines use “precision” differently—one as runtime bitwidth, the other as efficient reuse of slow physical weights—but both make precision a selectable operating point rather than a fixed property of the model.

7. Applications, ambiguities, and recurring themes

The application range attached to the PNN label is unusually broad. Physical and photonic PNNs are demonstrated on Iris classification, RF fingerprinting of six nominally identical ZigBee Pro transmitters at $\Theta=\mathbf{C}^{-1}$47, MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100; the same photonic hardware-aware methods are also positioned for photonic switching, LiDAR, and radio-frequency beamforming (Zhang et al., 2024, Sunada et al., 26 Feb 2025, Xu et al., 2024). Precision-matrix PNNs are evaluated on synthetic regression and on brain-age prediction from cortical thickness in ADNI2 and ABIDE (Cavallo et al., 18 Sep 2025). Proximal PNNs are developed for Gaussian denoising and then plugged into forward–backward deblurring algorithms (Le et al., 2023). Mechanical PNNs target smart tech, metamaterials, medical devices, and soft robotics by combining computation with direct physical actuation (Ben-Haim et al., 2024).

A recurrent source of confusion is terminological. In different papers, PNN denotes physical, precision-matrix, porcupine, polynomial, proximal, or switchable-precision neural networks. This suggests that any technical discussion of PNNs must specify what “precision” refers to: analog hardware fidelity, inverse-covariance structure, geometric constraints on weights, spectral treatment of high frequencies, operator-theoretic proximality, or numerical bitwidth. The literature also shows that higher precision need not arise only from more bits or tighter feedback control. In several photonic and analog settings, precision is improved by training the system into flat, low-sensitivity, or low-power operating regions rather than by enforcing ideal behavior externally (Xu et al., 2024, Zhang et al., 2024, Xu et al., 2024).

Taken together, these uses define PNNs not as a single model family but as a family of structurally constrained neural representations. In some cases the constraint is physical, in others statistical, geometric, proximal, polynomial, or quantized. The consequence is the same: network behavior is shaped by an explicit structural prior that makes accuracy, robustness, or efficiency analyzable at a level that standard unconstrained parameterizations often do not expose.