Large Language Models (LLMs) have significantly advanced AI with their exceptional performance across a wide range of tasks. However, their extensive computational requirements restrict their use on devices with limited resources. While recent compression methods based on low-rank matrices show potential solutions, they often suffer from significant loss of accuracy or introduce substantial overhead in parameters and inference time. In this paper, we introduce Modular De- composition (MoDeGPT), a new, efficient, and structured compression framework that overcomes these limitations. MoDeGPT jointly decomposes pairs of consecu- tive subcomponents within Transformer blocks, reduces hidden dimensions through output reconstruction on a larger structural scale than conventional low-rank meth- ods, and repurposes three classical matrix decomposition algorithms—Nyström approximation, CR decomposition, and SVD—to ensure bounded errors in our novel decomposition approach. Our experiments show that MoDeGPT, without relying on backward propagation, consistently matches or surpasses the performance of prior techniques that depend on gradient information, while achieving a 98% reduction in compute costs when compressing a 13B-parameter model. On LLaMA-2/3 and OPT models, MoDeGPT retains 90-95% of zero-shot performance with compression rates of 25-30%. The compression process can be completed on a single GPU in a few hours, boosting inference throughput by up to 46%.

Sparse Mixture-of-Experts (SMoE) models represent a significant advancement in large language model (LLM) development through their efficient parameter utilization. These models achieve substantial performance improvements at reduced inference costs. However, the deployment of SMoE models faces constraints from extensive memory requirements of expert components in resource-limited environments. To address these limitations, this paper introduces Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), a task-agnostic expert merging framework for parameter reduction without retraining. HC-SMoE introduces a novel hierarchical clustering approach based on expert outputs to ensure merging robustness independent of routing decisions. The proposed output-based clustering method enables effective capture of functional relationships between experts for largescale architectures. We provide theoretical analysis and comprehensive evaluations across multiple zero-shot language tasks to demonstrate HC-SMoE’s effectiveness in state-of-the-art models including Qwen and Mixtral. The experimental results validate HC-SMoE’s superior performance and practical applicability for real-world deployments. Our implementation is available at https://github.com/wazenmai/HC-SMoE.

James Seale Smith, Chi-Heng Lin, Shikhar Tuli, Haris Jeelani, Shangqian Gao, Yilin Shen, Hongxia Jin, Yen-Chang Hsu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Shangqian Gao, Ting Hua, Yen-Chang Hsu, Yilin Shen, Hongxia Jin. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Shikhar Tuli, Chi-Heng Lin, Yen-Chang Hsu, Niraj Jha, Yilin Shen, Hongxia Jin. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Large Language Models (LLMs) have achieved remarkable success in various natural language processing tasks, including language modeling, understanding, and generation. However, the increased memory and computational costs associated with these models pose significant challenges for deployment on resource-limited devices. Structural pruning has emerged as a promising solution to reduce the costs of LLMs without requiring post-processing steps. Prior structural pruning methods either follow the dependence of structures at the cost of limiting flexibility, or introduce non-trivial additional parameters by incorporating different projection matrices. In this work, we propose a novel approach that relaxes the constraint imposed by regular structural pruning methods and eliminates the structural dependence along the embedding dimension. Our dimension-independent structural pruning method offers several benefits. Firstly, our method enables different blocks to utilize different subsets of the feature maps. Secondly, by removing structural dependence, we facilitate each block to possess varying widths along its input and output dimensions, thereby significantly enhancing the flexibility of structural pruning. We evaluate our method on various LLMs, including OPT, LLaMA, LLaMA-2, Phi-1.5, and Phi-2. Experimental results demonstrate that our approach outperforms other state-of-the-art methods, showing for the first time that structural pruning can achieve an accuracy similar to semi-structural pruning.

Vision Transformers (ViTs) have emerged as powerful backbones in computer vision, outperforming many traditional CNNs. However, their computational overhead, largely attributed to the self-attention mechanism, makes deployment on resource-constrained edge devices challenging. Multiple solutions rely on token pruning or token merging. In this paper, we introduce "Token Fusion" (ToFu), a method that amalgamates the benefits of both token pruning and token merging. Token pruning proves advantageous when the model exhibits sensitivity to input interpolations, while token merging is effective when the model manifests close to linear responses to inputs. We combine this to propose a new scheme called Token Fusion. Moreover, we tackle the limitations of average merging, which doesn’t preserve the intrinsic feature norm, resulting in distributional shifts. To mitigate this, we introduce MLERP merging, a variant of the SLERP technique, tailored to merge multiple tokens while maintaining the norm distribution. ToFu is versatile, applicable to ViTs with or without additional training. Our empirical evaluations indicate that ToFu establishes new benchmarks in both classification and image generation tasks concerning computational efficiency and model accuracy.

In this paper, we establish a connection between the parameterization of flow-based and energy-based generative models, and present a new flow-based modeling approach called energy-based normalizing flow (EBFlow). We demonstrate that by optimizing EBFlow with score-matching objectives, the computation of Jacobian determinants for linear transformations can be entirely bypassed. This feature enables the use of arbitrary linear layers in the construction of flow-based models without increasing the computational time complexity of each training iteration from $\mathcal{O}(D^2L)$ to $\mathcal{O}(D^3L)$ for an $L$-layered model that accepts $D$-dimensional inputs. This makes the training of EBFlow more efficient than the commonly-adopted maximum likelihood training method. In addition to the reduction in runtime, we enhance the training stability and empirical performance of EBFlow through a number of techniques developed based on our analysis of the score-matching methods. The experimental results demonstrate that our approach achieves a significant speedup compared to maximum likelihood estimation while outperforming prior methods with a noticeable margin in terms of negative log-likelihood (NLL).

Recent multi-modal detectors based on transformers and modality encoders have successfully achieved impressive results on end-to-end visual object detection conditioned on a raw text query. However, they require a large model size and an enormous amount of computations to achieve high performance, which makes it difficult to deploy mobile applications that are limited by tight hardware resources. In this paper, we present a Lightweight modulated detector, Lite-MDETR, to facilitate efficient end-to-end multi-modal understanding on mobile devices. The key primitive is that Dictionary-Lookup-Transformormations (DLT) is proposed to replace Linear Transformation (LT) in multi-modal detectors where each weight in Linear Transformation (LT) is approximately factorized into a smaller dictionary, index, and coefficient. This way, the enormous linear projection with weights is converted into efficient linear projection with dictionaries, a few lookups and scalings with indices and coefficients. DLT can be applied to any pretrained multi-modal detectors, removing the need to perform expensive training from scratch. To tackle the challenging training of DLT due to non-differentiable index, we convert the index and coefficient into a sparse matrix, train this sparse matrix during the fine-tuning phase, and recover it back to index and coefficient during the inference phase. Our experiments on phrase grounding, referring expression comprehension and segmentation, and VQA show that our Lite-MDETR achieves similar accuracy as the prior multi-modal detectors with up to ~ 4.1 × model size reduction.

Singular value decomposition (SVD) is one of the most popular compression methods that approximate a target matrix with smaller matrices.However, standard SVD treats the parameters within the matrix with equal importance, which is a simple but unrealistic assumption.The parameters of a trained neural network model may affect the task performance unevenly, which suggests non-equal importance among the parameters.Compared to SVD, the decomposition method aware of parameter importance is the more practical choice in real cases.Unlike standard SVD, weighted value decomposition is a non-convex optimization problem that lacks a closed-form solution.We systematically investigated multiple optimization strategies to tackle the problem and examined our method by compressing Transformer-based language models.Further, we designed a metric to predict when the SVD may introduce a significant performance drop, for which our method can be a rescue strategy.The extensive evaluations demonstrate that our method can perform better than current SOTA methods in compressing Transformer-based language models.

Yilin Shen, Yen-Chang Hsu, Avik Ray, Hongxia Jin. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Modern computer vision applications suffer from catastrophic forgetting when incrementally learning new concepts over time. The most successful approaches to alleviate this forgetting require extensive replay of previously seen data, which is problematic when memory constraints or data legality concerns exist. In this work, we consider the high-impact problem of Data-Free Class-Incremental Learning (DFCIL), where an incremental learning agent must learn new concepts over time without storing generators or training data from past tasks. One approach for DFCIL is to replay synthetic images produced by inverting a frozen copy of the learner’s classification model, but we show this approach fails for common class-incremental benchmarks when using standard distillation strategies. We diagnose the cause of this failure and propose a novel incremental distillation strategy for DFCIL, contributing a modified cross-entropy training and importance-weighted feature distillation, and show that our method results in up to a 25.1% increase in final task accuracy (absolute difference) compared to SOTA DFCIL methods for common class-incremental benchmarks. Our method even outperforms several standard replay based methods which store a coreset of images. Our code is available at https://github.com/GT-RIPL/AlwaysBeDreaming-DFCIL

Rehearsal is a critical component for class-incremental continual learning, yet it requires a substantial memory budget. Our work investigates whether we can significantly reduce this memory budget by leveraging unlabeled data from an agent's environment in a realistic and challenging continual learning paradigm. Specifically, we explore and formalize a novel semi-supervised continual learning (SSCL) setting, where labeled data is scarce yet non-i.i.d. unlabeled data from the agent's environment is plentiful. Importantly, data distributions in the SSCL setting are realistic and therefore reflect object class correlations between, and among, the labeled and unlabeled data distributions. We show that a strategy built on pseudo-labeling, consistency regularization, Out-of- Distribution (OoD) detection, and knowledge distillation reduces forgetting in this setting. Our approach, DistillMatch, increases performance over the state-of-the-art by no less than 8.7% average task accuracy and up to 54.5% average task accuracy in SSCL CIFAR-100 experiments. Moreover, we demonstrate that DistillMatch can save up to 0.23 stored images per processed unlabeled image compared to the next best method which only saves 0.08. Our results suggest that focusing on realistic correlated distributions is a significantly new perspective, which accentuates the importance of leveraging the world's structure as a continual learning strategy. Our code is available at https://github.com/GT-RIPL/DistillMatch-SSCL

Ting Hua, Yilin Shen, Changsheng Zhao, Yen-Chang Hsu, Hongxia Jin. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

It is well known that vision classification models suffer from poor calibration in the face of data distribution shifts. In this paper, we take a geometric approach to this problem. We propose Geometric Sensitivity Decomposition (GSD) which decomposes the norm of a sample feature embedding and the angular similarity to a target classifier into an instance-dependent and an instance-independent com-ponent. The instance-dependent component captures the sensitive information about changes in the input while the instance-independent component represents the insensitive information serving solely to minimize the loss on the training dataset. Inspired by the decomposition, we analytically derive a simple extension to current softmax-linear models, which learns to disentangle the two components during training. On several common vision models, the disentangled model out-performs other calibration methods on standard calibration metrics in the face of out-of-distribution (OOD) data and corruption with significantly less complexity. Specifically, we surpass the current state of the art by 30.8% relative improvement on corrupted CIFAR100 in Expected Calibration Error.

Deep neural networks have attained remarkable performance when applied to data that comes from the same distribution as that of the training set, but can significantly degrade otherwise. Therefore, detecting whether an example is out-of-distribution (OoD) is crucial to enable a system that can reject such samples or alert users. Recent works have made significant progress on OoD benchmarks consisting of small image datasets. However, many recent methods based on neural networks rely on training or tuning with both in-distribution and out-of-distribution data. The latter is generally hard to define a-priori, and its selection can easily bias the learning. We base our work on a popular method ODIN, proposing two strategies for freeing it from the needs of tuning with OoD data, while improving its OoD detection performance. We specifically propose to decompose confidence scoring as well as a modified input pre-processing method. We show that both of these significantly help in detection performance. Our further analysis on a larger scale image dataset shows that the two types of distribution shifts, specifically semantic shift and non-semantic shift, present a significant difference in the difficulty of the problem, providing an analysis of when ODIN-like strategies do or do not work.

Neural Networks can perform poorly when the training label distribution is heavily imbalanced, as well as when the testing data differs from the training distribution. In order to deal with shift in the testing label distribution, which imbalance causes, we motivate the problem from the perspective of an optimal Bayes classifier and derive a prior rebalancing technique that can be solved through a KL-divergence based optimization. This method allows a flexible post-training hyper-parameter to be efficiently tuned on a validation set and effectively modify the classifier margin to deal with this imbalance. We further combine this method with existing likelihood shift methods, re-interpreting them from the same Bayesian perspective, and demonstrating that our method can deal with both problems in a unified way. The resulting algorithm can be conveniently used on probabilistic classification problems agnostic to underlying architectures. Our results on six different datasets and five different architectures show state of art accuracy, including on large-scale imbalanced datasets such as iNaturalist for classification and Synthia for semantic segmentation. Please see https://github.com/GT-RIPL/UNO-IC.git for implementation.

This work proposed a novel learning objective to train a deep neural network to perform end-to-end image pixel clustering. We applied the approach to instance segmentation, which is at the intersection of image semantic segmentation and object detection. We utilize the most fundamental property of instance labeling- the pairwise relationship between pixels-as the supervision to formulate the learning objective, then apply it to train a fully convolutional network (FCN) for learning to perform pixel-wise clustering. The resulting clusters can be used as the instance labeling directly. To support labeling of an unlimited number of instance, we further formulate ideas from graph coloring theory into the proposed learning objective. The evaluation on the Cityscapes dataset demonstrates strong performance and therefore proof of the concept. Moreover, our approach won the second place in the lane detection competition of 2017 CVPR Autonomous Driving Challenge, and was the top performer without using external data.