Neural Network Compression

This paper considers neural channel state information (CSI) compression for time-varying massive multiple-input multiple-output (MIMO) channels in frequency division duplex (FDD) systems with limited feedback resources. The main challenge lies in obtaining a compact and efficient representation of the CSI given that it exhibits strong temporal correlation across successive snapshots. Existing memoryless compression models do not exploit this property, while simple temporal extensions often incorporate multiple observations without explicitly modeling the latent dynamics. We propose a context-aware compression framework based on a k-memory Markov variational autoencoder (k-MMVAE), which uses a finite temporal window to capture the evolution of CSI in the latent space. The model introduces Markov-structured latent dynamics with finite memory, enabling efficient use of temporal dependencies for compression. Simulation results show that the proposed approach improves target CSI reconstruction performance compared to memoryless and weakly sequential baselines, particularly at low and moderate compression rates. These results suggest that explicit latent temporal modeling can provide an effective mechanism for CSI compression under limited feedback constraints.

Pretraining language models with extended context windows enhances their ability to leverage rich information during generation. Existing methods split input sequences into chunks, broadcast them across multiple devices, and compute attention block by block which incurs significant communication overhead. While feasible in high-speed clusters, these methods are impractical for decentralized training over low-bandwidth connections. We propose a compression method for communication-efficient context parallelism in decentralized settings, achieving a remarkable compression rate of over 95% with negligible overhead and no loss in convergence. Our key insight is to exploit the intrinsic low-rank structure of activation outputs by dynamically constraining them to learned mixtures of subspaces via efficient reparameterizations. We demonstrate scaling billion-parameter decentralized models to context lengths exceeding 100K tokens on networks as slow as 300Mbps, matching the wall-clock convergence speed of centralized models on 100Gbps interconnects.

Embedded devices from wildlife monitoring stations to clinical wearables require local AI inference due to latency, communication, or privacy constraints. Optimizing models for heterogeneous microcontrollers (MCUs) requires simultaneously satisfying hard physical constraints on memory, power, and temperature while preserving accuracy, a multidimensional optimization that is today performed manually by experts. We ask whether an LLM agent can autonomously navigate this complex, multi-turn pipeline guided by real hardware feedback, and introduce a hardware-in-the-loop agent arena in which the agent iteratively refines both model and firmware -- compiling, flashing, and measuring on real hardware -- to enable closed-loop optimization. Frontier models, including Claude Opus 4.7 and Gemini 3.1 Pro, fail entirely without hardware feedback (0% deployment success), whereas our hardware-in-the-loop formulation achieves the first successful deployment within three iterations and can surpass human expert results within seven. This agentic co-optimization achieves 250x compression for vision models with <3.3% accuracy loss and 400x for audio with <6% Feature Error Rate loss, enabling battery-free operation on a commercial MCU via solar harvesting. We demonstrate practical impact in two real-world systems: an elk-detection camera trap (96.7% accuracy) and a phonetic-transcription wearable (8.44% FER) for child development research.

Spiking Vision Transformers (SViTs) have emerged as alternative low-power ViT models, but their large sizes hinder their deployments on resource-constrained embedded AI systems. To address this, state-of-the-art works proposed quantization techniques to compress SViT models, but their manual, human-guided approach needs a huge design time and power/energy consumption to find the appropriate quantization setting for each given network, making this approach not scalable for quantizing multiple networks. Toward this, we propose AQ4SViT, a novel automated quantization framework for SViTs that can provide quick quantization settings with good trade-offs between accuracy and memory. To achieve this, AQ4SViT employs the following key ideas: quantization search strategy that evaluates the quantization setting candidates while considering the accuracy constraint; and search gating policy that quickly evaluates and selects promising quantization candidates by leveraging membrane potential drift as a performance proxy. In the search gating policy, AQSViT employs two search algorithm variants to provide trade-off options: Greedy search, which performs fast but may lead to local optima; and Beam search, which performs slower but has better performance in finding global optima selection due to a wider search space. Experimental results show that AQ4SViT-Greedy quickly finds the appropriate quantization settings, achieving up to 6.6x faster search time and up to 82.5% memory saving compared to the state-of-the-art; while AQ4SViT-Beam further reduces the memory footprint by up to 90% compared to the state-of-the-art, but with 4.5x longer search time; all these results are obtained while maintaining high accuracy within 1.5% from the original/non-quantized models on the ImageNet dataset. These results highlight that AQ4SViT framework offers advancements toward SViT deployments on embedded AI systems.

In the present article, an improved Knowledge Distillation (KD) framework has been proposed for efficient compression of deep convolutional neural networks for land-use image classification task. Motivated by the need to achieve competitive classification accuracy while reducing computational complexity, a teacher-student learning paradigm is adopted in which a VGG16 network transfers knowledge to a lightweight MobileNetV2 model. The proposed framework integrates hard supervision from ground truth labels with a soft supervision strategy that combines Kullback-Leibler divergence and Cosine Similarity losses. Experiments conducted on three land-use datasets show that the proposed KD-based method yields improved performance, and achieves an accuracy of 99.04%, outperforming both baseline student training and single-loss distillation approaches, while retaining substantial model compression.

Spiking Neural Networks (SNNs) offer an energy-efficient paradigm for edge AI, making them attractive for hardware acceleration. However, deploying dense SNNs on FPGAs is constrained by limited on-chip memory for synaptic weight storage. To address this bottleneck, we propose VQ4SNN, a hardware-aware architecture that reduces memory requirements through Vector Quantization (VQ)-based weight sharing. To the best of our knowledge, this is the first application of VQ to pipelined spatial-dataflow SNN accelerators on FPGAs. VQ4SNN replaces conventional weight storage with a two-level memory organization consisting of compact pointers and a shared codebook of quantized weight vectors. The proposed design integrates FPGA-aware memory mapping with analytical VQ parameter selection, enabling efficient deployment on such accelerators while preserving inference accuracy. The experimental results show a reduction of 52-61% in the total number of BRAMs compared to the state-of-the-art uncompressed FPGA SNNs without increasing overall logic utilization.

Unstructured pruning produces sparse weight tensors, but the standard implementation keeps tensor shapes unchanged so the deployed model is no smaller than before pruning. We present an exact structural rewrite, which we call minimization, that converts a masked network into a smaller dense network with the same forward function up to floating-point rounding. The Squeeze-Release cycle iterates pruning and minimization with an intermediate release step that re-enables the exact-zero positions inside the compacted tensors as small calibrated noise, turning otherwise wasted capacity back into trainable parameters. Successive cycles use that capacity to find structural redundancy a single pass cannot reach. We additionally introduce CompensatedLayerNorm, a function-preserving replacement for LayerNorm that extends minimization to channel reduction across LayerNorm-equipped residual streams. Squeeze-Release compresses the deployable network to 39x smaller than the unpruned model on a fully-connected model network and 14.8x smaller on modern CNN (ConvNeXt-Tiny), at comparable accuracy. In addition we prove that the rewrite can be extended to transformer architectures.

Wearable healthcare devices are the fastest-growing Internet of Things (IoT) sector. Many automated healthcare services rely on two crucial biological signals, namely ECG and EEG, which reflect the activity of the heart and brain, respectively. Although deep neural networks are considered the primary way to process and analyze these signals, the very tight energy and computational power constraints in wearable devices are far below the computational, energy, and memory bandwidth demands of DNN models, thereby impeding the deployment of deep learning in many practical wearable services. This paper investigates the feasibility of deploying state-of-the-art DNN models in resource-constrained wearable devices. Notably, we explore the trade-off between accuracy and computational complexity of DNNs when parameter quantization and electrode reduction methods are used. Our investigation centers on several state-of-the-art DNN models designed for EEG signal analysis, specifically for detecting epileptic seizures. Our findings demonstrate that, when applied judiciously, these techniques can significantly reduce the complexity of the DNNs under consideration with minimal adverse effects on accuracy. These results reveal the explicit trade-offs between accuracy and complexity reduction encountered when adapting DNN-based online EEG analysis for wearable devices.

This paper presents a novel data-free and training-free compression approach for speech foundation models using channelwise clustering via k-means. More fine-grained, mixed sparsity pruning by layer-level varying number of parameter clusters is also explored. Experiments conducted on the LibriSpeech dataset suggest that when operating with pruning sparsity of 50% on HuBERT-large, consistent WER reductions of 27.73%/18.61% absolute (34.37%/21.91% relative) over the magnitude-based pruning were obtained on the test-clean and test-other subsets before fine-tuning and 0.19%/0.79% absolute (3.36%/4.62% relative) after fine-tuning with only 3 epochs. Similar WER reductions of 2.86%/5.02% absolute (59.21%/55.29% relative) were observed against magnitudebased pruning on Whisper-large-v3 at 10% sparsity, all with no significant WER increase relative to the uncompressed baseline.

Deep learning EEG denoising architectures have scaled from tens of thousands to tens of millions of parameters, yet no prior study has isolated model capacity as the experimental variable or tested whether reconstruction metrics predict downstream neural-signal utility. We address both gaps by fixing architecture, loss, data split, and training recipe while sweeping only channel width from 1.05K to 40.26K parameters in a minimal depthwise-separable convolutional U-Net. Models were evaluated on the EEGDenoiseNet benchmark, cross-dataset BCI transfer tests, controlled baseline retraining, and downstream motor-imagery classification with five decoder families across all nine BCI Competition IV-2a subjects. Reconstruction performance saturated by 3-6.5K parameters, with post-elbow gains of at most 0.015 correlation coefficient per log10-parameter unit. An 8.46M-parameter baseline retrained under the same pipeline matched the 40.26K compact variant on EOG--a 200x parameter gap yielding no advantage--while a Patch-Transformer control reproduced the same diminishing-return shape. Downstream evaluation exposed a classifier-dependent metric-utility gap: reconstruction-optimized denoising significantly degraded CSP+LDA classification across all nine subjects and three artifact types (best denoised accuracy 0.547 vs. 0.612 noisy baseline; Bonferroni p=0.0488), persisting on naturally recorded trials (Delta=-0.047; BH-FDR q=0.0049). End-to-end neural decoders showed variable or neutral effects. Standard EEG denoising benchmarks are saturated far below current model capacity, and reconstruction metrics do not predict BCI utility. Ultra-compact models at 33-46 KB and 1.27-2.61M FLOPs/segment are practical for edge deployment. These findings argue for capacity-controlled evaluation, harder task-aware benchmarks, and mandatory downstream validation.

Recently, the efficiency of Large Language Models (LLMs) deployment has become a critical concern in practical applications. While post-training quantization (PTQ) and structural pruning are established techniques for reducing memory footprint and inference latency, most existing PTQ approaches optimize quantization errors on a per-layer basis, overlooking how errors accumulate and propagate through the network, often resulting in suboptimal solutions. Traditional pipelines also tend to apply pruning and quantization in isolation or sequentially, further compounding sub-optimality. We introduce a novel end-to-end framework that addresses these limitations in two key ways. First, we propose a novel mixed-precision PTQ strategy that directly minimizes global error propagation across the entire model, rather than isolating layer-wise errors. Building on this, we develop a novel joint optimization approach that simultaneously learns structural pruning decisions and mixed-precision quantization policies within a unified search space. Extensive experiments show that, at ultra-low precisions (1-3 bits), our quantization method reduces WikiText perplexity by up to 21% compared to state-of-the-art (SoTA) weight-activation quantization baselines. Against leading weight-only quantization methods, it achieves up to 59% and 85% lower perplexity on WikiText and C4, respectively. Compared to the SoTA joint pruning-and-quantization techniques, our proposed method delivers superior perplexity and reasoning performance at ultra-low bits.

Pipeline parallelism enables training of large language models that exceed single-device memory, yet inter-stage activation communication becomes the dominant bottleneck when trained on low-bandwidth networks. Recent work in this area has proposed using fixed orthogonal projections to compress activations. However, this still results in a significant performance degradation and requires a number of non-standard adaptations to constrain the optimization. A natural alternative is to learn a low rank projection for each pipeline stage, however maintaining the necessary orthogonality of these projectors during training remains a challenge. We present Manifold Aware Projection Learning (MAPL), a method that treats inter-stage compression as a learnable orthogonal projection under explicit Stiefel manifold (orthogonal matrices) constraints. Rather than prescribing a fixed global subspace, MAPL lets each pipeline stage discover and continuously adapt its own task-optimal compression subspace via manifold-constrained steepest descent. To recover token-specific signals at stage boundaries, we introduce per-stage factorized anchor embeddings that allow for full-rank activation reconstruction with negligible communication overhead. We further show that we can incorporate residual vector quantization after projection with a streaming codebook synchronization protocol that amortizes dictionary communication. Across LLaMA models from 150M to 1B parameters we show that MAPL can be easily applied to the existing pipeline and can achieve high compression with neglibile performance degradation with a drastically improved tradeoffs in performance vs. compression compared to Subspace Networks.

This work presents a data-efficient variant of the Attention-Based Dual-Branch Complex Feature Fusion Network (CFFN) for hyperspectral image classification. The proposed model, termed DE-CFFN, retains the original two-stream structure: the Real-Valued Neural Network (RVNN) processes standard hyperspectral patches, while the Complex-Valued Neural Network (CVNN) handles their Fourier-transformed counterparts. The main contribution of this work lies in the feature extraction process and architectural enhancement. Factor Analysis is used for dimensionality reduction, offering improved latent feature representation over Principal Component Analysis. Additionally, both the RVNN and CVNN streams are structurally modified by successively halving the number of filters in the 3D convolutional layers to reduce complexity. The outputs of both branches are concatenated and passed through a Squeeze and Excitation (SE) block to enhance joint feature representation. Evaluated on the Pavia University and Salinas datasets, DE-CFFN achieves classification performance comparable to CFFN, while significantly reducing model size, memory consumption, and inference latency, making it suitable for real-time hyperspectral imaging applications.

Many researchers investigated neural networks with some of their weights fixed to values randomly drawn from a given distribution, e.g., $N(0, I)$. Our proposed HalfNet draws random weights from $N(0, Σ)$, where $Σ$, which defines the geometry of the distribution, has a low-rank factorization that we learn from data. Experiments on MNIST and CIFAR-10 demonstrate that HalfNet can match the performance of fully trained multilayer perceptrons while using substantially fewer parameters. Spectral analysis indicates that much of the predictive power of neural networks lies in the geometry of their weight space rather than in the precise values of individual parameters, and we observe that accuracy scales smoothly with rank. HalfNet is not a neural architecture trick for low-rank structure; it implements a data-dependent random embedding that can also be interpreted through supervised metric learning, or random-feature and kernel perspectives.

The large sizes of Spiking Vision Transformers (SViTs) still hinder their embedded implementation, highlighting the need for model compression. State-of-the-art works compress SViT models through unstructured pruning, which needs specialized hardware accelerators for their specific sparsity patterns to maximize efficiency gains. Moreover, their manual approach requires a huge design time to find an appropriate pruning setting for each network, thus making this approach not scalable. To address this limitation, we propose PrimeSVT, a novel framework that performs automated memory-aware structured pruning on pre-trained SViT models, thereby maximizing their efficiency gains during inference amenable to widely-used computing architectures. To achieve this, PrimeSVT first sorts the SViT layers based on their sizes (i.e., number of parameters), identifies the targeted pruning layers based on their robustness under different pruning rates, then leverages this order for compressing the model layer-by-layer sequentially from the largest one to the smallest one (i.e., so-called prioritized compression policy), while considering the user-defined constraints (i.e., acceptable accuracy and memory saving). In each layer, PrimeSVT employs channel-wise filter pruning based on their L2-norm values to structurally remove the non-significant weights. Experimental results show that PrimeSVT saves 26.68% memory through automated single-shot pruning, while preserving accuracy within 3% (70.3% without fine-tuning and 72.9% with fine-tuning) from the original unpruned SViT model (73.3%), thus meeting the accuracy and memory constraints. These show that our PrimeSVT framework enables design automation for SViTs and their embedded implementation.

Neural network compression is commonly achieved by pruning parameters based on local importance scores, e.g., magnitude-based pruning. We propose a complementary approach that compresses models by aggregating neurons with similar functional behavior rather than removing weights independently. Our method encodes a trained network as a polynomial ODE system and applies a lumping method called Approximate Forward Differential Equivalence to identify neurons with approximately matching induced dynamics. A single tolerance parameter, $\varepsilon$, controls the compression level and induces a smooth trade-off between model size and predictive accuracy. We evaluate the method on synthetic datasets derived from nonlinear dynamical systems with known ground-truth behavior and on public regression benchmarks. Across both settings, the proposed approach achieves substantial parameter reduction while preserving accuracy, and consistently compares favorably with magnitude-based pruning and Wanda at similar compression levels. These results suggest that differential equivalence-based aggregation is a principled and effective alternative to conventional weight-centric pruning.

Pruning is a process designed to reduce the number of weights in a large neural network. This can substantially speed up inference but might cause a considerable reduction in the model's accuracy, and thus it is usually followed by a healing process that regains some of the lost accuracy. In this paper, we propose a new healing method, STARFISH, that can recover (most of) the accuracy of any pruned network efficiently. The main idea of STARFISH is to optimize the pruned network to align with the original network's internal state representations using a tiny calibration set of unlabeled examples. For the common case of removing 50% of the weights, STARFISH healing improves the recovered accuracy by up to 22% over the state-of-the-art methods on ViT-based networks. Its advantage is even more pronounced under aggressive pruning. For example, after eliminating 75% of the weights in a DeiT-B network for ImageNet, STARFISH uses only 0.4% of the number of training images as a calibration set and recovers 82% of the original dense accuracy, whereas competing recovery techniques reach only 40% of the dense model accuracy.

Deep neural networks often contain redundant hidden units. Removing individual weights can reduce parameter count, but unstructured sparsity is not always easy to exploit in standard dense implementations. This paper develops a structured pruning framework in which complete neurons are removed using multi-armed bandit (MAB) algorithms. Each candidate neuron is treated as an arm; pulling an arm temporarily masks that neuron, measures the change in loss on a sampled mini-batch, restores the neuron, and updates an estimate of its safe-removal reward. The framework supports stochastic policies, including Epsilon-Greedy, Softmax, UCB1 and Thompson Sampling, and multiplicative-weight policies, including Hedge-style multiplicative weights and EXP3. We evaluate the method on tabular classification, tabular regression and deep neural-network benchmarks covering image, text and reasoning tasks. Statistical comparisons using the Friedman test followed by the Nemenyi post-hoc test show significant differences between methods. On tabular classification tasks, UCB1 obtains the highest mean rank among pruning policies and improves on the unpruned neural network. On regression tasks, UCB1 obtains the highest mean rank and is statistically competitive with, or superior to, several standard regression models according to R^2. On deep-learning tasks, UCB1 and Thompson Sampling obtain the strongest ranks, and several MAB policies significantly outperform the unpruned model, magnitude-based neuron pruning and greedy activation-variation pruning. The results show that MAB-based neuron pruning is an effective and computationally practical approach for structured model reduction.

Large deep neural networks are costly to store and deploy because inference must move and evaluate many parameters. This paper studies \emph{Automatically Differentiable Nonlinear Tensor Networks} (ADNTNs), compact differentiable weight generators for replacing selected dense, convolutional, and attention layers. An ADNTN maps a small set of trainable tensor cores to a full weight tensor through hierarchical contractions and learnable nonlinearities; the generated layer is then used as an ordinary linear or convolutional operator. We investigate three multilayered topologies: Tree Tensor Networks, augmented Tree Tensor Networks with boundary disentanglers, and MERA-style multi-scale decoders. Compared with flat brick-wall automatically differentiable tensor networks, these hierarchies provide logarithmic-depth communication between tensorised modes and optional lateral mixing, which can improve long-range structure without large increases in stored parameters. We give a unified forward--adjoint formulation showing how reverse-mode automatic differentiation computes pre-activation adjoints and contracted-environment gradients for all trainable cores. The formulation supports task losses, reconstruction losses, distillation, quantisation-aware terms, batching, and modern optimisers. Proof-of-concept experiments on selected AlexNet and VGG-16 layers on CIFAR-10 datasets achieve per-layer parameter-compression ratios from about $2{,}000\times$ to $430{,}000\times$. Several VGG-16 compressed models match or slightly exceed the dense baseline, whereas AlexNet shows moderate degradation under more restrictive redundancy. These results indicate that nonlinear tensor-network generators are a promising structured route to compact pattern-recognition models, while also showing that contraction schedules and hardware-aware implementations remain essential for practical speedups.

Sparse autoencoders (SAEs) have become a central tool for interpreting language models. However, two key SAE analyses that remain difficult to scale are (1) matching semantically similar features across multi-layers and (2) compressing large feature circuits into interpretable supernodes. Although these have been treated as separate problems, we show that both are instances of a more fundamental challenge, which we frame as the estimation of semantic distances between SAE features that lie on different activation manifolds. We introduce a distributional framework for this problem, in which each feature is represented not by a single decoder vector like in the literature, but by an activation-weighted distribution over the hidden states that express it. By projecting these distributions into a shared reference space and comparing them with Wasserstein distance, our method provides a unified semantic metric for cross-layer feature comparison. We prove that our representation is invariant to activation rescaling, stable under perturbations, and recovers true matches under finite-sample margin conditions. Empirically, our method outperforms decoder-vector and LLM-based baselines and captures subtle functional distinctions between related features. Notably, our method compresses large feature circuits into interpretable supernodes automatically.

Image semantic communication is a critical component in next-generation wireless communication systems. However, such systems typically suffer from large memory footprints and high computational complexity, making them difficult to deploy on resource-constrained devices. To address these challenges, we propose a vision transformer (ViT)-enabled image semantic communication system. In this system, a recursive structure is introduced to iteratively refine semantic features and reduce the parameter count. In addition, three dynamic adjustment strategies are designed to adaptively reduce computational complexity: dynamic depth adjustment, dynamic width adjustment, and joint width-depth optimization. Dynamic depth adjustment adaptively determines the number of recursive modules according to image content and channel conditions, while dynamic width adjustment selectively preserves important neurons and attention heads. The joint width-depth optimization further enables flexible computation configurations. Simulation results verify that the proposed recursive ViT-based system, combined with the three dynamic adjustment strategies, reduces the parameter count by 48.7% and achieves higher reconstruction quality than existing baselines under comparable computational complexity.

Efficient onboard multi-field sparse reconstruction is essential for the autonomous operation of aerospace vehicles. While existing deep learning models exhibit promise for single-field reconstruction, deploying multiple independent models leads to prohibitive model size growth and fails to exploit cross-field correlations, particularly under few-shot conditions. To address these challenges, we first propose a lightweight multi-task Fourier neural operator (MTL-FNO), an end-to-end joint training framework based on hard parameter sharing. In each layer, the parameters are divided into shared and task-specific components to capture common features across fields while preserving task-specific characteristics. Moreover, the task-specific fine-tuning parameters are implemented as low-rank terms, achieving substantial model compression. Second, to address the difficulty of co-optimizing shared and task-specific parameters along with their real and imaginary parts, we revisit the FNO's spectral weight from a polar-form perspective and devise a physically meaningful decoupled optimization scheme. Specifically, we apply polar decomposition to slice-wise disentangle the spectral weight into a unitary tensor encoding phase information and a positive semi-definite tensor characterizing amplitude. By decoupling the optimization of phase and amplitude, our method can effectively mitigate tasks conflict. Meanwhile, to preserve unitary geometric fidelity during training, the Cayley transform is introduced to reparameterize the unitary tensor, converting the constrained optimization problem to an unconstrained one. Finally, the effectiveness of the proposed method under few-shot conditions is validated on two representative engineering cases. Results show that MTL-FNO achieves accuracy comparable to or even surpassing that of standard FNO, while reducing total model size by 76% and 60%, respectively.

At very high sparsity, neural network pruning does more than decide which weights remain. It also determines where pruning induced damage is placed across the network, and whether that damage can be recovered by a fixed lightweight repair procedure. We study this problem through the lens of repair conditioned sparsity allocation. We introduce Relative Repairability (RR), a calibration based diagnostic that compares the raw activation distortion caused by layerwise pruning with the residual distortion left after channelwise variance matching repair. RR estimates the fraction of local damage that remains after repair, using only unlabeled calibration data. Across ResNet18, ResNet34, and VGG16 BN on CIFAR10 and CIFAR100, we find that RR is not a universally dominant allocation rule. Instead, it is most useful near an architecture dependent recoverability transition, where standard structural or magnitude based allocation priors begin to lose reliability but post repair recovery has not yet fully collapsed. On CIFAR100 ResNet18, a fine grained sweep shows that RR improves over ERK across the central transition band and surpasses LAMP near the upper part of this band. A projection forced ablation further shows that capped ERK can over protect projection layers, shifting excessive sparsity onto regular convolutions and reducing post repair recovery. These results suggest that high sparsity pruning should allocate not only retained weights, but also repairable damage.

Large language models (LLMs) are dominated by dense linear transformations, whose storage, memory and computational overheads hinder efficient adaptation and deployment while masking the functional impacts of structural simplification. Here we present Tensor Mixture (MixT), a general tensor-structured compression scheme that replaces targeted dense linear layers with natively executable mixtures of tensor operators. Operating directly on generic linear projections instead of model-specific components, MixT is potentially applicable across Transformer-based LLMs and other dense neural mappings. We evaluate MixT on Qwen3-8B and LLaMA2-7B under a unified recovery protocol, identifying a broad compressible regime in which MMLU accuracy is largely preserved before an abrupt transition at model-specific boundaries. This transition coincides with coordinated shifts in output entropy, prediction entropy and inter-layer geometry. At the LLaMA2-7B transition boundary, MixT reduces full-model parameters by 47.5%, inference FLOPs by 37.1%, training FLOPs by 52.1% and peak inference memory by 60.4%, demonstrating its practical potential for lower-cost LLM compression.

Neural network weights are increasingly a bottleneck for deployment, yet most compression pipelines treat layers independently and overlook cross-layer redundancy induced by function-preserving symmetries. We propose Motion-Compensated Weight Compression (MCWC), a weight-only codec that aligns permutation-symmetric blocks (e.g., hidden units and attention heads) to maximize cross-layer correspondence, turning depth into a predictable sequence. In the aligned coordinate system, MCWC uses a lightweight layer-sequential predictor with periodic keyframes and encodes only quantized prediction residuals using a learned entropy model trained under a rate distortion objective. A simple decoder reconstructs deployable weights by entropy decoding, dequantization, predictor-driven reconstruction, and inverse alignment, enabling fast weight materialization for inference. Across Transformer language modeling and vision classification, MCWC improves the rate accuracy Pareto frontier over strong quantization and learned weight-codec baselines, while maintaining competitive decode time. Ablations confirm that alignment, prediction, entropy modeling, and keyframe scheduling are each necessary for the full gains. Our code is available via https://github.com/Ism-ail11/MCWC.

Deploying vision models across devices with varying resource constraints, or even on a single device where available compute fluctuates due to battery state, thermal throttling, or latency deadlines, typically requires training and maintaining separate models. Width-adaptive inference addresses this by training a single set of shared weights containing multiple nested subnetworks of increasing capacity, but prior CNN-based approaches required switchable batch normalization, while recent scalable methods have focused on Vision Transformers. We present Slimmable ConvNeXt, which shows that ConvNeXt's modern design, specifically LayerNorm and inverted bottlenecks, makes it particularly suited for channel-width slimming, eliminating the normalization overhead of classical slimmable networks and producing a simpler training pipeline than both prior CNN and ViT approaches. On ImageNet-1k, Slimmable ConvNeXt-T with 3 subnetworks achieves 80.8% top-1 accuracy at 4.5 GMACs and 77.4% at 1.2 GMACs, trained from scratch for 600 epochs. At comparable compute, this exceeds HydraViT's 6-head subnetwork (78.4% at 4.6 GMACs) by 2.4 percentage points and its 3-head configuration (73.0% at 1.3 GMACs) by 4.4 percentage points, while also outperforming MatFormer-S (78.6%) and SortedNet-S (78.2%) at the same GMACs. Scaling to Slimmable ConvNeXt-B further improves maximum accuracy to 82.8% at 15.35 GMACs.

Ensembles of neural networks typically outperform individual networks but incur large computational costs, whereas weight aggregation produces less costly, yet also less accurate, aggregate models. We introduce partial fusion of networks, which interpolates between ensembles and weight aggregation and thus allows for a flexible tradeoff between computational cost and performance. A direct way to achieve this is to extend existing weight aggregation methods based on neuron-level similarity between different networks, where partial fusion then only aggregates weights of neurons which are most similar. We showcase one particular method to jointly identify which neurons are most similar and match them via partial optimal transport. Further, we consider the more general perspective of weight aggregation and partial fusion as generalized pruning of ensemble models, where neurons cannot just be deleted, but also linearly combined. Finally, we show that generalized pruning applied to a single network yields similar benefits as partial fusion by allowing for a tradeoff between isolating, deleting, and linearly combining neurons based on similarity. Our code is available at https://github.com/Fabian-Mor/partial_fusion_nn.

One-shot magnitude pruning can cause severe accuracy collapse in the high-sparsity regime, even when the pruning mask preserves the largest weights. We argue that this failure reflects a granularity mismatch in post-pruning repair. Under global magnitude pruning, nearly collapsed channels can coexist with channels that retain informative activation variance within the same layer. Existing layer-wise activation repair methods apply a single correction to the whole layer, and can therefore over-amplify damaged channels while trying to restore the layer-level signal. We propose Adaptive Signal Resuscitation (ASR), a training-free channel-wise repair method that matches the granularity of repair to the granularity of damage. ASR estimates a variance-matching correction for each output channel and stabilizes it with a data-driven shrinkage rule, suppressing unreliable corrections for channels with weak post-pruning signal while preserving corrections for healthier channels. Applied before BatchNorm recalibration, ASR requires only forward passes on a small calibration set and no retraining. Across three datasets, four convolutional architectures, and both unstructured and structured sparsity settings, ASR generally improves over layer-wise repair, with the clearest gains in high-sparsity regimes. On ResNet-50 at 90% sparsity, ASR recovers 55.6% top-1 accuracy on CIFAR-10, compared with 41.0% for layer-wise repair and 28.0% for BatchNorm-only recalibration. Ablations show that naive channel-wise variance matching is insufficient, and that shrinkage stabilizes post-pruning repair.

We propose a scalable tensorization framework for neural network compression based on slice-wise feature distillation. Unlike conventional tensor decomposition methods that rely on costly global finetuning, our approach decomposes the network into slices consisting of either individual layers or blocks (e.g., convolutional layers or MLPs), or small groups of consecutive layers, and tensorizes each slice independently to reproduce the intermediate representations of the original pretrained model. This modular strategy improves accuracy recovery, reduces data requirements, and enables efficient parallel optimization. Experiments on ResNet-34 show significant gains over conventional global tensorization, achieving near-lossless compression at moderate compression rates with faster optimization. Results on GPT-2 XL further demonstrate the scalability of the method and its applicability to large-scale models, particularly in distributed settings.

End-to-end training with full-depth backpropagation remains the dominant paradigm for optimizing deep neural networks, but its efficiency deteriorates as models grow deeper. Since every block must be executed and differentiated under a single global objective, full-depth BP introduces substantial parameter redundancy, activation-memory cost, and training latency, especially when neighboring layers exhibit highly correlated learning patterns. Directly skipping or removing layers can reduce cost, but often weakens representation capacity or requires architecture-specific reuse designs. In this paper, we propose Replacement Learning (RepL), a training-time paradigm that reduces full-depth redundancy by replacing selected blocks rather than simply discarding them. For each removed block, RepL inserts a lightweight computing layer that synthesizes a surrogate operator from the parameters of its adjacent preceding and succeeding blocks through a learnable transformation, and applies the synthesized operator to the preceding activation. In this way, RepL preserves local contextual continuity while avoiding unnecessary full-layer computation. We instantiate RepL for CNNs and ViTs with tailored parameter-fusion blocks that handle convolutional channels, feature resolutions, and transformer submodules. Extensive experiments on CIFAR-10, SVHN, STL-10, ImageNet, COCO, and CityScapes show that RepL reduces trainable parameters, GPU memory usage, and training time while matching or surpassing standard end-to-end training across classification, detection, and segmentation. Additional results on WikiText-2, transfer learning, inference throughput, checkpointing, stochastic depth, and INT8 quantization further demonstrate its generality and compatibility.