feature-wise modulation

Non-differentiable operations in ultra-low-bit quantization (e.g., 1-bit) and highly sparse pruning (>99% sparsity) severely destabilize backpropagation, causing conventional methods like the Straight-Through Estimator (STE) to fail catastrophically during training. Method: We propose a perturbation modeling paradigm that unifies quantization and pruning as structured perturbations injected during training, and introduce a differentiable denoising affine transformation for gradient approximation. Further, we design a robust ridge regression–based training framework featuring a piecewise-constant backbone network to guarantee performance lower bounds, coupled with an adaptive noise suppression mechanism enabling end-to-end optimization across arbitrary bit-widths and sparsity levels. Contribution/Results: Our approach achieves, for the first time, stable convergence under simultaneous 1-bit quantization and >99% sparsity—demonstrating unprecedented training stability. It significantly narrows the gap between artificial neural networks and biological neurons in terms of dynamic binary learning mechanisms.

Low-bit post-training quantization (PTQ) often incurs substantial accuracy degradation, exacerbated by conventional uniform affine transformations. To address this, we propose Clustering-based Affine Transformation (CAT), a method that learns cluster-specific affine parameters for distinct output clusters—thereby aligning quantized and full-precision output distributions with near-zero parameter overhead. CAT operates as a plug-and-play module requiring no fine-tuning or retraining, enabling seamless integration into existing PTQ pipelines. On ImageNet-1K, CAT achieves 53.18% Top-1 accuracy for W2A2 ResNet-18—surpassing the state-of-the-art by over 3%. It demonstrates consistent robustness across diverse architectures and quantization configurations. The core innovation lies in coupling clustering analysis with cluster-level affine calibration, effectively mitigating distribution mismatch—a critical challenge in low-bit PTQ.

This work addresses the limited interpretability and controllability of image generation models by proposing an internal mechanism intervention method based on parameterized activation functions. Specifically, we replace standard activations (e.g., ReLU) in mainstream generative architectures—such as StyleGAN2 and BigGAN—with learnable, semantically interpretable parameterized variants (e.g., generalized Swish with shape and bias controls). This enables direct, fine-grained manipulation of activation behavior for targeted image editing, without altering network architecture or requiring additional training. We demonstrate effective, attribute-specific control—including illumination, texture, and pose—on FFHQ and ImageNet. Experimental results confirm that our intervention preserves model fidelity while offering both human-understandable semantics and quantitative effectiveness. The approach establishes a novel paradigm for transparent, plug-and-play control over generative models’ internal representations.

Deep neural network model fusion is hindered by linear mode connectivity (LMC) barriers arising from weight-space solution dispersion, primarily due to inconsistent neuron permutations across distinct training configurations. To address this, we propose Training-time Neuron Alignment (TNA), a mechanism that enhances fusion performance without increasing inference overhead. TNA introduces permutation subspaces into the training phase—enabling lossless neuron alignment for the first time—and instantiates the TNA-PFN algorithm, which we theoretically prove reduces LMC barriers and supports federated fusion under heterogeneous data. Experiments demonstrate substantial improvements: enhanced generalization of Vision Transformers (ViTs) in Model Soup, superior performance of large language models (LLMs) in ColD fusion, and high-accuracy, low-communication wide-model fusion in federated learning settings.

To address the limited representational capacity of implicit neural representations (INRs), this paper proposes the Reversible Symmetric Power (RSP) transformation, which jointly enforces target-range constraints and enhances symmetry through data redistribution—without requiring additional storage. We introduce the novel “Range-Defined Symmetric Hypothesis” and design a bias-aware calibration module alongside an adaptive soft-boundary mechanism to preserve output continuity while improving robustness. Extensive experiments on 1D audio, 2D image, and 3D video fitting tasks demonstrate that RSP consistently achieves higher PSNR and SSIM compared to baselines—including random permutation and standard normalization—across diverse modalities. The method exhibits strong generalizability and stability, validating both its theoretical foundation and practical efficacy in enhancing INR performance.

Existing generative models lack a unified theoretical framework for concept manipulation and behavioral alignment in intermediate representations. This work proposes MidSteer, a general-purpose steering method based on affine transformations, which unifies concept erasure (e.g., LEACE) and concept switching under a closed-form affine intervention theory for the first time. MidSteer enables minimal-perturbation, targeted modifications to intermediate layer activations while circumventing restrictive assumptions inherent in prior approaches. The method is applicable across both diffusion models and large language models, demonstrating consistent superiority over existing techniques across diverse tasks, modalities, and architectures, thereby validating its effectiveness and broad generality.

Existing methods struggle to precisely characterize the continuous piecewise affine regions induced by practical neural networks over bounded input domains. This work proposes AffineLens, a novel framework that, for the first time, supports modern architectures incorporating convolutional layers, batch normalization, and residual connections. By iteratively computing intersections between neuron-induced hyperplanes and input polytopes layer by layer, AffineLens exactly enumerates all non-empty maximal affine subregions. The approach integrates polyhedral operations, hyperplane arrangements, and inter-layer propagation algorithms to enable verifiable visualization of region partitions and decision boundaries. Furthermore, it quantifies regional complexity across different architectures, offering insights into how network design choices influence the geometric structure of the learned functions.

This work investigates how to reconstruct input images from neural network outputs to uncover the features underlying model decisions. To this end, two novel inversion methods are proposed: a forward inversion approach leveraging the input Jacobian matrix combined with root-finding algorithms, and a backward inversion technique that iteratively inverts layer-by-layer while injecting random vectors into the nullspace of each layer’s linear transformation. For the first time, high-fidelity input reconstructions are achieved on Transformers and linear sequence networks. The generated images, though appearing random, consistently yield near-100% classification confidence and densely span the feasible input space. This approach substantially outperforms existing methods and effectively exposes the model’s reliance on non-semantic features and its inherent vulnerabilities.

Existing parameter reparameterization methods are often confined to a single objective—either parameter-efficient fine-tuning or model compression—making it challenging to simultaneously address both demands under resource constraints. This work proposes CRISP, a unified framework that jointly achieves model compression and parameter-efficient fine-tuning within a single architecture. CRISP decomposes pre-trained weights into shared base matrices and lightweight mixture coefficients, enhanced by cross-layer base sharing and an interpolation-based gated coefficient recombination mechanism. Requiring fewer than 200 trainable parameters, CRISP outperforms existing approaches by 1% in joint compression and fine-tuning tasks, surpasses state-of-the-art methods by up to 1.5% in pure parameter-efficient fine-tuning, and achieves a consistent 4–5% improvement in overall dual-task performance.

This work addresses the trade-off between background preservation and content generation in flow-matching-based image editing, where feature injection often leads to mutual suppression of these objectives. To resolve this, the authors propose a training-free, adaptive editing framework that dynamically modulates feature fusion and editing strength across both temporal and channel dimensions. Specifically, a progressive injection scheduling mechanism employs continuous decay functions—such as sigmoid or cosine—to gradually attenuate source feature injection over time, while a channel-selective latent perturbation strategy estimates channel importance based on latent distribution discrepancies and applies differentiated perturbations accordingly. The framework is compatible with various ODE solvers and achieves improved performance on PIE-Bench, reducing LPIPS by 8.7% and increasing SSIM and PSNR by 2.6% and 2.3%, respectively, all while maintaining high CLIP similarity, thereby enabling plug-and-play applicability.