Residual Learning

The Universal Approximation Theorem (UAT) guarantees universal function approximation but does not explain how residual models distribute approximation across layers. We reframe residual networks as a layer-wise approximation process that builds an approximation trajectory from input to target, and prove the existence of progressive trajectories where error decreases monotonically with depth. It reveals that residual networks can implement structured, step-by-step refinement rather than end-to-end (E2E) black-box mapping. Building on this, we propose Layer-wise Progressive Approximation (LPA), a theoretically grounded training principle that explicitly aligns each layer with its residual target to realize such trajectories. LPA is architecture-agnostic: we observe progressive behavior in residual FNNs, ResNets, and Transformers across tasks including complex surface fitting, image classification, and NLP with LLMs for generation and classification. Crucially, this enables ``train once, use $N$ models": a single network yields useful predictions at every depth, supporting efficient shallow inference without retraining. Our work unifies approximation theory with practical deep learning, providing a new lens on representation learning and a flexible framework for multi-depth deployment. The source code will be released unpon acceptance at https://(open_upon_acceptance).

The benefits usually attributed to skip connections are optimization-theoretic: a smoother loss landscape and better gradient propagation. We ask a representational question instead: given a residual block x -> x + MLP(x), does a residual-free MLP of the same width compute the same function? The answer is no, unconditionally and at every depth, for every activation used in current frontier language models (ReLU^2, ReGLU, SwiGLU, GeGLU). For ungated ReLU and GELU absorption is possible, but only on a set of weights of measure zero. The two families are therefore generically disjoint: removing a skip connection and retraining cannot recover exactly the same function at equal width.

Accurate dynamical modeling is essential for simulation and control of embodied systems, yet first-principles models of electromechanical systems often fail to capture complex dissipative effects such as joint friction, stray losses, and structural damping. While residual-learning physics-informed neural networks (PINNs) can effectively augment imperfect first-principles models with data-driven components, the residual terms are typically implemented as unconstrained multilayer perceptrons (MLPs), which may inadvertently inject artificial energy into the system. To more faithfully model the dissipative dynamics, we propose DiLaR-PINN, a dissipative latent residual PINN designed to learn unmodeled dissipative effects in a physically consistent manner. Structurally, the residual network operates only on unmeasurable (latent) state components and is parameterized in a skew-dissipative form that guarantees non-increasing energy for any choice of network parameters. To enable stable and data-efficient training under partial measurability of the state, we further develop a recurrent rollout scheme with a curriculum-based sequence length extension strategy. We validate DiLaR-PINN on a real-world helicopter system and compare it against four baselines: a pure physical model (without a residual network), an unstructured residual MLP, a DiLaR variant with a soft dissipativity constraint, and a black-box LSTM. The results demonstrate that DiLaR-PINN more accurately captures dissipative effects and achieves superior long-horizon extrapolation performance.

Cross-dose denoising for low-dose positron emission tomography (LDPET) has been proposed to address the limited generalization of models trained at a single noise level. However, neural networks trained on a specific dose level often fail to generalize to other dose conditions due to variations in noise magnitude and statistical properties. Conventional "one-size-for-all" models attempt to mitigate this variability but tend to learn averaged representations across noise levels, resulting in degraded performance. In this work, we analyze this limitation and show that standard training formulations implicitly optimize an expectation over heterogeneous noise distributions, causing the network to learn an averaged denoising mapping that cannot accurately model dose-specific noise characteristics. We propose a unified residual noise learning framework that estimates noise directly from low-dose PET images rather than predicting full-dose images. Experiments on large-scale multi-dose PET datasets from two medical centers demonstrate that the proposed method outperforms the "one-size-for-all" model, individual dose-specific U-Net models, and dose-conditioned approaches, achieving improved denoising performance. These results indicate that residual noise learning effectively mitigates the averaging effect and enhances generalization for cross-dose PET denoising.

Long-tailed classification, where a small number of frequent classes dominate many rare ones, remains challenging because models systematically favor frequent classes at inference time. Existing post-hoc methods such as logit adjustment address this by adding a fixed classwise offset to the base-model logits. However, the correction required to restore the relative ranking of two classes need not be constant across inputs, and a fixed offset cannot adapt to such variation. We study this problem through Bayes-optimal reranking on a base-model top-k shortlist. The gap between the optimal score and the base score, the residual correction, decomposes into a classwise component that is constant within each class, and a pairwise component that depends on the input and competing labels. When the residual is purely classwise, a fixed offset suffices to recover the Bayes-optimal ordering. We further show that when the same label pair induces incompatible ordering constraints across contexts, no fixed offset can achieve this recovery. This decomposition leads to testable predictions regarding when pairwise correction can improve performance and when cannot. We develop REPAIR (Reranking via Pairwise residual correction), a lightweight post-hoc reranker that combines a shrinkage-stabilized classwise term with a linear pairwise term driven by competition features on the shortlist. Experiments on nine benchmarks confirm that the decomposition explains where pairwise correction helps and where classwise correction alone suffices. These span text classification, visual recognition, and multimodal rare-disease diagnosis.

Autonomous Aerial Manipulators (AAMs) are inherently coupled, nonlinear systems that exhibit nonstationary and multiscale residual dynamics, particularly during manipulator reconfiguration and abrupt payload variations. Conventional analytical dynamic models rely on fixed parametric structures, while static data-driven model assume stationary dynamics and degrade under configuration changes and payload variations. Moreover, existing learning architectures do not explicitly factorize cross-variable coupling and multi-scale temporal effects, conflating instantaneous inertial dynamics with long-horizon regime evolution. We propose a predictive-adaptive framework for real-time residual modeling and compensation in AAMs. The core of this framework is the Factorized Dynamics Transformer (FDT), which treats physical variables as independent tokens. This design enables explicit cross-variable attention while structurally separating short-horizon inertial dependencies from long-horizon aerodynamic effects. To address deployment-time distribution shifts, a Latent Residual Adapter (LRA) performs rapid linear adaptation in the latent space via Recursive Least Squares, preserving the offline nonlinear representation without prohibitive computational overhead. The adapted residual forecast is directly integrated into a residual-compensated adaptive controller. Real-world experiments on an aerial manipulator subjected to unseen payloads demonstrate higher prediction fidelity, accelerated disturbance attenuation, and superior closed-loop tracking precision compared to state-of-the-art learning baselines, all while maintaining strict real-time feasibility.

Scaling network depth has been a central driver behind the success of modern foundation models, yet recent investigations suggest that deep layers are often underutilized. This paper revisits the default mechanism for deepening neural networks, namely residual connections, from an optimization perspective. Rigorous analysis proves that the layout of residual connections can fundamentally shape convergence behavior, and even induces an exponential gap in convergence rates. Prompted by this insight, we introduce adaptive neural connection reassignment (ANCRe), a principled and lightweight framework that parameterizes and learns residual connectivities from the data. ANCRe adaptively reassigns residual connections with negligible computational and memory overhead ($<1\%$), while enabling more effective utilization of network depth. Extensive numerical tests across pre-training of large language models, diffusion models, and deep ResNets demonstrate consistently accelerated convergence, boosted performance, and enhanced depth efficiency over conventional residual connections.

We propose Self-Augmented Residual 3D Gaussian Splatting (SA-ResGS), a novel framework to stabilize uncertainty quantification and enhancing uncertainty-aware supervision in next-best-view (NBV) selection for active scene reconstruction. SA-ResGS improves both the reliability of uncertainty estimates and their effectiveness for supervision by generating Self-Augmented point clouds (SA-Points) via triangulation between a training view and a rasterized extrapolated view, enabling efficient scene coverage estimation. While improving scene coverage through physically guided view selection, SA-ResGS also addresses the challenge of under-supervised Gaussians, exacerbated by sparse and wide-baseline views, by introducing the first residual learning strategy tailored for 3D Gaussian Splatting. This targeted supervision enhances gradient flow in high-uncertainty Gaussians by combining uncertainty-driven filtering with dropout- and hard-negative-mining-inspired sampling. Our contributions are threefold: (1) a physically grounded view selection strategy that promotes efficient and uniform scene coverage; (2) an uncertainty-aware residual supervision scheme that amplifies learning signals for weakly contributing Gaussians, improving training stability and uncertainty estimation across scenes with diverse camera distributions; (3) an implicit unbiasing of uncertainty quantification as a consequence of constrained view selection and residual supervision, which together mitigate conflicting effects of wide-baseline exploration and sparse-view ambiguity in NBV planning. Experiments on active view selection demonstrate that SA-ResGS outperforms state-of-the-art baselines in both reconstruction quality and view selection robustness.

Real-world time-series regression often involves non-stationarity, heteroscedasticity, and regime changes, under which recent prediction errors may contain structured information about local temporal mismatch between model predictions and observations. Learning how to represent and reuse these errors can therefore provide useful information for subsequent prediction. We introduce the Variability-Aware Recursive Neural Network (VARNN), a residual-aware architecture for supervised time-series regression that learns an explicit residual-memory state from recent prediction errors and uses it to condition subsequent predictions. Specifically, VARNN maps scalar prediction innovations into a learned nonlinear, vector-valued residual representation over a short context. Across nine datasets spanning energy, healthcare, and environmental domains, VARNN achieves lower test MSE than the compared static, lag-based, and sequence-model baselines. Targeted ablations further show that learned projected residual memory improves predictive accuracy over direct scalar residual feedback, supporting the benefit of a learned nonlinear representation of prediction deviations.

Residual connections are foundational to modern neural networks, yet the standard additive update provides no explicit mechanism for structured inter-channel interaction. We introduce KITINet (KInetics Theory Inspired Network), a training-time module with no additional trainable parameters that augments residual updates with stochastic pairwise feature interactions inspired by collision sampling in kinetic theory. KITINet reshapes channel groups as particles and mixes their features according to relative distance and velocity, while leaving the underlying residual layer trainable in the usual way. The module is disabled at inference, so the deployed model is identical to the original backbone and incurs no additional inference cost. Across language pre-training and fine-tuning, image classification, and PDE operator learning, KITINet demonstrates broad and robust performance gains across diverse tasks and architectures. We further observe enhanced parameter condensation across several settings and develop a simplified stochastic analysis that provides a mechanistic account of how collision-inspired interactions can promote condensation dynamics.

Scaling factors in residual branches have emerged as a prevalent method for boosting neural network performance, especially in normalization-free architectures. While prior work has primarily examined scaling effects from an optimization perspective, this paper investigates their role in residual architectures through the lens of generalization theory. Specifically, we establish that wide residual networks (ResNets) with constant scaling factors become asymptotically unlearnable as depth increases. In contrast, when the scaling factor exhibits rapid depth-wise decay combined with early stopping, over-parameterized ResNets achieve minimax-optimal generalization rates. To establish this, we demonstrate that the generalization capability of wide ResNets can be approximated by kernel regression associated with the Neural Tangent Kernel (NTK). Our theoretical findings are validated through experiments on synthetic data and real-world classification tasks, including MNIST and CIFAR-100.

Recent studies revealed the mathematical connection between deep neural networks (DNNs) and dynamic systems. However, the specific dynamics that DNNs, especially deep residual networks (ResNets), tend to learn during training remain insufficiently characterized. To this end, we model the forward propagation of deep residual networks using continuity equations, in which the measure is conserved and infinite curves in the measure space connect the input distribution to the output one of a ResNet. We find ResNets with $L_2$ regularization attempt to learn the geodesic curve in the Wasserstein space, induced by the optimal transport map. Compared with plain networks, ResNets can better approximate the geodesic curve, which explains why ResNets can be optimized and generalize better. Numerical experiments show that the data tracks of a ResNet tend to be line-shaped in terms of the line-shape score, and the map learned by a ResNet is closer to the optimal transport map in terms of the optimal transport score. In a word, we conclude that ResNets learn the geodesic curve in the Wasserstein space and discretely engineer the data transformation in high-dimensional spaces.

Looped (weight-tied) Transformers apply a shared residual block $N$ times ($h \leftarrow h + \varepsilon\,f(h)$, same $f$ at each step), increasing effective depth without adding parameters. Prior depth-scaling analyses prescribe $\varepsilon = 1/\!\sqrt{L}$ for depth-$L$ residual networks. We show that this is insufficient for looped architectures: weight sharing makes residual updates correlated across iterations, requiring the stronger scaling $\varepsilon = 1/N$. For multi-layer blocks ($L$ unique layers looped $N$ times), we derive a factored parameterization $\varepsilon = \lambda/(N\!\sqrt{L})$ that separates the two sources of growth: $1/N$ controls the within-layer loop correlation, and $1/\!\sqrt{L}$ controls the across-layer variance. A key consequence is that the optimal learning rate depends only on the number of unique layers $L$, not on the loop count $N$, enabling direct hyperparameter transfer from small to large $N$ without retuning. Experiments on looped Transformers confirm that $1/N$ scaling improves trainability and yields better loss than $1/\!\sqrt{N}$ scaling across loop counts.