In addition to our main results presented in Section 4 of the paper, we also performed various exploratory experiments to investigate further application cases of the METASIN activations. The experiments cover image classification where we show favorable results of using convolutional METASIN networks over baseline RELU networks, as well as various overfitting experiments to explore the use of METASIN activations with MLPs. In all MLP experiments, we use METASIN with K = 10 sine components, and distribute the frequencies evenly across the range [1,35]. The initialization of the remaining parameters follows the description provided in Section 3. Bicycle Figure 5: Visualization of selected reconstructed frames of the video. To fully appreciate the details and visual cues presented in the figure, we recommend visualizing the figures in color and zooming in for a more comprehensive analysis.

Weight-Decomposed Low-Rank Adaptation (DoRA) extends LoRA by decoupling weight magnitude from direction, but its forward pass requires the row-wise norm of W + sBA, a computation that every major framework we surveyed implements by materializing the dense [d_out, d_in] product BA. At d_in = 8192 and rank r = 384, a single module's norm requires about 512 MB of transient working memory in bf16, making high-rank DoRA costly and often infeasible on common single-GPU setups once hundreds of adapted modules and checkpointing are involved. We present two systems contributions. A factored norm decomposes the squared norm into base, cross, and Gram terms computable through O(d_out r + r^2) intermediates, eliminating the dense product. Fused Triton kernels collapse the four-kernel DoRA composition into a single pass, reducing memory traffic by about 4x and using a numerically stable form that avoids catastrophic cancellation in the near-unity rescaling regime where magnitude scales concentrate in practice. Across six 8-32B vision-language models (VLMs) on three NVIDIA GPUs (RTX 6000 PRO, H200, B200) at r = 384 in bf16, the fused implementation is 1.5-2.0x faster than Hugging Face PEFT's DoRA implementation for inference and 1.5-1.9x faster for gradient computation (optimizer step excluded), with up to 7 GB lower peak VRAM. Microbenchmarks on six GPUs spanning four architecture generations (L40S, A100, RTX 6000 PRO, H200, B200, B300) confirm 1.5-2.7x compose-kernel speedup. Final-logit cosine similarity exceeds 0.9999 across all model/GPU pairs, and multi-seed training curves match within 7.1 x 10^-4 mean per-step loss delta over 2000 steps.

Mobile devices have become essential enablers for AI applications, particularly in scenarios that require real-time performance. Vision Transformer (ViT) has become a fundamental cornerstone in this regard due to its high accuracy. Recent efforts have been dedicated to developing various transformer architectures that offer improved accuracy while reducing the computational requirements. However, existing research primarily focuses on reducing the theoretical computational complexity through methods such as local attention and model pruning, rather than considering realistic performance on mobile hardware. Although these optimizations reduce computational demands, they either introduce additional overheads related to data transformation (e.g., Reshape and Transpose) or irregular computation/data-access patterns.

The large communication and computation overhead of federated learning (FL) is one of the main challenges facing its practical deployment over resource-constrained clients and systems. In this work, SpaFL: a communication-efficient FL framework is proposed to optimize sparse model structures with low computational overhead. In SpaFL, a trainable threshold is defined for each filter/neuron to prune its all connected parameters, thereby leading to structured sparsity. To optimize the pruning process itself, only thresholds are communicated between a server and clients instead of parameters, thereby learning how to prune. Further, global thresholds are used to update model parameters by extracting aggregated parameter importance. The generalization bound of SpaFL is also derived, thereby proving key insights on the relation between sparsity and performance. Experimental results show that SpaFL improves accuracy while requiring much less communication and computing resources compared to sparse baselines. The code is available at https://github.com/news-vt/SpaFL

In this paper, we present a simple optimization-based preprocessing technique called Weight Magnitude Reduction (MagR) to improve the performance of post-training quantization. For each linear layer, we adjust the pre-trained floating-point weights by solving an $\ell_\infty$-regularized optimization problem. This process greatly diminishes the maximum magnitude of the weights and smooths out outliers, while preserving the layer's output. The preprocessed weights are centered more towards zero, which facilitates the subsequent quantization process. To implement MagR, we address the $\ell_\infty$-regularization by employing an efficient proximal gradient descent algorithm. Unlike existing preprocessing methods that involve linear transformations and subsequent post-processing steps, which can introduce significant overhead at inference time, MagR functions as a non-linear transformation, eliminating the need for any additional post-processing. This ensures that MagR introduces no overhead whatsoever during inference. Our experiments demonstrate that MagR achieves state-of-the-art performance on the Llama family of models. For example, we achieve a Wikitext2 perplexity of 6.7 on the LLaMA2-70B model for per-channel INT2 weight quantization without incurring any inference overhead.

KV cache stores key and value states from previous tokens to avoid re-computation, yet it demands substantial storage space, especially for long sequences. Adaptive KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance.

Deploying trustworthy AI systems requires principled uncertainty quantification. Conformal prediction (CP) is a widely used framework for constructing prediction sets with distribution-free coverage guarantees. In many practical settings, including healthcare, finance, and mobile sensing, the calibration data required for CP are distributed across multiple clients, each with its own local data distribution. In this federated setting, data can often be partitioned into, potentially overlapping, groups, which may reflect client-specific strata or cross-cutting attributes such as demographic or semantic categories. We propose group-conditional federated conformal prediction (GC-FCP), a novel protocol that provides group-conditional coverage guarantees. GC-FCP constructs mergeable, group-stratified coresets from local calibration scores, enabling clients to communicate compact weighted summaries that support efficient aggregation and calibration at the server. Experiments on synthetic and real-world datasets validate the performance of GC-FCP compared to centralized calibration baselines.

Recently, several works have shown that natural modifications of the classical conditional gradient method (aka Frank-Wolfe algorithm) for constrained convex optimization, provably converge with a linear rate when the feasible set is a polytope, and the objective is smooth and strongly-convex. However, all of these results suffer from two significant shortcomings: i) large memory requirement due to the need to store an explicit convex decomposition of the current iterate, and as a consequence, large running-time overhead per iteration ii) the worst case convergence rate depends unfavorably on the dimension In this work we present a new conditional gradient variant and a corresponding analysis that improves on both of the above shortcomings. In particular, both memory and computation overheads are only linear in the dimension, and in addition, in case the optimal solution is sparse, the new convergence rate replaces a factor which is at least linear in the dimension in previous works, with a linear dependence on the number of non-zeros in the optimal solution At the heart of our method, and corresponding analysis, is a novel way to compute decomposition-invariant away-steps. While our theoretical guarantees do not apply to any polytope, they apply to several important structured polytopes that capture central concepts such as paths in graphs, perfect matchings in bipartite graphs, marginal distributions that arise in structured prediction tasks, and more. Our theoretical findings are complemented by empirical evidence that shows that our method delivers state-of-the-art performance.

We study the collaborative PAC learning problem recently proposed in Blum et al.~\cite{BHPQ17}, in which we have $k$ players and they want to learn a target function collaboratively, such that the learned function approximates the target function well on all players' distributions simultaneously. The quality of the collaborative learning algorithm is measured by the ratio between the sample complexity of the algorithm and that of the learning algorithm for a single distribution (called the overhead). We obtain a collaborative learning algorithm with overhead $O(\ln k)$, improving the one with overhead $O(\ln^2 k)$ in \cite{BHPQ17}. We also show that an $\Omega(\ln k)$ overhead is inevitable when $k$ is polynomial bounded by the VC dimension of the hypothesis class. Finally, our experimental study has demonstrated the superiority of our algorithm compared with the one in Blum et al.~\cite{BHPQ17} on real-world datasets.