arXiv subjects

Can Yaras

Publications and source records attributed to Can Yaras.

13 recordsLinked to original sources

arXiv2026

The Effect of Training Task Diversity on In-Context Learning through the Lens of Low-Dimensional Subspaces

The transformer's emergent ability to perform in-context learning (ICL) has sparked a wide range of studies designed to understand its underlying mechanisms. Existing works often study how training task diversity, defined either as the number of ICL training task vectors or as the number of function classes from which the task vectors are drawn, shapes both the learning dynamics and generalization capabilities of ICL. While both definitions have uncovered many interesting phenomena, many observations under the latter definition remain theoretically unexplained. This paper presents a minimal analytical model under which these phenomena provably emerge from the properties of the training data. By modeling the training task vectors as a mixture of low-rank Gaussians, we show how training task diversity, defined by the number of non-overlapping columns between subspaces that parameterize the covariance matrices, improves both the generalization and optimization trajectory of ICL with linear attention. In particular, we show that our model can explain (i) why training with task diversity shortens the ICL plateau and (ii) why ICL appears to achieve out-of-distribution generalization. We conclude by empirically demonstrating how our results extend to nonlinear transformers and nonlinear function classes. Overall, our work presents a tractable framework to unify existing observations.

stat.ML↗

arXiv2026

Stochastic Sparse Attention for Memory-Bound Inference

Autoregressive decoding becomes bandwidth-limited at long contexts, as generating each token requires reading all $n_k$ key and value vectors from KV cache. We present Stochastic Additive No-mulT Attention (SANTA), a method that sparsifies value-cache access by sampling $S \ll n_k$ indices from the post-softmax distribution and aggregates only those value rows. This yields an unbiased estimator of the post-softmax value aggregation while replacing value-stage multiply-accumulates with gather-and-add. We introduce stratified and systematic sampling to design variance-reduced, GPU-friendly variants. Evaluated on Llama-3.1-8B-Instruct at 32k-token contexts, S$^2$ANTA matches baseline accuracy while achieving up to $1.5\times$ decode-step attention-kernel speedup over FlashInfer and FlashDecoding on an NVIDIA RTX 6000 Ada. In batched long-context generation, these kernel gains translate to up to $1.25\times$ end-to-end decode-latency speedup. Finally, we propose Bernoulli $qK^\mathsf{T}$ sampling as a complementary technique to sparsify the score stage, reducing key-feature access through stochastic ternary queries. Both methods are complementary to upstream quantization, low-rank projection, KV-cache compression, and KV-cache selection methods. Together, they point toward sparse, multiplier-free, and energy-efficient inference. We open-source our kernels at: https://github.com/OPUSLab/SANTA.git

cs.LG↗

arXiv2026

Emergent Low-Rank Training Dynamics in MLPs with Smooth Activations

Recent empirical evidence has demonstrated that the training dynamics of large-scale deep neural networks occur within low-dimensional subspaces. While this has inspired new research into low-rank training, compression, and adaptation, theoretical justification for these dynamics in nonlinear networks remains limited. %compared to deep linear settings. To address this gap, this paper analyzes the learning dynamics of multi-layer perceptrons (MLPs) under gradient descent (GD). We demonstrate that the weight dynamics concentrate within invariant low-dimensional subspaces throughout training. Theoretically, we precisely characterize these invariant subspaces for two-layer networks with smooth nonlinear activations, providing insight into their emergence. Experimentally, we validate that this phenomenon extends beyond our theoretical assumptions. Leveraging these insights, we empirically show there exists a low-rank MLP parameterization that, when initialized within the appropriate subspaces, matches the classification performance of fully-parameterized counterparts on a variety of classification tasks.

cs.LG↗

arXiv2025

MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention

Transformers have achieved state-of-the-art performance across various tasks, but suffer from a notable quadratic complexity in sequence length due to the attention mechanism. In this work, we propose MonarchAttention -- a novel approach to sub-quadratic attention approximation via Monarch matrices, an expressive class of structured matrices. Based on the variational form of softmax, we describe an efficient optimization-based algorithm to compute an approximate projection of softmax attention onto the class of Monarch matrices with $\Theta(N\sqrt{N} d)$ computational complexity and $\Theta(Nd)$ memory/IO complexity. Unlike previous approaches, MonarchAttention is both (1) transferable, yielding minimal performance loss with no additional training, even when replacing every attention layer of the Transformer, and (2) hardware-efficient, utilizing the highest-throughput tensor core units on modern GPUs. With optimized kernels, MonarchAttention achieves substantial speed-ups in wall-time over FlashAttention-2: $1.4\times$ for shorter sequences $(N=256)$, $4.5\times$ for medium-length sequences $(N=4K)$, and $8.2\times$ for longer sequences $(N=16K)$. We demonstrate the quality of MonarchAttention on diverse tasks and architectures in vision and language problems, showing that it flexibly and accurately approximates softmax attention in a variety of contexts. Our code is available at https://github.com/cjyaras/monarch-attention.

cs.LG↗

arXiv2025

Out-of-Distribution Generalization of In-Context Learning: A Low-Dimensional Subspace Perspective

The transformer's remarkable ability to perform in-context learning (ICL) has sparked a wide range of studies designed to understand its strengths and limitations. However, a theoretical understanding of when ICL can and cannot generalize beyond its pre-training data still remains unclear. This paper puts forth a minimal mathematical model that provably identifies when ICL can generalize out-of-distribution (OOD). By studying linear regression tasks parameterized with low-rank covariance matrices, we model distribution shifts as varying angles between subspaces and derive conditions under which a single-layer linear attention model interpolates across all angles. We show that if pre-training task vectors are drawn from a union of subspaces, transformers can generalize to all angle shifts--enabling ICL even in regions with zero probability mass in the training distribution. On the other hand, if the pre-training tasks are drawn from a single Gaussian, the test risk shows a non-negligible dependence on the angle, implying that ICL cannot generalize OOD. We empirically show that our results also hold for models such as GPT-2, and present experiments on how our results extend to nonlinear function classes.

stat.ML↗

arXiv2025

An Overview of Low-Rank Structures in the Training and Adaptation of Large Models

The substantial computational demands of modern large-scale deep learning present significant challenges for efficient training and deployment. Recent research has revealed a widespread phenomenon wherein deep networks inherently learn low-rank structures in their weights and representations during training. This tutorial paper provides a comprehensive review of advances in identifying and exploiting these low-rank structures, bridging mathematical foundations with practical applications. We present two complementary theoretical perspectives on the emergence of low-rankness: viewing it through the optimization dynamics of gradient descent throughout training, and understanding it as a result of implicit regularization effects at convergence. Practically, these theoretical perspectives provide a foundation for understanding the success of techniques such as Low-Rank Adaptation (LoRA) in fine-tuning, inspire new parameter-efficient low-rank training strategies, and explain the effectiveness of masked training approaches like dropout and masked self-supervised learning.

cs.LG↗

arXiv2025

Linearly Separable Features in Shallow Nonlinear Networks: Width Scales Polynomially with Intrinsic Data Dimension

Deep neural networks have attained remarkable success across diverse classification tasks. Recent empirical studies have shown that deep networks learn features that are linearly separable across classes. However, these findings often lack rigorous justifications, even under relatively simple settings. In this work, we address this gap by examining the linear separation capabilities of shallow nonlinear networks. Specifically, inspired by the low intrinsic dimensionality of image data, we model inputs as a union of low-dimensional subspaces (UoS) and demonstrate that a single nonlinear layer can transform such data into linearly separable sets. Theoretically, we show that this transformation occurs with high probability when using random weights and quadratic activations. Notably, we prove this can be achieved when the network width scales polynomially with the intrinsic dimension of the data rather than the ambient dimension. Experimental results corroborate these theoretical findings and demonstrate that similar linear separation properties hold in practical scenarios beyond our analytical scope. This work bridges the gap between empirical observations and theoretical understanding of the separation capacity of nonlinear networks, offering deeper insights into model interpretability and generalization.

cs.LG↗

arXiv2024

Explaining and Mitigating the Modality Gap in Contrastive Multimodal Learning

Multimodal learning has recently gained significant popularity, demonstrating impressive performance across various zero-shot classification tasks and a range of perceptive and generative applications. Models such as Contrastive Language-Image Pretraining (CLIP) are designed to bridge different modalities, such as images and text, by learning a shared representation space through contrastive learning. Despite their success, the working mechanisms underlying multimodal learning are not yet well understood. Notably, these models often exhibit a modality gap, where different modalities occupy distinct regions within the shared representation space. In this work, we conduct an in-depth analysis of the emergence of modality gap by characterizing the gradient flow learning dynamics. Specifically, we identify the critical roles of mismatched data pairs and a learnable temperature parameter in causing and perpetuating the modality gap during training. Furthermore, our theoretical insights are validated through experiments on practical CLIP models. These findings provide principled guidance for mitigating the modality gap, including strategies such as appropriate temperature scheduling and modality swapping. Additionally, we demonstrate that closing the modality gap leads to improved performance on tasks such as image-text retrieval.

cs.LG↗

arXiv2024

Compressible Dynamics in Deep Overparameterized Low-Rank Learning & Adaptation

While overparameterization in machine learning models offers great benefits in terms of optimization and generalization, it also leads to increased computational requirements as model sizes grow. In this work, we show that by leveraging the inherent low-dimensional structures of data and compressible dynamics within the model parameters, we can reap the benefits of overparameterization without the computational burdens. In practice, we demonstrate the effectiveness of this approach for deep low-rank matrix completion as well as fine-tuning language models. Our approach is grounded in theoretical findings for deep overparameterized low-rank matrix recovery, where we show that the learning dynamics of each weight matrix are confined to an invariant low-dimensional subspace. Consequently, we can construct and train compact, highly compressed factorizations possessing the same benefits as their overparameterized counterparts. In the context of deep matrix completion, our technique substantially improves training efficiency while retaining the advantages of overparameterization. For language model fine-tuning, we propose a method called "Deep LoRA", which improves the existing low-rank adaptation (LoRA) technique, leading to reduced overfitting and a simplified hyperparameter setup, while maintaining comparable efficiency. We validate the effectiveness of Deep LoRA on natural language tasks, particularly when fine-tuning with limited data. Our code is available at https://github.com/cjyaras/deep-lora-transformers.

cs.LG↗

arXiv2023

Understanding Deep Representation Learning via Layerwise Feature Compression and Discrimination

Over the past decade, deep learning has proven to be a highly effective tool for learning meaningful features from raw data. However, it remains an open question how deep networks perform hierarchical feature learning across layers. In this work, we attempt to unveil this mystery by investigating the structures of intermediate features. Motivated by our empirical findings that linear layers mimic the roles of deep layers in nonlinear networks for feature learning, we explore how deep linear networks transform input data into output by investigating the output (i.e., features) of each layer after training in the context of multi-class classification problems. Toward this goal, we first define metrics to measure within-class compression and between-class discrimination of intermediate features, respectively. Through theoretical analysis of these two metrics, we show that the evolution of features follows a simple and quantitative pattern from shallow to deep layers when the input data is nearly orthogonal and the network weights are minimum-norm, balanced, and approximate low-rank: Each layer of the linear network progressively compresses within-class features at a geometric rate and discriminates between-class features at a linear rate with respect to the number of layers that data have passed through. To the best of our knowledge, this is the first quantitative characterization of feature evolution in hierarchical representations of deep linear networks. Empirically, our extensive experiments not only validate our theoretical results numerically but also reveal a similar pattern in deep nonlinear networks which aligns well with recent empirical studies. Moreover, we demonstrate the practical implications of our results in transfer learning. Our code is available at https://github.com/Heimine/PNC_DLN.

cs.LG↗

arXiv2023

The Law of Parsimony in Gradient Descent for Learning Deep Linear Networks

Over the past few years, an extensively studied phenomenon in training deep networks is the implicit bias of gradient descent towards parsimonious solutions. In this work, we investigate this phenomenon by narrowing our focus to deep linear networks. Through our analysis, we reveal a surprising "law of parsimony" in the learning dynamics when the data possesses low-dimensional structures. Specifically, we show that the evolution of gradient descent starting from orthogonal initialization only affects a minimal portion of singular vector spaces across all weight matrices. In other words, the learning process happens only within a small invariant subspace of each weight matrix, despite the fact that all weight parameters are updated throughout training. This simplicity in learning dynamics could have significant implications for both efficient training and a better understanding of deep networks. First, the analysis enables us to considerably improve training efficiency by taking advantage of the low-dimensional structure in learning dynamics. We can construct smaller, equivalent deep linear networks without sacrificing the benefits associated with the wider counterparts. Second, it allows us to better understand deep representation learning by elucidating the linear progressive separation and concentration of representations from shallow to deep layers. We also conduct numerical experiments to support our theoretical results. The code for our experiments can be found at https://github.com/cjyaras/lawofparsimony.

cs.LG↗

arXiv2023

Neural Collapse with Normalized Features: A Geometric Analysis over the Riemannian Manifold

When training overparameterized deep networks for classification tasks, it has been widely observed that the learned features exhibit a so-called "neural collapse" phenomenon. More specifically, for the output features of the penultimate layer, for each class the within-class features converge to their means, and the means of different classes exhibit a certain tight frame structure, which is also aligned with the last layer's classifier. As feature normalization in the last layer becomes a common practice in modern representation learning, in this work we theoretically justify the neural collapse phenomenon for normalized features. Based on an unconstrained feature model, we simplify the empirical loss function in a multi-class classification task into a nonconvex optimization problem over the Riemannian manifold by constraining all features and classifiers over the sphere. In this context, we analyze the nonconvex landscape of the Riemannian optimization problem over the product of spheres, showing a benign global landscape in the sense that the only global minimizers are the neural collapse solutions while all other critical points are strict saddles with negative curvature. Experimental results on practical deep networks corroborate our theory and demonstrate that better representations can be learned faster via feature normalization.

cs.LG↗

arXiv2021

Randomized Histogram Matching: A Simple Augmentation for Unsupervised Domain Adaptation in Overhead Imagery

Modern deep neural networks (DNNs) are highly accurate on many recognition tasks for overhead (e.g., satellite) imagery. However, visual domain shifts (e.g., statistical changes due to geography, sensor, or atmospheric conditions) remain a challenge, causing the accuracy of DNNs to degrade substantially and unpredictably when testing on new sets of imagery. In this work, we model domain shifts caused by variations in imaging hardware, lighting, and other conditions as non-linear pixel-wise transformations, and we perform a systematic study indicating that modern DNNs can become largely robust to these types of transformations, if provided with appropriate training data augmentation. In general, however, we do not know the transformation between two sets of imagery. To overcome this, we propose a fast real-time unsupervised training augmentation technique, termed randomized histogram matching (RHM). We conduct experiments with two large benchmark datasets for building segmentation and find that despite its simplicity, RHM consistently yields similar or superior performance compared to state-of-the-art unsupervised domain adaptation approaches, while being significantly simpler and more computationally efficient. RHM also offers substantially better performance than other comparably simple approaches that are widely used for overhead imagery.

cs.CV↗