DYS-Net: Diverse Neural Architectures in AI

• DYS-Net is a polysemous research label covering distinct neural architectures applied to survival analysis, image restoration, decision-focused learning, and EEG dyslexia modeling.
• In survival analysis, DyS-based models offer interpretable, feature-sparse, discrete-time survival curves that adapt to time-varying risks without relying on proportional hazards.
• Optimization-focused DYS-Net variants employ Davis–Yin splitting and differentiable layers to achieve efficient convergence in image restoration and decision-making applications.

DYS-Net is not a single universally standardized model name in the arXiv literature. The available usage spans several technically distinct constructions: a neural generalized additive survival model based on DyS for censored time-to-event prediction; a Davis–Yin splitting-based plug-and-play deep network for image restoration; a fully neural differentiable optimization layer for quadratically regularized linear programs in decision-focused learning; and, more loosely, a dyslexia-related network model derived from EEG phase–amplitude coupling. The shared label reflects the recurrence of “DyS/DYS” as either DyNAMic Survival or Davis–Yin splitting, rather than a single canonical architecture (Ness et al., 2024, Wu et al., 2024, Mandi et al., 15 Aug 2025, Gallego-Molina et al., 2021).

1. Terminological scope and disambiguation

The available papers suggest that “DYS-Net” is best treated as a polysemous research label. In some cases it is an informal abbreviation supplied in later exposition rather than a term explicitly used by the original authors.

Context	DYS-Net refers to	Source
Survival analysis	Neural implementation of DyS, a feature-sparse GA²M/NAM for censored survival prediction	(Ness et al., 2024)
Image restoration	Deep-unfolded extrapolated plug-and-play Davis–Yin splitting network	(Wu et al., 2024)
Decision-focused learning	Fully neural differentiable optimization layer for quadratically regularized LPs	(Mandi et al., 15 Aug 2025)
EEG dyslexia modelling	Conceptual network framework built from PAC-derived complex networks	(Gallego-Molina et al., 2021)

A common misconception is that DYS-Net denotes one architecture with a fixed mathematical signature. The papers instead indicate a family of constructions tied either to DyS in survival modelling or to DYS as Davis–Yin three-operator splitting. A second misconception is historical: in the survival and dyslexia papers, the term “DYS-Net” does not explicitly appear in the original text; it is a conceptual relabelling of the underlying method (Ness et al., 2024, Gallego-Molina et al., 2021).

2. DyS as a survival-analysis DYS-Net

In survival analysis, DyS is introduced as a glass-box survival model that is simultaneously interpretable and feature-sparse. Its functional form is a generalized additive model with interactions, implemented as a neural additive model:

$$f(X) = \sum_{j=1}^p f_j(X_j) + \sum_{j\neq \ell} f_{j,\ell}(X_j,X_\ell).$$

Here each main-effect or interaction module outputs a vector in $\mathbb{R}^K$, one logit per discrete time bin, and the final event-time mass function is obtained by a softmax over the additive logits (Ness et al., 2024).

The model uses a discrete-time survival formulation rather than a Cox-style proportional hazards parameterization. For evaluation times $\{t_1,\dots,t_K\}$, DyS computes

$$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$

and then derives the survival curve by cumulative summation,

$$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$

The paper emphasizes that DyS does not assume proportional hazards, so relative risk rankings may change over time. This is central to its interpretability claims, because the learned feature effects are explicitly time-specific rather than compressed into a single PH-compatible risk score (Ness et al., 2024).

Training is based on the discrete Ranked Probability Score. For a sample $(X,T,\delta)$,

$$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$

This formulation handles censoring directly: uncensored observations receive penalties both before and after the event time, whereas censored observations are penalized only for survival underestimation before censoring. The paper’s interpretation is that DyS optimizes survival curves themselves rather than an auxiliary discrimination surrogate (Ness et al., 2024).

Feature selection is integrated into training through smooth-step gates attached to each shape function. The sparse model takes the form

$$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$

with a sparsity penalty

$$L_{\mathrm{SS}}(\mu)=\lambda\left(\sum_j s(\mu_j)+\alpha\sum_{j,\ell}s(\mu_{j,\ell})\right),$$

and an entropy regularizer that encourages hard $0/1$ gate values. This yields feature selection at the level of entire learned shape functions rather than individual coefficients. The training objective combines RPS, sparsity, and entropy terms, and optimization uses Adam with learning rate $\mathbb{R}^K$0, up to 200 epochs, with early stopping if validation performance does not improve for 5 epochs (Ness et al., 2024).

Interpretability follows from three structural choices: additive decomposition, explicit interaction surfaces, and feature sparsity. The paper defines time-specific feature importance by averaging the absolute value of the logits across samples, and it visualizes main effects as time-indexed shape functions and pairwise interactions as 2D surfaces. On a heart failure dataset of 672,296 patients with 2,410 features and 97.5% censoring, DyS selects $\mathbb{R}^K$1 features, achieves near state-of-the-art AUC, and exposes clinically legible short-term versus long-term effects. On metabric, the paper reports CoxPH $\mathbb{R}^K$2 versus DyS $\mathbb{R}^K$3 mean time-dependent AUC; on the same heart failure dataset, DeepHit attains $\mathbb{R}^K$4 and DyS $\mathbb{R}^K$5, with DyS using only 45–65 features rather than all 2,410 (Ness et al., 2024).

3. Extrapolated plug-and-play DYS-Net for image restoration

In image restoration, DYS refers to Davis–Yin three-operator splitting for problems of the form

$\mathbb{R}^K$6

where $\mathbb{R}^K$7 and $\mathbb{R}^K$8 are smooth, possibly nonconvex, and $\mathbb{R}^K$9 is proper, closed, and possibly nonsmooth/nonconvex. The extrapolated DYS algorithm introduces the inertial point

$\{t_1,\dots,t_K\}$0

followed by a proximal step for $\{t_1,\dots,t_K\}$1, a proximal step for $\{t_1,\dots,t_K\}$2, and a residual-style update (Wu et al., 2024).

The plug-and-play construction replaces one proximal map with a learned denoiser. The denoiser is the Gradient Step denoiser

$\{t_1,\dots,t_K\}$3

with

$\{t_1,\dots,t_K\}$4

where $\{t_1,\dots,t_K\}$5 is a differentiable neural network. A key proposition in the paper states that $\{t_1,\dots,t_K\}$6 for a specific nonconvex functional $\{t_1,\dots,t_K\}$7, so the denoiser is not merely heuristic but can be reinterpreted as an exact proximal operator of an explicit nonconvex prior (Wu et al., 2024).

This yields two extrapolated PnP-DYS schemes. In the smooth variant,

$\{t_1,\dots,t_K\}$8

When unrolled, these iterations define what the paper informally characterizes as a DYS-based PnP network: each layer consists of a proximal data-fidelity block, a gradient step on $\{t_1,\dots,t_K\}$9, a denoising block $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$0, and an extrapolated skip connection (Wu et al., 2024).

The convergence theory is KL-based. Under explicit conditions $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$1 and $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$2, the paper proves monotone decrease of an extrapolated envelope, summability of increments, an $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$3 stationarity bound, and convergence of the whole sequence to a critical point when the objective is a KL function and the iterates are bounded. This is one of the central technical distinctions of this DYS-Net usage: the unrolled architecture is derived from an operator-splitting method with a nonconvex convergence guarantee, rather than from unconstrained architectural heuristics (Wu et al., 2024).

Empirically, the method is applied to deblurring and super-resolution. On Set3C with Ker1 deblurring, degraded images at noise level $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$4 have PSNR $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$5, while DeTik and DeBox achieve $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$6 and $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$7, respectively. On Set5 super-resolution at scale $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$8 with Ker1 and noise level $$\{\hat P(T=t_k\mid X)\}_{k=1}^K=\mathrm{softmax}(f(X)),$$9, bicubic gives $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$0, while DeTik and DeBox achieve $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$1 and $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$2. The paper reports that larger $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$3 values near the theoretical maximum significantly reduce iterations and time without degrading PSNR or SSIM (Wu et al., 2024).

4. DYS-Net as a differentiable optimization layer for decision-focused learning

In decision-focused learning, DYS-Net denotes a fully neural differentiable optimization layer for quadratically regularized linear programs. The underlying LP is

$$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$4

and differentiability is induced by the smoothed QP

$$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$5

The purpose of the layer is to approximate both the solution $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$6 and its gradient with respect to the predicted costs $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$7, while avoiding the matrix factorizations required by exact differentiable solvers such as OptNet or CvxpyLayers (Mandi et al., 15 Aug 2025).

The forward pass is a fixed-point iteration derived from Davis–Yin splitting:

$$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$8

where $$\hat S(t_k\mid X)=1-\sum_{m:t_m\le t_k}\hat P(T=t_m\mid X).$$9 and $(X,T,\delta)$0. The projections are

$(X,T,\delta)$1

Because each iteration uses only matrix–vector products, pseudo-inverse multiplication, elementwise $(X,T,\delta)$2, and affine combinations, the layer can be implemented as a stack of feedforward operations and run efficiently on GPUs. The paper uses 100 iterations, $(X,T,\delta)$3, and a problem-dependent smoothing strength $(X,T,\delta)$4 (Mandi et al., 15 Aug 2025).

Backward differentiation uses Jacobian-free backpropagation, in which the Jacobian of the fixed-point mapping with respect to the iterate is approximated by the identity. This makes the backward pass substantially cheaper than exact implicit differentiation. The paper’s central finding, however, is that differentiability alone does not resolve the zero-gradient problem of regret minimization. Even after smoothing, the mapping $(X,T,\delta)$5 remains nearly piecewise constant across large regions, so gradients of regret with respect to $(X,T,\delta)$6 are zero or negligible (Mandi et al., 15 Aug 2025).

The proposed remedy is to train the predictor through surrogate losses, especially SCE and SPO+, rather than through regret or SqDE directly. With DYS-Net as the differentiable solver, the combination of SCE and DYS-Net is denoted SCEDYS. On shortest path, knapsack, and capacitated facility location benchmarks, the paper reports that minimizing regret or SqDE with CvxpyLayers or DYS-Net yields higher test regret than minimizing SPO+ or SCE. SCEDYS is described as achieving regret comparable to or better than strong DFL baselines while reducing training time by factors of 3–5 in several settings; on larger CFL instances, SCEDYS achieves both the lowest regret and the lowest runtime among the compared DFL methods (Mandi et al., 15 Aug 2025).

5. Operator-splitting theory underlying DYS-based networks

Two papers provide the mathematical substrate for DYS-based architectures without themselves defining a neural DYS-Net. In the convex setting, Davis–Yin splitting is analyzed through scaled relative graphs for problems

$(X,T,\delta)$7

with $(X,T,\delta)$8 convex, closed, and proper, and $(X,T,\delta)$9 differentiable with $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$0-Lipschitz gradient. The DYS operator takes the form

$$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$1

so a DYS block consists of two proximal maps and one gradient map. The paper derives explicit Lipschitz and contraction factors, including the compact rate

$$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$2

and characterizes these as state-of-the-art linear convergence rates for convex optimization problems (Yi et al., 2022).

In the nonconvex setting, a complementary analysis treats problems

$$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$3

with smooth possibly nonconvex $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$4 and $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$5, and proper closed possibly nonsmooth/nonconvex $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$6. The DYS iteration alternates a proximal step on $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$7, a proximal step on $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$8, and an explicit gradient step on $$L_{\mathrm{RPS}}(\hat S,T,\delta)=\sum_{t<T}(1-\hat S(t\mid X))^2+\delta\sum_{t\ge T}\hat S(t\mid X)^2.$$9. A specially constructed energy function $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$0 decreases when the step size satisfies a computable threshold condition, and the paper proves that every cluster point is stationary. Under the KL property, the whole sequence converges, with finite-length trajectories and local rates determined by the KL exponent (Bian et al., 2020).

These two analyses imply that DYS-Net architectures based on Davis–Yin updates inherit a well-developed operator-theoretic lineage. A plausible implication is that the term “DYS-Net” is often meaningful only when paired with its underlying optimization regime: convex contractive splitting, nonconvex KL-controlled splitting, or a task-specific unrolling that uses those primitives.

6. Dyslexia-related network modelling and interpretive issues

A fourth usage is more tentative. In the dyslexia paper, “DYS-Net” is not named by the authors, but the study constructs a network model that can serve as the core of such a framework. The data comprise 48 children, with 32 controls and 16 dyslexic readers, recorded using a Brainvision actiCHamp Plus system with 32 active electrodes at 500 Hz during 15-minute sessions. Auditory stimulation consists of white noise amplitude-modulated at 4.8 Hz, 16 Hz, and 40 Hz. EEG is segmented into 15.02 s windows, phase–amplitude coupling is estimated using Tort’s Modulation Index with 18 phase bins, and significance is assessed against 200 surrogate MI values with threshold $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$1 (Gallego-Molina et al., 2021).

The resulting networks are frequency-centric rather than spatial. Nodes are the five EEG bands Delta, Theta, Alpha, Beta, and Gamma; edges indicate whether at least one electrode exhibits significant PAC from a phase-providing band to an amplitude-modulated band in a given segment. Graph metrics include degree, density, clustering coefficient, characteristic path length, small-worldness, and betweenness centrality. The study reports that dyslexic networks tend to have fewer edges, reduced clustering at particular time points, and a transient loss of small-world topology under the 4.8 Hz stimulus (Gallego-Molina et al., 2021).

For classification, the paper uses graph-metric time series as SVM features with stratified 5-fold cross-validation. The most discriminative configuration is 4.8 Hz small-worldness, with accuracy $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$2, sensitivity $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$3, specificity $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$4, and AUC $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$5; permutation testing yields $$f(X)=\sum_j f_j(X_j)s(\mu_j)+\sum_{j,\ell}f_{j,\ell}(X_j,X_\ell)s(\mu_{j,\ell}),$$6. By contrast, 16 Hz and 40 Hz small-worldness do not surpass chance under the same test. The authors explicitly frame the work as exploratory and emphasize that it is a methodological step toward objective, early diagnosis rather than a ready clinical classifier (Gallego-Molina et al., 2021).

This usage highlights the broadest interpretive issue around the term. Here, “DYS-Net” would denote not a named model family but an inferred EEG-to-PAC-to-graph-to-classifier pipeline. That differs categorically from the named survival, imaging, and differentiable-optimization constructions. The unifying idea is therefore weakly architectural rather than taxonomic: DYS-Net tends to denote a networked realization of a DyS/DYS mechanism within a particular domain, but the mechanism itself changes across domains.