Product-Unit Residual Networks (PURe)

• Product-Unit Residual Networks (PURe) are residual architectures that incorporate learned multiplicative product units to explicitly model feature interactions.
• PURe variants modify classical residual blocks by replacing standard convolutions with product units and utilizing stabilization techniques like log-domain clipping and positivity mappings.
• Empirical studies demonstrate that PURe models achieve faster convergence, improved noise robustness, and competitive performance across image classification, regression, and segmentation tasks.

Product-Unit Residual Networks (PURe) are residual neural architectures that place multiplicative product units inside skip-connected blocks so that feature interactions are modeled explicitly rather than only through stacked additive transformations. In the literature, the term covers several closely related designs: convolutional residual classifiers in which the second convolution of each basic block is replaced by a 2D product unit and explicit activations are removed (Li et al., 7 May 2025); tabular and scientific regression networks in which affine layers surround real- or complex-valued product-unit maps inside shallow residual blocks (Li et al., 5 Jun 2026); and dense prediction variants such as PU-UNet, which insert stabilized product-unit residual blocks into selected low-resolution stages of a residual U-Net (Li et al., 18 Jun 2026). The immediate conceptual precursor is the windowed product unit neural network, which showed that local multiplicative operations can be trained with standard optimizers when the product is kept small and parameter-free (Godfrey et al., 2018).

1. Lineage from classical product units to residual formulations

The intellectual starting point is the classical product unit neural network, in which a unit computes

$y=\prod_{i=1}^{N} x_i^{\theta_i}.$

This parameterization has strong representational capacity: the 2018 work states that a single layer can represent arbitrary polynomials. Its practical difficulty is optimization. The forward pass uses many exponent and log operations, and the derivatives

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$

depend on the full product, making gradients highly sensitive to input scale. That paper characterizes the resulting error surface as chaotic and notes that prior work typically relied on evolutionary or global methods rather than SGD-type optimizers (Godfrey et al., 2018).

Windowed product unit neural networks (WPUNNs) were introduced as a simplification. Instead of learned exponents, all exponents are fixed to $1$, and each unit multiplies only a small window of inputs:

$$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$

The critical residual-style case is $w=s=2$, for which adjacent pairs are multiplied. The product layer is parameter-free, nonlinear in each input, and trainable with standard SGD or Adam because the derivative reduces to $y_i/x_k$ within a window and $0$ outside it. The same study reports that small windows, especially $w=2$, train well, whereas larger windows recover PUNN-like behavior and degrade optimization. This local-product principle is the main bridge from classical product units to later residual constructions: multiplicative interactions are retained, but their numerical footprint is constrained (Godfrey et al., 2018).

The same precursor also established that windowed products can be inserted between standard linear layers and can generalize gated units in recurrent networks. That observation is important historically because later PURe designs repeatedly use the same division of labor: additive layers handle scaling, grouping, and channel mixing, while product units supply the nonlinearity and the interaction structure.

2. Architectural patterns across PURe variants

In convolutional classification, PURe is defined as a deep residual architecture in which the standard $3\times3$ convolution in the second layer of each basic residual block is replaced by a 2D product-unit layer, denoted ConvPU, and all explicit nonlinear activation functions such as ReLU are removed. The high-level ResNet organization is preserved: an initial stem, four residual stages, global average pooling, and a fully connected classifier. For ImageNet-scale models, PURe18 and PURe34 use the same block configurations as ResNet18 and ResNet34, namely $[2,2,2,2]$ and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$0. For CIFAR-10, the stem becomes a single $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$1 convolution with stride $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$2, and Block D is removed. Each product-unit residual block also carries one extra scalar parameter, the log-threshold used for numerical stability (Li et al., 7 May 2025).

The residual block itself changes from the usual

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$3

to

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$4

with output

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$5

Thus the main path becomes multiplicative, the skip path remains additive, and the overall block remains residual.

In tabular and scientific regression, PURe denotes a different but structurally analogous family. The real-valued residual block is

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$6

where, for $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$7,

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$8

The full real-valued network uses an input projection with ReLU, two residual PURe blocks in sequence, and a final affine readout. The complex-valued variant keeps the same residual structure but uses complex affine layers and a complex product-unit transform. Safe preprocessing is imposed through

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$9

followed by complex $1$0 and $1$1. To compare fairly with real-valued models, the hidden dimension is $1$2 for real networks and $1$3 for complex networks, reflecting matched parameter budgets because each complex parameter corresponds to two real parameters (Li et al., 5 Jun 2026).

In dense prediction, PU-UNet is a U-Net instantiation of the same residual principle. The baseline is a residual U-Net with depth $1$4, base width $1$5, channel dimensions $1$6, GroupNorm with $1$7 groups, and SiLU activations. PU-UNet replaces selected residual blocks by product-unit residual blocks whose first stage remains standard,

$1$8

and whose second stage uses a stabilized ConvPU followed by GroupNorm:

$1$9

The main reported configuration inserts these blocks only in the final encoder block at $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$0 resolution, the bottleneck, and the first two decoder blocks at $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$1 and $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$2 resolution, leaving higher-resolution stages purely additive (Li et al., 18 Jun 2026).

3. Mathematical mechanisms and numerical stabilization

The shared mechanism behind PURe is the log-linear-exponential realization of multiplicative interactions. In the convolutional classifier formulation, a scalar product unit is

$$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$3

and the 2D product-unit convolution at location $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$4 is

$$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$5

Implementation therefore reduces to log, standard convolution, and exp. Because images and feature maps may contain zero or negative values, a trainable threshold is introduced:

$$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$6

after which $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$7 is taken. ConvPU weights are initialized with Kaiming uniform with a scaling factor $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$8, and $$M=\left\lfloor\frac{N-w}{s}\right\rfloor+1,\qquad y_i=\prod_{j=0}^{w-1} x_{si+j}.$$9 is initialized to $w=s=2$0. The paper explicitly attributes stable training to the combination of this thresholding and residual learning (Li et al., 7 May 2025).

The tabular real-valued PURe adopts a different positivity mechanism. Instead of thresholding raw hidden features, it uses $w=s=2$1, which guarantees strictly positive inputs for the logarithm while also making inactive components multiplicatively neutral because $w=s=2$2. The complex-valued variant avoids the real-domain sign restriction by moving to $w=s=2$3, but it still protects against zero and unstable magnitudes through first-quadrant rectification with offset $w=s=2$4. Additional stabilization comes from zero-initializing the second residual block so that it begins near the identity. All tabular models are trained with Adam, learning rate $w=s=2$5, batch size $w=s=2$6, and $w=s=2$7 epochs, with Kaiming initialization applied to real parameters and independently to real and imaginary parts in complex parameters (Li et al., 5 Jun 2026).

PU-UNet refines the stabilization strategy further. Its ConvPU first applies a smooth positivity mapping

$w=s=2$8

then performs log-domain convolution,

$w=s=2$9

then clips the log-domain response,

$y_i/x_k$0

and finally reconstructs multiplicative outputs by

$y_i/x_k$1

The clipping constant is chosen so that $y_i/x_k$2, keeping activations in a safe FP16 range. The paper reports that removing log-domain clipping causes gradient explosion, while replacing the smooth positivity mapping by a simpler lower-bound truncation degrades Dice and IoU across all datasets. GroupNorm before and after the multiplicative stage, together with the residual shortcut, is treated as an integral part of stable deep training (Li et al., 18 Jun 2026).

A recurrent theme across these formulations is that the product unit is never deployed as an unconstrained multiplicative stack. It is wrapped by linear or convolutional transforms, normalization, positivity controls, and residual addition. This suggests that “PURe” is best understood not as a single neuron type but as a stabilized residual design pattern for multiplicative computation.

4. Expressive role and inductive bias

Product units are introduced because they encode multiplicative structure directly. The vision PURe work emphasizes that, unlike standard summation neurons, product units can represent high-order monomials, sparse polynomials, power laws, roots, and fractional powers. A 2D product-unit kernel therefore models a monomial over its receptive field, whereas a standard convolution models a weighted sum before an external nonlinearity. The argument is that such multiplicative operators increase expressivity per parameter, allowing richer local feature couplings at equal or lower depth (Li et al., 7 May 2025).

The earlier WPUNN formulation makes this claim more concrete. It shows that $y_i/x_k$3 can be represented exactly by duplicating a scalar $y_i/x_k$4 through a fully connected layer and applying a windowed product of size $y_i/x_k$5, and that arbitrary polynomials can be represented exactly with sufficient width and with depth roughly logarithmic in the polynomial degree. The same work also identifies a direct connection between pairwise products and gating in recurrent architectures: operations of the form $y_i/x_k$6 are exactly windowed products with $y_i/x_k$7. A plausible implication is that later PURe blocks inherit both a polynomial inductive bias and a natural capacity for gating-like modulation (Godfrey et al., 2018).

In tabular regression, this inductive bias is made explicit at the level of feature interactions. Each PU unit

$y_i/x_k$8

acts as a learned multiplicative coupling among transformed features. Because the inputs to the PU are already affine mixtures of earlier hidden variables, the model can realize high-order interactions among original predictors in a compact residual block. The 2026 study frames this as an alternative to the implicit and entangled interaction formation of standard MLPs and evaluates the resulting models partly through SHAP-based interaction maps, arguing that PURe learns more concentrated and structurally coherent interaction patterns than MLP baselines (Li et al., 5 Jun 2026).

In dense prediction, the same logic is transposed from tabular feature pairs to spatial-semantic couplings. PU-UNet argues that multiplicative interactions are most beneficial in low-resolution stages, where features are semantically richer and encode object-level structure rather than only local edge statistics. Its ablation results support that placement hypothesis: using product-unit blocks only in deep encoder, bottleneck, and early decoder stages is better than using them only in the bottleneck and better than replacing all stages (Li et al., 18 Jun 2026).

5. Empirical performance across domains

In image classification, the 2025 PURe study evaluates Galaxy10 DECaLS, ImageNet, and CIFAR-10. On Galaxy10 DECaLS, PURe34 attains the highest reported test accuracy, $y_i/x_k$9, with mean $0$0, and the smallest Poisson-noise performance drop, $0$1. PURe18 reaches $0$2 best test accuracy and outperforms ResNet18 and ResNet34. Convergence is also faster: time to $0$3 validation accuracy is $0$4 s for PURe34 versus $0$5 s for ResNet152, and for similar accuracy levels the paper reports that PURe is about $0$6 faster to reach best validation performance, $0$7 s versus $0$8 s. On ImageNet, PURe18 achieves $0$9 top-1 and $w=2$0 top-5 with $w=2$1M parameters, while PURe34 reaches $w=2$2 top-1 and $w=2$3 top-5 with $w=2$4M parameters, exceeding ResNet34 and ResNet50 and slightly surpassing ResNet101 in both top-1 and top-5 while approaching ResNet152 with one-third of the parameters. On CIFAR-10, PURe consistently improves over matched-depth ResNets: PURe110 reaches $w=2$5, PURe110$w=2$6 reaches $w=2$7, and PURe272$w=2$8 reaches $w=2$9 with $3\times3$0M parameters, compared with $3\times3$1 for ResNet1001 at $3\times3$2M parameters (Li et al., 7 May 2025).

In interaction-driven regression, the 2026 PURe study compares real-valued and complex-valued MLP and PURe models on Friedman 1, Concrete Compressive Strength, and California Housing. On Friedman 1, the best clean and noisy test MSEs are both obtained by CV-PURe: $3\times3$3 at $3\times3$4 and $3\times3$5 at $3\times3$6. On Concrete, the numerically best clean and noisy results are again CV-PURe, $3\times3$7 and $3\times3$8, while RV-PURe improves over RV-MLP. On California Housing, CV-PURe achieves $3\times3$9 clean and $[2,2,2,2]$0 under Gaussian feature noise, outperforming all baselines. Sample-efficiency experiments on Concrete further show that CV-PURe is best at every training fraction from $[2,2,2,2]$1 to $[2,2,2,2]$2, and at $[2,2,2,2]$3 data it reaches $[2,2,2,2]$4 compared with $[2,2,2,2]$5 for RV-MLP (Li et al., 5 Jun 2026).

In medical image segmentation, PU-UNet reports substantial gains over a matched Residual U-Net while keeping computational cost nearly unchanged. On ISIC 2018, PU-UNet reaches Dice $[2,2,2,2]$6 and IoU $[2,2,2,2]$7, compared with $[2,2,2,2]$8 and $[2,2,2,2]$9 for the baseline. On Kvasir-SEG, it reaches Dice $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$00 and IoU $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$01, compared with $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$02 and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$03. On BUSI including normal images, lesion Dice and IoU are $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$04 and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$05, with image-level false-positive rate reduced from $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$06 to $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$07; excluding normal cases, Dice rises to $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$08 and IoU to $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$09. At input size $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$10, PU-UNet uses $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$11M parameters, $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$12G FLOPs, and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$13 ms latency, compared with $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$14M, $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$15G, and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$16 ms for the baseline. At $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$17, the corresponding values are $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$18M, $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$19G, and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$20 ms versus $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$21M, $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$22G, and $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$23 ms (Li et al., 18 Jun 2026).

6. Interpretation, limitations, and open questions

A recurring empirical theme is robustness. In astronomy classification, PURe34 exhibits the smallest relative performance degradation under Poisson-distributed photon noise, $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$24, compared with $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$25 for the best ResNet baseline and up to $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$26 for ResNet34. In tabular regression, PURe models, especially CV-PURe, show smaller MSE increases under Gaussian feature noise. In medical segmentation, PU-UNet reduces the image-level false-positive rate on normal BUSI cases from $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$27 to zero. These results are presented as evidence that explicit multiplicative structure can improve robustness when the target depends on coupled features or when nuisance variation is signal-dependent (Li et al., 7 May 2025, Li et al., 5 Jun 2026, Li et al., 18 Jun 2026).

Interpretability is most explicitly addressed in the tabular study. Using SHAP-based interaction analysis, the authors compute normalized entropy

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$28

and Herfindahl-Hirschman Index

$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$29

over the top-$$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$30 interaction pairs, with $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$31. Lower entropy and higher HHI indicate more concentrated interaction structure. On Friedman 1, for example, RV-MLP yields $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$32 and HHI $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$33, whereas CV-PURe yields $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$34 and HHI $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$35. The same paper reports that PURe interaction maps are more sharply focused on plausible or known feature couplings than MLP baselines (Li et al., 5 Jun 2026).

Several misconceptions are clarified by the literature itself. First, PURe is not a single block formula. The 2025 classifier uses exponent-weighted ConvPU layers with all ReLUs removed; the 2026 tabular models retain ReLU outside the product-unit sub-block; PU-UNet retains GN and SiLU around the multiplicative stage. Second, PURe is not identical to WPUNN. WPUNN uses parameter-free local products with fixed exponents $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$36, whereas the later PURe formulations use learned exponents in log-linear-exp maps. Third, residual connections are not by themselves the source of the observed gains. On the Concrete ablation, the real-valued Res-MLP baseline scores $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$37, worse than the MLP baseline at $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$38, while the non-residual PU model improves to $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$39 and the full PURe reaches the best result, $$\frac{\partial y}{\partial x_k}=y\frac{\theta_k}{x_k}, \qquad \frac{\partial y}{\partial \theta_k}=y\ln x_k$$40 (Li et al., 5 Jun 2026).

The limitations are also consistent across studies. Product units remain sensitive to initialization, scale, and positivity handling. The convolutional vision PURe requires thresholded logs; PU-UNet requires both smooth positivity mapping and log-domain clipping, and the unclipped variant exhibits gradient explosion. In classification, performance saturates beyond PURe272 on CIFAR-10, and PURe486 is reported as suffering from gradient explosion and non-convergence. The tabular study is intentionally narrow, covering only three regression tasks, and its interpretability analysis is described as SHAP-based and mainly qualitative in character. PU-UNet is limited to 2D medical image segmentation with a U-Net-style backbone (Li et al., 7 May 2025, Li et al., 5 Jun 2026, Li et al., 18 Jun 2026).

The current literature therefore supports a precise characterization of PURe: it is a family of residual architectures that use stabilized multiplicative operators to encode explicit feature interactions. The family is heterogeneous in implementation but consistent in motivation. Its reported advantages are strongest when interaction structure is central to the task, when robustness to noise matters, or when parameter efficiency is important. Its unresolved issues concern depth scaling, numerical stabilization, and transfer beyond the currently studied regimes.