Diana Marculescu

Feng Liang, Bichen Wu, Xiaoliang Dai et al.

Open-vocabulary semantic segmentation aims to segment an image into semantic regions according to text descriptions, which may not have been seen during training. Recent two-stage methods first generate class-agnostic mask proposals and then leverage pre-trained vision-language models, e.g., CLIP, to classify masked regions. We identify the performance bottleneck of this paradigm to be the pre-trained CLIP model, since it does not perform well on masked images. To address this, we propose to finetune CLIP on a collection of masked image regions and their corresponding text descriptions. We collect training data by mining an existing image-caption dataset (e.g., COCO Captions), using CLIP to match masked image regions to nouns in the image captions. Compared with the more precise and manually annotated segmentation labels with fixed classes (e.g., COCO-Stuff), we find our noisy but diverse dataset can better retain CLIP's generalization ability. Along with finetuning the entire model, we utilize the "blank" areas in masked images using a method we dub mask prompt tuning. Experiments demonstrate mask prompt tuning brings significant improvement without modifying any weights of CLIP, and it can further improve a fully finetuned model. In particular, when trained on COCO and evaluated on ADE20K-150, our best model achieves 29.6% mIoU, which is +8.5% higher than the previous state-of-the-art. For the first time, open-vocabulary generalist models match the performance of supervised specialist models in 2017 without dataset-specific adaptations.

Feng Liang, Yangguang Li, Diana Marculescu

Recently, self-supervised Masked Autoencoders (MAE) have attracted unprecedented attention for their impressive representation learning ability. However, the pretext task, Masked Image Modeling (MIM), reconstructs the missing local patches, lacking the global understanding of the image. This paper extends MAE to a fully supervised setting by adding a supervised classification branch, thereby enabling MAE to learn global features from golden labels effectively. The proposed Supervised MAE (SupMAE) only exploits a visible subset of image patches for classification, unlike the standard supervised pre-training where all image patches are used. Through experiments, we demonstrate that SupMAE is not only more training efficient but it also learns more robust and transferable features. Specifically, SupMAE achieves comparable performance with MAE using only 30% of compute when evaluated on ImageNet with the ViT-B/16 model. SupMAE's robustness on ImageNet variants and transfer learning performance outperforms MAE and standard supervised pre-training counterparts. Codes are available at https://github.com/enyac-group/supmae.

Natalia Frumkin, Dibakar Gope, Diana Marculescu

Quantization scale and bit-width are the most important parameters when considering how to quantize a neural network. Prior work focuses on optimizing quantization scales in a global manner through gradient methods (gradient descent \& Hessian analysis). Yet, when applying perturbations to quantization scales, we observe a very jagged, highly non-smooth test loss landscape. In fact, small perturbations in quantization scale can greatly affect accuracy, yielding a $0.5-0.8\%$ accuracy boost in 4-bit quantized vision transformers (ViTs). In this regime, gradient methods break down, since they cannot reliably reach local minima. In our work, dubbed Evol-Q, we use evolutionary search to effectively traverse the non-smooth landscape. Additionally, we propose using an infoNCE loss, which not only helps combat overfitting on the small calibration dataset ($1,000$ images) but also makes traversing such a highly non-smooth surface easier. Evol-Q improves the top-1 accuracy of a fully quantized ViT-Base by $10.30\%$, $0.78\%$, and $0.15\%$ for $3$-bit, $4$-bit, and $8$-bit weight quantization levels. Extensive experiments on a variety of CNN and ViT architectures further demonstrate its robustness in extreme quantization scenarios. Our code is available at https://github.com/enyac-group/evol-q

Yuedong Yang, Hung-Yueh Chiang, Guihong Li et al.

The increasing scale of vision transformers (ViT) has made the efficient fine-tuning of these large models for specific needs a significant challenge in various applications. This issue originates from the computationally demanding matrix multiplications required during the backpropagation process through linear layers in ViT. In this paper, we tackle this problem by proposing a new Low-rank BackPropagation via Walsh-Hadamard Transformation (LBP-WHT) method. Intuitively, LBP-WHT projects the gradient into a low-rank space and carries out backpropagation. This approach substantially reduces the computation needed for adapting ViT, as matrix multiplication in the low-rank space is far less resource-intensive. We conduct extensive experiments with different models (ViT, hybrid convolution-ViT model) on multiple datasets to demonstrate the effectiveness of our method. For instance, when adapting an EfficientFormer-L1 model on CIFAR100, our LBP-WHT achieves 10.4% higher accuracy than the state-of-the-art baseline, while requiring 9 MFLOPs less computation. As the first work to accelerate ViT adaptation with low-rank backpropagation, our LBP-WHT method is complementary to many prior efforts and can be combined with them for better performance.

Tanvir Mahmud, Diana Marculescu

Audio separation in real-world scenarios, where mixtures contain a variable number of sources, presents significant challenges due to limitations of existing models, such as over-separation, under-separation, and dependence on predefined training sources. We propose OpenSep, a novel framework that leverages large language models (LLMs) for automated audio separation, eliminating the need for manual intervention and overcoming source limitations. OpenSep uses textual inversion to generate captions from audio mixtures with off-the-shelf audio captioning models, effectively parsing the sound sources present. It then employs few-shot LLM prompting to extract detailed audio properties of each parsed source, facilitating separation in unseen mixtures. Additionally, we introduce a multi-level extension of the mix-and-separate training framework to enhance modality alignment by separating single source sounds and mixtures simultaneously. Extensive experiments demonstrate OpenSep's superiority in precisely separating new, unseen, and variable sources in challenging mixtures, outperforming SOTA baseline methods. Code is released at https://github.com/tanvir-utexas/OpenSep.git

Natalia Frumkin, Dibakar Gope, Diana Marculescu

When considering post-training quantization, prior work has typically focused on developing a mixed precision scheme or learning the best way to partition a network for quantization. In our work, CPT-V, we look at a general way to improve the accuracy of networks that have already been quantized, simply by perturbing the quantization scales. Borrowing the idea of contrastive loss from self-supervised learning, we find a robust way to jointly minimize a loss function using just 1,000 calibration images. In order to determine the best performing quantization scale, CPT-V contrasts the features of quantized and full precision models in a self-supervised fashion. Unlike traditional reconstruction-based loss functions, the use of a contrastive loss function not only rewards similarity between the quantized and full precision outputs but also helps in distinguishing the quantized output from other outputs within a given batch. In addition, in contrast to prior works, CPT-V proposes a block-wise evolutionary search to minimize a global contrastive loss objective, allowing for accuracy improvement of existing vision transformer (ViT) quantization schemes. For example, CPT-V improves the top-1 accuracy of a fully quantized ViT-Base by 10.30%, 0.78%, and 0.15% for 3-bit, 4-bit, and 8-bit weight quantization levels. Extensive experiments on a variety of other ViT architectures further demonstrate its robustness in extreme quantization scenarios. Our code is available at <link>.

Ahmet Inci, Siri Garudanagiri Virupaksha, Aman Jain et al.

As the machine learning and systems community strives to achieve higher energy-efficiency through custom DNN accelerators and model compression techniques, there is a need for a design space exploration framework that incorporates quantization-aware processing elements into the accelerator design space while having accurate and fast power, performance, and area models. In this work, we present QAPPA, a highly parameterized quantization-aware power, performance, and area modeling framework for DNN accelerators. Our framework can facilitate the future research on design space exploration of DNN accelerators for various design choices such as bit precision, processing element type, scratchpad sizes of processing elements, global buffer size, device bandwidth, number of total processing elements in the the design, and DNN workloads. Our results show that different bit precisions and processing element types lead to significant differences in terms of performance per area and energy. Specifically, our proposed lightweight processing elements achieve up to 4.9x more performance per area and energy improvement when compared to INT16 based implementation.

Chi-Chih Chang, Yuan-Yao Sung, Shixing Yu et al.

Vision Transformers (ViT) have recently demonstrated success across a myriad of computer vision tasks. However, their elevated computational demands pose significant challenges for real-world deployment. While low-rank approximation stands out as a renowned method to reduce computational loads, efficiently automating the target rank selection in ViT remains a challenge. Drawing from the notable similarity and alignment between the processes of rank selection and One-Shot NAS, we introduce FLORA, an end-to-end automatic framework based on NAS. To overcome the design challenge of supernet posed by vast search space, FLORA employs a low-rank aware candidate filtering strategy. This method adeptly identifies and eliminates underperforming candidates, effectively alleviating potential undertraining and interference among subnetworks. To further enhance the quality of low-rank supernets, we design a low-rank specific training paradigm. First, we propose weight inheritance to construct supernet and enable gradient sharing among low-rank modules. Secondly, we adopt low-rank aware sampling to strategically allocate training resources, taking into account inherited information from pre-trained models. Empirical results underscore FLORA's efficacy. With our method, a more fine-grained rank configuration can be generated automatically and yield up to 33% extra FLOPs reduction compared to a simple uniform configuration. More specific, FLORA-DeiT-B/FLORA-Swin-B can save up to 55%/42% FLOPs almost without performance degradtion. Importantly, FLORA boasts both versatility and orthogonality, offering an extra 21%-26% FLOPs reduction when integrated with leading compression techniques or compact hybrid structures. Our code is publicly available at https://github.com/shadowpa0327/FLORA.

Dennis Menn, Yuedong Yang, Bokun Wang et al.

Current video generation models suffer from high computational latency, making real-time applications prohibitively costly. In this paper, we address this limitation by exploiting the temporal redundancy inherent in video latent patches. To this end, we propose the Latent Inter-frame Pruning with Attention Recovery (LIPAR) framework, which detects and skips recomputing duplicated latent patches. Additionally, we introduce a novel Attention Recovery mechanism that approximates the attention values of pruned tokens, thereby removing visual artifacts arising from naively applying the pruning method. Empirically, our method increases video editing throughput by $1.45\times$, on average achieving 12.2 FPS on an NVIDIA A6000 compared to the baseline 8.4 FPS. The proposed method does not compromise generation quality and can be seamlessly integrated with the model without additional training. Our approach effectively bridges the gap between traditional compression algorithms and modern generative pipelines.

Tanvir Mahmud, Diana Marculescu

An audio-visual event (AVE) is denoted by the correspondence of the visual and auditory signals in a video segment. Precise localization of the AVEs is very challenging since it demands effective multi-modal feature correspondence to ground the short and long range temporal interactions. Existing approaches struggle in capturing the different scales of multi-modal interaction due to ineffective multi-modal training strategies. To overcome this limitation, we introduce AVE-CLIP, a novel framework that integrates the AudioCLIP pre-trained on large-scale audio-visual data with a multi-window temporal transformer to effectively operate on different temporal scales of video frames. Our contributions are three-fold: (1) We introduce a multi-stage training framework to incorporate AudioCLIP pre-trained with audio-image pairs into the AVE localization task on video frames through contrastive fine-tuning, effective mean video feature extraction, and multi-scale training phases. (2) We propose a multi-domain attention mechanism that operates on both temporal and feature domains over varying timescales to fuse the local and global feature variations. (3) We introduce a temporal refining scheme with event-guided attention followed by a simple-yet-effective post processing step to handle significant variations of the background over diverse events. Our method achieves state-of-the-art performance on the publicly available AVE dataset with 5.9% mean accuracy improvement which proves its superiority over existing approaches.

Hung-Yueh Chiang, Chi-Chih Chang, Yu-Chen Lu et al.

Deploying large language model (LLM) models on mobile platforms faces significant challenges due to the limited memory and shared computational resources of the device. Resource availability may be an issue as it is directly impacted by the current device workload, adding to the uncertainty of model deployment. We introduce UniQL, a unified post-training quantization and low-rank compression framework with on-device configurable pruning rates for edge LLMs. UniQL is a general framework that integrates quantization and low-rank compression for Transformers, State Space Models (SSMs), and hybrid models to support diverse edge applications. In our proposed joint framework, we introduce an efficient structured weight-sorting method that speeds up computation by 20x, quantization-aware singular value decomposition (SVD) to minimize quantization errors, state-aware weight sorting for SSMs, and a fused rotary positional embedding (RoPE) kernel for pruned models. Our framework performs weight-sorting, fine-tuning, and quantization in the cloud in a single-pass workflow, while enabling on-device configurable pruning rates up to 35%. Our experiments show that quantized and pruned models achieve a memory reduction of 4x-5.7x and a token-throughput improvement of 2.7x-3.4x, maintaining accuracy within 5% of the original models at 15% pruning across Transformers (Llama3 and Qwen2.5), SSMs (Mamba2), and hybrid models (Nemotron-H and Bamba-v2). The code and quantized models are available at: https://github.com/enyac-group/UniQL.

Hung-Yueh Chiang, Natalia Frumkin, Feng Liang et al.

Transfer learning on edge is challenging due to on-device limited resources. Existing work addresses this issue by training a subset of parameters or adding model patches. Developed with inference in mind, Inverted Residual Blocks (IRBs) split a convolutional layer into depthwise and pointwise convolutions, leading to more stacking layers, e.g., convolution, normalization, and activation layers. Though they are efficient for inference, IRBs require that additional activation maps are stored in memory for training weights for convolution layers and scales for normalization layers. As a result, their high memory cost prohibits training IRBs on resource-limited edge devices, and making them unsuitable in the context of transfer learning. To address this issue, we present MobileTL, a memory and computationally efficient on-device transfer learning method for models built with IRBs. MobileTL trains the shifts for internal normalization layers to avoid storing activation maps for the backward pass. Also, MobileTL approximates the backward computation of the activation layer (e.g., Hard-Swish and ReLU6) as a signed function which enables storing a binary mask instead of activation maps for the backward pass. MobileTL fine-tunes a few top blocks (close to output) rather than propagating the gradient through the whole network to reduce the computation cost. Our method reduces memory usage by 46% and 53% for MobileNetV2 and V3 IRBs, respectively. For MobileNetV3, we observe a 36% reduction in floating-point operations (FLOPs) when fine-tuning 5 blocks, while only incurring a 0.6% accuracy reduction on CIFAR10. Extensive experiments on multiple datasets demonstrate that our method is Pareto-optimal (best accuracy under given hardware constraints) compared to prior work in transfer learning for edge devices.

Ahmet Inci, Siri Garudanagiri Virupaksha, Aman Jain et al.

As the machine learning and systems communities strive to achieve higher energy-efficiency through custom deep neural network (DNN) accelerators, varied precision or quantization levels, and model compression techniques, there is a need for design space exploration frameworks that incorporate quantization-aware processing elements into the accelerator design space while having accurate and fast power, performance, and area models. In this work, we present QUIDAM, a highly parameterized quantization-aware DNN accelerator and model co-exploration framework. Our framework can facilitate future research on design space exploration of DNN accelerators for various design choices such as bit precision, processing element type, scratchpad sizes of processing elements, global buffer size, number of total processing elements, and DNN configurations. Our results show that different bit precisions and processing element types lead to significant differences in terms of performance per area and energy. Specifically, our framework identifies a wide range of design points where performance per area and energy varies more than 5x and 35x, respectively. With the proposed framework, we show that lightweight processing elements achieve on par accuracy results and up to 5.7x more performance per area and energy improvement when compared to the best INT16 based implementation. Finally, due to the efficiency of the pre-characterized power, performance, and area models, QUIDAM can speed up the design exploration process by 3-4 orders of magnitude as it removes the need for expensive synthesis and characterization of each design.

Yang Zhou, Feng Liang, Ting-wu Chin et al.

Machine learning (ML) has entered the mobile era where an enormous number of ML models are deployed on edge devices. However, running common ML models on edge devices continuously may generate excessive heat from the computation, forcing the device to "slow down" to prevent overheating, a phenomenon called thermal throttling. This paper studies the impact of thermal throttling on mobile phones: when it occurs, the CPU clock frequency is reduced, and the model inference latency may increase dramatically. This unpleasant inconsistent behavior has a substantial negative effect on user experience, but it has been overlooked for a long time. To counter thermal throttling, we propose to utilize dynamic networks with shared weights and dynamically shift between large and small ML models seamlessly according to their thermal profile, i.e., shifting to a small model when the system is about to throttle. With the proposed dynamic shifting, the application runs consistently without experiencing CPU clock frequency degradation and latency increase. In addition, we also study the resulting accuracy when dynamic shifting is deployed and show that our approach provides a reasonable trade-off between model latency and model accuracy.

Ahmet Inci, Siri Garudanagiri Virupaksha, Aman Jain et al.

As the machine learning and systems communities strive to achieve higher energy-efficiency through custom deep neural network (DNN) accelerators, varied bit precision or quantization levels, there is a need for design space exploration frameworks that incorporate quantization-aware processing elements (PE) into the accelerator design space while having accurate and fast power, performance, and area models. In this work, we present QADAM, a highly parameterized quantization-aware power, performance, and area modeling framework for DNN accelerators. Our framework can facilitate future research on design space exploration and Pareto-efficiency of DNN accelerators for various design choices such as bit precision, PE type, scratchpad sizes of PEs, global buffer size, number of total PEs, and DNN configurations. Our results show that different bit precisions and PE types lead to significant differences in terms of performance per area and energy. Specifically, our framework identifies a wide range of design points where performance per area and energy varies more than 5x and 35x, respectively. We also show that the proposed lightweight processing elements (LightPEs) consistently achieve Pareto-optimal results in terms of accuracy and hardware-efficiency. With the proposed framework, we show that LightPEs achieve on par accuracy results and up to 5.7x more performance per area and energy improvement when compared to the best INT16 based design.

Tanvir Mahmud, Yapeng Tian, Diana Marculescu

Visual sound source localization poses a significant challenge in identifying the semantic region of each sounding source within a video. Existing self-supervised and weakly supervised source localization methods struggle to accurately distinguish the semantic regions of each sounding object, particularly in multi-source mixtures. These methods often rely on audio-visual correspondence as guidance, which can lead to substantial performance drops in complex multi-source localization scenarios. The lack of access to individual source sounds in multi-source mixtures during training exacerbates the difficulty of learning effective audio-visual correspondence for localization. To address this limitation, in this paper, we propose incorporating the text modality as an intermediate feature guide using tri-modal joint embedding models (e.g., AudioCLIP) to disentangle the semantic audio-visual source correspondence in multi-source mixtures. Our framework, dubbed T-VSL, begins by predicting the class of sounding entities in mixtures. Subsequently, the textual representation of each sounding source is employed as guidance to disentangle fine-grained audio-visual source correspondence from multi-source mixtures, leveraging the tri-modal AudioCLIP embedding. This approach enables our framework to handle a flexible number of sources and exhibits promising zero-shot transferability to unseen classes during test time. Extensive experiments conducted on the MUSIC, VGGSound, and VGGSound-Instruments datasets demonstrate significant performance improvements over state-of-the-art methods. Code is released at https://github.com/enyac-group/T-VSL/tree/main

Ahmet Inci, Mehmet Meric Isgenc, Diana Marculescu

Embedded machine learning (ML) systems have now become the dominant platform for deploying ML serving tasks and are projected to become of equal importance for training ML models. With this comes the challenge of overall efficient deployment, in particular low power and high throughput implementations, under stringent memory constraints. In this context, non-volatile memory (NVM) technologies such as STT-MRAM and SOT-MRAM have significant advantages compared to conventional SRAM due to their non-volatility, higher cell density, and scalability features. While prior work has investigated several architectural implications of NVM for generic applications, in this work we present DeepNVM++, a comprehensive framework to characterize, model, and analyze NVM-based caches in GPU architectures for deep learning (DL) applications by combining technology-specific circuit-level models and the actual memory behavior of various DL workloads. DeepNVM++ relies on iso-capacity and iso-area performance and energy models for last-level caches implemented using conventional SRAM and emerging STT-MRAM and SOT-MRAM technologies. In the iso-capacity case, STT-MRAM and SOT-MRAM provide up to 3.8x and 4.7x energy-delay product (EDP) reduction and 2.4x and 2.8x area reduction compared to conventional SRAM, respectively. Under iso-area assumptions, STT-MRAM and SOT-MRAM provide up to 2.2x and 2.4x EDP reduction and accommodate 2.3x and 3.3x cache capacity when compared to SRAM, respectively. We also perform a scalability analysis and show that STT-MRAM and SOT-MRAM achieve orders of magnitude EDP reduction when compared to SRAM for large cache capacities. DeepNVM++ is demonstrated on STT-/SOT-MRAM technologies and can be used for the characterization, modeling, and analysis of any NVM technology for last-level caches in GPUs for DL applications.

Natalia Frumkin, Bokun Wang, Hung-Yueh Chiang et al.

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to auto-regressive (AR) models, offering greater expressive capacity and potential for parallel generation and faster inference. However, open-source dLLMs remain immature, lagging behind AR models in both efficiency and quality. We identify an underexplored property of dLLMs: *token-wise redundancy* in bi-directional self-attention. Self-attention activations are highly correlated across tokens, and temporal changes in query representations can predict redundancy in corresponding key, value, and output activations. We introduce DARE, with two complementary mechanisms: DARE-KV, which reuses cached key-value (KV) activations, and DARE-O, which reuses output activations to reduce redundant computation while preserving quality. DARE achieves up to 1.20x per-layer latency reduction and reuses up to 87% of attention activations, with negligible degradation on reasoning and code-generation benchmarks. DARE-KV and DARE-O incur average performance drops of only 2.0% and 1.2%, respectively. Combined with techniques such as prefix caching and Fast-dLLM, DARE provides additive gains without retraining. These results establish token-wise reuse as an effective strategy for improving the efficiency of diffusion-based LLMs while preserving generation fidelity. Code: https://github.com/enyac-group/DARE

Ning-Chi Huang, Chi-Chih Chang, Wei-Cheng Lin et al.

$N{:}M$ sparsity is an emerging model compression method supported by more and more accelerators to speed up sparse matrix multiplication in deep neural networks. Most existing $N{:}M$ sparsity methods compress neural networks with a uniform setting for all layers in a network or heuristically determine the layer-wise configuration by considering the number of parameters in each layer. However, very few methods have been designed for obtaining a layer-wise customized $N{:}M$ sparse configuration for vision transformers (ViTs), which usually consist of transformer blocks involving the same number of parameters. In this work, to address the challenge of selecting suitable sparse configuration for ViTs on $N{:}M$ sparsity-supporting accelerators, we propose ELSA, Exploiting Layer-wise $N{:}M$ Sparsity for ViTs. Considering not only all $N{:}M$ sparsity levels supported by a given accelerator but also the expected throughput improvement, our methodology can reap the benefits of accelerators supporting mixed sparsity by trading off negligible accuracy loss with both memory usage and inference time reduction for ViT models. For instance, our approach achieves a noteworthy 2.9$\times$ reduction in FLOPs for both Swin-B and DeiT-B with only a marginal degradation of accuracy on ImageNet. Our code will be released upon paper acceptance.

Tanvir Mahmud, Burhaneddin Yaman, Chun-Hao Liu et al.

As deep neural networks evolve from convolutional neural networks (ConvNets) to advanced vision transformers (ViTs), there is an increased need to eliminate redundant data for faster processing without compromising accuracy. Previous methods are often architecture-specific or necessitate re-training, restricting their applicability with frequent model updates. To solve this, we first introduce a novel property of lightweight ConvNets: their ability to identify key discriminative patch regions in images, irrespective of model's final accuracy or size. We demonstrate that fully-connected layers are the primary bottleneck for ConvNets performance, and their suppression with simple weight recalibration markedly enhances discriminative patch localization performance. Using this insight, we introduce PaPr, a method for substantially pruning redundant patches with minimal accuracy loss using lightweight ConvNets across a variety of deep learning architectures, including ViTs, ConvNets, and hybrid transformers, without any re-training. Moreover, the simple early-stage one-step patch pruning with PaPr enhances existing patch reduction methods. Through extensive testing on diverse architectures, PaPr achieves significantly higher accuracy over state-of-the-art patch reduction methods with similar FLOP count reduction. More specifically, PaPr reduces about 70% of redundant patches in videos with less than 0.8% drop in accuracy, and up to 3.7x FLOPs reduction, which is a 15% more reduction with 2.5% higher accuracy. Code is released at https://github.com/tanvir-utexas/PaPr.

Hung-Yueh Chiang, Chi-Chih Chang, Natalia Frumkin et al.

State Space Models (SSMs) are emerging as a compelling alternative to Transformers because of their consistent memory usage and high performance. Despite this, scaling up SSMs on cloud services or limited-resource devices is challenging due to their storage requirements and computational power. To overcome this, quantizing SSMs with low bit-width data formats can reduce model size and benefit from hardware acceleration. As SSMs are prone to quantization-induced errors, recent efforts have focused on optimizing a particular model or bit-width for efficiency without sacrificing performance. However, distinct bit-width configurations are essential for different scenarios, like W4A8 for boosting large-batch decoding speed, and W4A16 for enhancing generation speed in short prompt applications for a single user. To this end, we present Quamba2, compatible with W8A8, W4A8, and W4A16 for both Mamba1 and Mamba2 backbones, addressing the growing demand for SSM deployment on various platforms. Based on the channel order preserving and activation persistence of SSMs, we propose an offline approach to quantize inputs of a linear recurrence in 8-bit by sorting and clustering for input $x$, combined with a per-state-group quantization for input-dependent parameters $B$ and $C$. To ensure compute-invariance in the SSM output, we rearrange weights offline according to the clustering sequence. The experiments show that Quamba2-8B outperforms two state-of-the-art SSM quantization methods and delivers 1.3$\times$ and 3$\times$ speed-ups in the pre-filling and generation stages, respectively, while offering 4$\times$ memory reduction with only a $1.6\%$ average accuracy drop. The evaluation on MMLU shows the generalizability and robustness of our framework. The code and quantized models will be released at: https://github.com/enyac-group/Quamba.

Tanvir Mahmud, Mustafa Munir, Radu Marculescu et al.

Video-to-video synthesis poses significant challenges in maintaining character consistency, smooth temporal transitions, and preserving visual quality during fast motion. While recent fully cross-frame self-attention mechanisms have improved character consistency across multiple frames, they come with high computational costs and often include redundant operations, especially for videos with higher frame rates. To address these inefficiencies, we propose an adaptive motion-guided cross-frame attention mechanism that selectively reduces redundant computations. This enables a greater number of cross-frame attentions over more frames within the same computational budget, thereby enhancing both video quality and temporal coherence. Our method leverages optical flow to focus on moving regions while sparsely attending to stationary areas, allowing for the joint editing of more frames without increasing computational demands. Traditional frame interpolation techniques struggle with motion blur and flickering in intermediate frames, which compromises visual fidelity. To mitigate this, we introduce KV-caching for jointly edited frames, reusing keys and values across intermediate frames to preserve visual quality and maintain temporal consistency throughout the video. With our adaptive cross-frame self-attention approach, we achieve a threefold increase in the number of keyframes processed compared to existing methods, all within the same computational budget as fully cross-frame attention baselines. This results in significant improvements in prediction accuracy and temporal consistency, outperforming state-of-the-art approaches. Code will be made publicly available at https://github.com/tanvir-utexas/AdaVE/tree/main

Ting-Wu Chin, Ari S. Morcos, Diana Marculescu

Slimmable neural networks provide a flexible trade-off front between prediction error and computational requirement (such as the number of floating-point operations or FLOPs) with the same storage requirement as a single model. They are useful for reducing maintenance overhead for deploying models to devices with different memory constraints and are useful for optimizing the efficiency of a system with many CNNs. However, existing slimmable network approaches either do not optimize layer-wise widths or optimize the shared-weights and layer-wise widths independently, thereby leaving significant room for improvement by joint width and weight optimization. In this work, we propose a general framework to enable joint optimization for both width configurations and weights of slimmable networks. Our framework subsumes conventional and NAS-based slimmable methods as special cases and provides flexibility to improve over existing methods. From a practical standpoint, we propose Joslim, an algorithm that jointly optimizes both the widths and weights for slimmable nets, which outperforms existing methods for optimizing slimmable networks across various networks, datasets, and objectives. Quantitatively, improvements up to 1.7% and 8% in top-1 accuracy on the ImageNet dataset can be attained for MobileNetV2 considering FLOPs and memory footprint, respectively. Our results highlight the potential of optimizing the channel counts for different layers jointly with the weights for slimmable networks. Code available at https://github.com/cmu-enyac/Joslim.

Ting-Wu Chin, Cha Zhang, Diana Marculescu

Fine-tuning through knowledge transfer from a pre-trained model on a large-scale dataset is a widely spread approach to effectively build models on small-scale datasets. In this work, we show that a recent adversarial attack designed for transfer learning via re-training the last linear layer can successfully deceive models trained with transfer learning via end-to-end fine-tuning. This raises security concerns for many industrial applications. In contrast, models trained with random initialization without transfer are much more robust to such attacks, although these models often exhibit much lower accuracy. To this end, we propose noisy feature distillation, a new transfer learning method that trains a network from random initialization while achieving clean-data performance competitive with fine-tuning. Code available at https://github.com/cmu-enyac/Renofeation.

Dimitrios Stamoulis, Ruizhou Ding, Di Wang et al.

Can we reduce the search cost of Neural Architecture Search (NAS) from days down to only few hours? NAS methods automate the design of Convolutional Networks (ConvNets) under hardware constraints and they have emerged as key components of AutoML frameworks. However, the NAS problem remains challenging due to the combinatorially large design space and the significant search time (at least 200 GPU-hours). In this work, we alleviate the NAS search cost down to less than 3 hours, while achieving state-of-the-art image classification results under mobile latency constraints. We propose a novel differentiable NAS formulation, namely Single-Path NAS, that uses one single-path over-parameterized ConvNet to encode all architectural decisions based on shared convolutional kernel parameters, hence drastically decreasing the search overhead. Single-Path NAS achieves state-of-the-art top-1 ImageNet accuracy (75.62%), hence outperforming existing mobile NAS methods in similar latency settings (~80ms). In particular, we enhance the accuracy-runtime trade-off in differentiable NAS by treating the Squeeze-and-Excitation path as a fully searchable operation with our novel single-path encoding. Our method has an overall cost of only 8 epochs (24 TPU-hours), which is up to 5,000x faster compared to prior work. Moreover, we study how different NAS formulation choices affect the performance of the designed ConvNets. Furthermore, we exploit the efficiency of our method to answer an interesting question: instead of empirically tuning the hyperparameters of the NAS solver (as in prior work), can we automatically find the hyperparameter values that yield the desired accuracy-runtime trade-off? We open-source our entire codebase at: https://github.com/dstamoulis/single-path-nas.

Zhuo Chen, Jiyuan Zhang, Ruizhou Ding et al.

In recent years, Convolutional Neural Networks (CNNs) have shown superior capability in visual learning tasks. While accuracy-wise CNNs provide unprecedented performance, they are also known to be computationally intensive and energy demanding for modern computer systems. In this paper, we propose Virtual Pooling (ViP), a model-level approach to improve speed and energy consumption of CNN-based image classification and object detection tasks, with a provable error bound. We show the efficacy of ViP through experiments on four CNN models, three representative datasets, both desktop and mobile platforms, and two visual learning tasks, i.e., image classification and object detection. For example, ViP delivers 2.1x speedup with less than 1.5% accuracy degradation in ImageNet classification on VGG-16, and 1.8x speedup with 0.025 mAP degradation in PASCAL VOC object detection with Faster-RCNN. ViP also reduces mobile GPU and CPU energy consumption by up to 55% and 70%, respectively. As a complementary method to existing acceleration approaches, ViP achieves 1.9x speedup on ThiNet leading to a combined speedup of 5.23x on VGG-16. Furthermore, ViP provides a knob for machine learning practitioners to generate a set of CNN models with varying trade-offs between system speed/energy consumption and accuracy to better accommodate the requirements of their tasks. Code is available at https://github.com/cmu-enyac/VirtualPooling.

Ting-Wu Chin, Ruizhou Ding, Cha Zhang et al.

Pruning convolutional filters has demonstrated its effectiveness in compressing ConvNets. Prior art in filter pruning requires users to specify a target model complexity (e.g., model size or FLOP count) for the resulting architecture. However, determining a target model complexity can be difficult for optimizing various embodied AI applications such as autonomous robots, drones, and user-facing applications. First, both the accuracy and the speed of ConvNets can affect the performance of the application. Second, the performance of the application can be hard to assess without evaluating ConvNets during inference. As a consequence, finding a sweet-spot between the accuracy and speed via filter pruning, which needs to be done in a trial-and-error fashion, can be time-consuming. This work takes a first step toward making this process more efficient by altering the goal of model compression to producing a set of ConvNets with various accuracy and latency trade-offs instead of producing one ConvNet targeting some pre-defined latency constraint. To this end, we propose to learn a global ranking of the filters across different layers of the ConvNet, which is used to obtain a set of ConvNet architectures that have different accuracy/latency trade-offs by pruning the bottom-ranked filters. Our proposed algorithm, LeGR, is shown to be 2x to 3x faster than prior work while having comparable or better performance when targeting seven pruned ResNet-56 with different accuracy/FLOPs profiles on the CIFAR-100 dataset. Additionally, we have evaluated LeGR on ImageNet and Bird-200 with ResNet-50 and MobileNetV2 to demonstrate its effectiveness. Code available at https://github.com/cmu-enyac/LeGR.

Dimitrios Stamoulis, Ruizhou Ding, Di Wang et al.

Can we automatically design a Convolutional Network (ConvNet) with the highest image classification accuracy under the runtime constraint of a mobile device? Neural architecture search (NAS) has revolutionized the design of hardware-efficient ConvNets by automating this process. However, the NAS problem remains challenging due to the combinatorially large design space, causing a significant searching time (at least 200 GPU-hours). To alleviate this complexity, we propose Single-Path NAS, a novel differentiable NAS method for designing hardware-efficient ConvNets in less than 4 hours. Our contributions are as follows: 1. Single-path search space: Compared to previous differentiable NAS methods, Single-Path NAS uses one single-path over-parameterized ConvNet to encode all architectural decisions with shared convolutional kernel parameters, hence drastically decreasing the number of trainable parameters and the search cost down to few epochs. 2. Hardware-efficient ImageNet classification: Single-Path NAS achieves 74.96% top-1 accuracy on ImageNet with 79ms latency on a Pixel 1 phone, which is state-of-the-art accuracy compared to NAS methods with similar constraints (<80ms). 3. NAS efficiency: Single-Path NAS search cost is only 8 epochs (30 TPU-hours), which is up to 5,000x faster compared to prior work. 4. Reproducibility: Unlike all recent mobile-efficient NAS methods which only release pretrained models, we open-source our entire codebase at: https://github.com/dstamoulis/single-path-nas.

Zhuo Chen, Ruizhou Ding, Ting-Wu Chin et al.

In recent years, supervised learning using Convolutional Neural Networks (CNNs) has achieved great success in image classification tasks, and large scale labeled datasets have contributed significantly to this achievement. However, the definition of a label is often application dependent. For example, an image of a cat can be labeled as "cat" or perhaps more specifically "Persian cat." We refer to this as label granularity. In this paper, we conduct extensive experiments using various datasets to demonstrate and analyze how and why training based on fine-grain labeling, such as "Persian cat" can improve CNN accuracy on classifying coarse-grain classes, in this case "cat." The experimental results show that training CNNs with fine-grain labels improves both network's optimization and generalization capabilities, as intuitively it encourages the network to learn more features, and hence increases classification accuracy on coarse-grain classes under all datasets considered. Moreover, fine-grain labels enhance data efficiency in CNN training. For example, a CNN trained with fine-grain labels and only 40% of the total training data can achieve higher accuracy than a CNN trained with the full training dataset and coarse-grain labels. These results point to two possible applications of this work: (i) with sufficient human resources, one can improve CNN performance by re-labeling the dataset with fine-grain labels, and (ii) with limited human resources, to improve CNN performance, rather than collecting more training data, one may instead use fine-grain labels for the dataset. We further propose a metric called Average Confusion Ratio to characterize the effectiveness of fine-grain labeling, and show its use through extensive experimentation. Code is available at https://github.com/cmu-enyac/Label-Granularity.

Feng Liang, Bichen Wu, Jialiang Wang et al.

Diffusion models have transformed the image-to-image (I2I) synthesis and are now permeating into videos. However, the advancement of video-to-video (V2V) synthesis has been hampered by the challenge of maintaining temporal consistency across video frames. This paper proposes a consistent V2V synthesis framework by jointly leveraging spatial conditions and temporal optical flow clues within the source video. Contrary to prior methods that strictly adhere to optical flow, our approach harnesses its benefits while handling the imperfection in flow estimation. We encode the optical flow via warping from the first frame and serve it as a supplementary reference in the diffusion model. This enables our model for video synthesis by editing the first frame with any prevalent I2I models and then propagating edits to successive frames. Our V2V model, FlowVid, demonstrates remarkable properties: (1) Flexibility: FlowVid works seamlessly with existing I2I models, facilitating various modifications, including stylization, object swaps, and local edits. (2) Efficiency: Generation of a 4-second video with 30 FPS and 512x512 resolution takes only 1.5 minutes, which is 3.1x, 7.2x, and 10.5x faster than CoDeF, Rerender, and TokenFlow, respectively. (3) High-quality: In user studies, our FlowVid is preferred 45.7% of the time, outperforming CoDeF (3.5%), Rerender (10.2%), and TokenFlow (40.4%).

Hung-Yueh Chiang, Diana Marculescu

Designing low-latency and high-efficiency hybrid networks for a variety of low-cost commodity edge devices is both costly and tedious, leading to the adoption of hardware-aware neural architecture search (NAS) for finding optimal architectures. However, unifying NAS for a wide range of edge devices presents challenges due to the variety of hardware designs, supported operations, and compilation optimizations. Existing methods often fix the search space of architecture choices (e.g., activation, convolution, or self-attention) and estimate latency using hardware-agnostic proxies (e.g., FLOPs), which fail to achieve proclaimed latency across various edge devices. To address this issue, we propose SCAN-Edge, a unified NAS framework that jointly searches for self-attention, convolution, and activation to accommodate the wide variety of edge devices, including CPU-, GPU-, and hardware accelerator-based systems. To handle the large search space, SCAN-Edge relies on with a hardware-aware evolutionary algorithm that improves the quality of the search space to accelerate the sampling process. Experiments on large-scale datasets demonstrate that our hybrid networks match the actual MobileNetV2 latency for 224x224 input resolution on various commodity edge devices.

Feng Liang, Akio Kodaira, Chenfeng Xu et al.

This paper introduces StreamV2V, a diffusion model that achieves real-time streaming video-to-video (V2V) translation with user prompts. Unlike prior V2V methods using batches to process limited frames, we opt to process frames in a streaming fashion, to support unlimited frames. At the heart of StreamV2V lies a backward-looking principle that relates the present to the past. This is realized by maintaining a feature bank, which archives information from past frames. For incoming frames, StreamV2V extends self-attention to include banked keys and values and directly fuses similar past features into the output. The feature bank is continually updated by merging stored and new features, making it compact but informative. StreamV2V stands out for its adaptability and efficiency, seamlessly integrating with image diffusion models without fine-tuning. It can run 20 FPS on one A100 GPU, being 15x, 46x, 108x, and 158x faster than FlowVid, CoDeF, Rerender, and TokenFlow, respectively. Quantitative metrics and user studies confirm StreamV2V's exceptional ability to maintain temporal consistency.

Hung-Yueh Chiang, Chi-Chih Chang, Natalia Frumkin et al.

State Space Models (SSMs) have emerged as an appealing alternative to Transformers for large language models, achieving state-of-the-art accuracy with constant memory complexity which allows for holding longer context lengths than attention-based networks. The superior computational efficiency of SSMs in long sequence modeling positions them favorably over Transformers in many scenarios. However, improving the efficiency of SSMs on request-intensive cloud-serving and resource-limited edge applications is still a formidable task. SSM quantization is a possible solution to this problem, making SSMs more suitable for wide deployment, while still maintaining their accuracy. Quantization is a common technique to reduce the model size and to utilize the low bit-width acceleration features on modern computing units, yet existing quantization techniques are poorly suited for SSMs. Most notably, SSMs have highly sensitive feature maps within the selective scan mechanism (i.e., linear recurrence) and massive outliers in the output activations which are not present in the output of token-mixing in the self-attention modules. To address this issue, we propose a static 8-bit per-tensor SSM quantization method which suppresses the maximum values of the input activations to the selective SSM for finer quantization precision and quantizes the output activations in an outlier-free space with Hadamard transform. Our 8-bit weight-activation quantized Mamba 2.8B SSM benefits from hardware acceleration and achieves a 1.72x lower generation latency on an Nvidia Orin Nano 8G, with only a 0.9% drop in average accuracy on zero-shot tasks. The experiments demonstrate the effectiveness and practical applicability of our approach for deploying SSM-based models of all sizes on both cloud and edge platforms.

Feng Liang, Haoyu Ma, Zecheng He et al.

Video personalization, which generates customized videos using reference images, has gained significant attention. However, prior methods typically focus on single-concept personalization, limiting broader applications that require multi-concept integration. Attempts to extend these models to multiple concepts often lead to identity blending, which results in composite characters with fused attributes from multiple sources. This challenge arises due to the lack of a mechanism to link each concept with its specific reference image. We address this with anchored prompts, which embed image anchors as unique tokens within text prompts, guiding accurate referencing during generation. Additionally, we introduce concept embeddings to encode the order of reference images. Our approach, Movie Weaver, seamlessly weaves multiple concepts-including face, body, and animal images-into one video, allowing flexible combinations in a single model. The evaluation shows that Movie Weaver outperforms existing methods for multi-concept video personalization in identity preservation and overall quality.

Dimitrios Stamoulis, Diana Marculescu

Geospatial Copilots hold immense potential for automating Earth observation (EO) and climate monitoring workflows, yet their reliance on large-scale models such as GPT-4o introduces a paradox: tools intended for sustainability studies often incur unsustainable costs. Using agentic AI frameworks in geospatial applications can amass thousands of dollars in API charges or requires expensive, power-intensive GPUs for deployment, creating barriers for researchers, policymakers, and NGOs. Unfortunately, when geospatial Copilots are deployed with open language models (OLMs), performance often degrades due to their dependence on GPT-optimized logic. In this paper, we present Geo-OLM, a tool-augmented geospatial agent that leverages the novel paradigm of state-driven LLM reasoning to decouple task progression from tool calling. By alleviating the workflow reasoning burden, our approach enables low-resource OLMs to complete geospatial tasks more effectively. When downsizing to small models below 7B parameters, Geo-OLM outperforms the strongest prior geospatial baselines by 32.8% in successful query completion rates. Our method performs comparably to proprietary models achieving results within 10% of GPT-4o, while reducing inference costs by two orders of magnitude from \$500-\$1000 to under \$10. We present an in-depth analysis with geospatial downstream benchmarks, providing key insights to help practitioners effectively deploy OLMs for EO applications.

Tanvir Mahmud, Saeed Amizadeh, Kazuhito Koishida et al.

Conditional sound separation in multi-source audio mixtures without having access to single source sound data during training is a long standing challenge. Existing mix-and-separate based methods suffer from significant performance drop with multi-source training mixtures due to the lack of supervision signal for single source separation cases during training. However, in the case of language-conditional audio separation, we do have access to corresponding text descriptions for each audio mixture in our training data, which can be seen as (rough) representations of the audio samples in the language modality. To this end, in this paper, we propose a generic bi-modal separation framework which can enhance the existing unsupervised frameworks to separate single-source signals in a target modality (i.e., audio) using the easily separable corresponding signals in the conditioning modality (i.e., language), without having access to single-source samples in the target modality during training. We empirically show that this is well within reach if we have access to a pretrained joint embedding model between the two modalities (i.e., CLAP). Furthermore, we propose to incorporate our framework into two fundamental scenarios to enhance separation performance. First, we show that our proposed methodology significantly improves the performance of purely unsupervised baselines by reducing the distribution shift between training and test samples. In particular, we show that our framework can achieve 71% boost in terms of Signal-to-Distortion Ratio (SDR) over the baseline, reaching 97.5% of the supervised learning performance. Second, we show that we can further improve the performance of the supervised learning itself by 17% if we augment it by our proposed weakly-supervised framework, that enables a powerful semi-supervised framework for audio separation.

Amulya Bhattaram, Janani Ramamoorthy, Ranit Gupta et al.

The rise of agentic AI workflows unlocks novel opportunities for computer systems design and optimization. However, for specialized domains such as program synthesis, the relative scarcity of HDL and proprietary EDA resources online compared to more common programming tasks introduces challenges, often necessitating task-specific fine-tuning, high inference costs, and manually-crafted agent orchestration. In this work, we present VeriMaAS, a multi-agent framework designed to automatically compose agentic workflows for RTL code generation. Our key insight is to integrate formal verification feedback from HDL tools directly into workflow generation, reducing the cost of gradient-based updates or prolonged reasoning traces. Our method improves synthesis performance by 5-7% for pass@k over fine-tuned baselines, while requiring only a few hundred training examples, representing an order-of-magnitude reduction in supervision cost.

Amulya Bhattaram, Justin Chung, Stanley Chung et al.

We present GeoFlow, a method that automatically generates agentic workflows for geospatial tasks. Unlike prior work that focuses on reasoning decomposition and leaves API selection implicit, our method provides each agent with detailed tool-calling objectives to guide geospatial API invocation at runtime. GeoFlow increases agentic success by 6.8% and reduces token usage by up to fourfold across major LLM families compared to state-of-the-art approaches.

Dennis Menn, Feng Liang, Hung-Yueh Chiang et al.

Artifact detection algorithms are crucial to correcting the output generated by diffusion models. However, because of the variety of artifact forms, existing methods require substantial annotated data for training. This requirement limits their scalability and efficiency, which restricts their wide application. This paper shows that the similarity of denoised images between consecutive time steps during the sampling process is related to the severity of artifacts in images generated by diffusion models. Building on this observation, we introduce the concept of Similarity Trajectory to characterize the sampling process and its correlation with the image artifacts presented. Using an annotated data set of 680 images, which is only 0.1% of the amount of data used in the prior work, we trained a classifier on these trajectories to predict the presence of artifacts in images. By performing 10-fold validation testing on the balanced annotated data set, the classifier can achieve an accuracy of 72.35%, highlighting the connection between the Similarity Trajectory and the occurrence of artifacts. This approach enables differentiation between artifact-exhibiting and natural-looking images using limited training data.

Yang Zhou, Zhen Dong, Ellick Chan et al.

Large-scale recommendation models are currently the dominant workload for many large Internet companies. These recommenders are characterized by massive embedding tables that are sparsely accessed by the index for user and item features. The size of these 1TB+ tables imposes a severe memory bottleneck for the training and inference of recommendation models. In this work, we propose a novel recommendation framework that is small, powerful, and efficient to run and train, based on the state-of-the-art Deep Learning Recommendation Model (DLRM). The proposed framework makes inference more efficient on the cloud servers, explores the possibility of deploying powerful recommenders on smaller edge devices, and optimizes the workload of the communication overhead in distributed training under the data parallelism settings. Specifically, we show that quantization-aware training (QAT) can impose a strong regularization effect to mitigate the severe overfitting issues suffered by DLRMs. Consequently, we achieved INT4 quantization of DLRM models without any accuracy drop. We further propose two techniques that improve and accelerate the conventional QAT workload specifically for the embedding tables in the recommendation models. Furthermore, to achieve efficient training, we quantize the gradients of the embedding tables into INT8 on top of the well-supported specified sparsification. We show that combining gradient sparsification and quantization together significantly reduces the amount of communication. Briefly, DQRM models with INT4 can achieve 79.07% accuracy on Kaggle with 0.27 GB model size, and 81.21% accuracy on the Terabyte dataset with 1.57 GB, which even outperform FP32 DLRMs that have much larger model sizes (2.16 GB on Kaggle and 12.58 on Terabyte).

Abhijit Chatterjee, Niraj K. Jha, Jonathan D. Cohen et al.

The field of artificial intelligence (AI) has taken a tight hold on broad aspects of society, industry, business, and governance in ways that dictate the prosperity and might of the world's economies. The AI market size is projected to grow from 189 billion USD in 2023 to 4.8 trillion USD by 2033. Currently, AI is dominated by large language models that exhibit linguistic and visual intelligence. However, training these models requires a massive amount of data scraped from the web as well as large amounts of energy (50--60 GWh to train GPT-4). Despite these costs, these models often hallucinate, a characteristic that prevents them from being deployed in critical application domains. In contrast, the human brain consumes only 20~W of power. What is needed is the next level of AI evolution in which lightweight domain-specific multimodal models with higher levels of intelligence can reason, plan, and make decisions in dynamic environments with real-time data and prior knowledge, while learning continuously and evolving in ways that enhance future decision-making capability. This will define the next wave of AI, progressing from today's large models, trained with vast amounts of data, to nimble energy-efficient domain-specific agents that can reason and think in a world full of uncertainty. To support such agents, hardware will need to be reimagined to allow energy efficiencies greater than 1000x over the state of the art. Such a vision of future AI systems is developed in this work.

Dennis Menn, Feng Liang, Diana Marculescu

Artifact detectors have been shown to enhance the performance of image-generative models by serving as reward models during fine-tuning. These detectors enable the generative model to improve overall output fidelity and aesthetics. However, training the artifact detector requires expensive pixel-level human annotations that specify the artifact regions. The lack of annotated data limits the performance of the artifact detector. A naive pseudo-labeling approach-training a weak detector and using it to annotate unlabeled images-suffers from noisy labels, resulting in poor performance. To address this, we propose an artifact corruption pipeline that automatically injects artifacts into clean, high-quality synthetic images on a predetermined region, thereby producing pixel-level annotations without manual labeling. The proposed method enables training of an artifact detector that achieves performance improvements of 13.2% for ConvNeXt and 3.7% for Swin-T, as verified on human-labeled data, compared to baseline approaches. This work represents an initial step toward scalable pixel-level artifact annotation datasets that integrate world knowledge into artifact detection.

Natalia Frumkin, Diana Marculescu

Text-to-image diffusion models are computationally intensive, often requiring dozens of forward passes through large transformer backbones. For instance, Stable Diffusion XL generates high-quality images with 50 evaluations of a 2.6B-parameter model, an expensive process even for a single batch. Few-step diffusion models reduce this cost to 2-8 denoising steps but still depend on large, uncompressed U-Net or diffusion transformer backbones, which are often too costly for full-precision inference without datacenter GPUs. These requirements also limit existing post-training quantization methods that rely on full-precision calibration. We introduce Q-Sched, a new paradigm for post-training quantization that modifies the diffusion model scheduler rather than model weights. By adjusting the few-step sampling trajectory, Q-Sched achieves full-precision accuracy with a 4x reduction in model size. To learn quantization-aware pre-conditioning coefficients, we propose the JAQ loss, which combines text-image compatibility with an image quality metric for fine-grained optimization. JAQ is reference-free and requires only a handful of calibration prompts, avoiding full-precision inference during calibration. Q-Sched delivers substantial gains: a 15.5% FID improvement over the FP16 4-step Latent Consistency Model and a 16.6% improvement over the FP16 8-step Phased Consistency Model, showing that quantization and few-step distillation are complementary for high-fidelity generation. A large-scale user study with more than 80,000 annotations further confirms Q-Sched's effectiveness on both FLUX.1[schnell] and SDXL-Turbo.

Tien-Yu Chi, Hung-Yueh Chiang, Diana Marculescu et al.

State space models (SSMs) reduce the quadratic complexity of transformers by leveraging linear recurrence. Recently, VMamba has emerged as a strong SSM-based vision backbone, yet remains bottlenecked by spatial redundancy in its four-directional scan. We propose QuarterMap, a post-training activation pruning method that removes redundant spatial activations before scanning and restores dimensions via nearest-neighbor upsampling. Our method improves throughput without retraining. On ImageNet-1K, QuarterMap achieves up to 11% speedup on VMamba with less than 0.9% accuracy drop, and yields similar gains on ADE20K segmentation. Beyond VMamba, we validate QuarterMap on MedMamba, a domain-specific model that shares the same four-directional scanning structure, where it consistently improves throughput while preserving accuracy across multiple medical imaging tasks. Compared to token merging methods like ToMe, QuarterMap is tailored for SSMs and avoids costly merge-unmerge operations. Our method offers a plug-and-play tool for deployment-time efficiency without compromising transferability.

Tanvir Mahmud, Shentong Mo, Yapeng Tian et al.

Recent advances in pre-trained vision transformers have shown promise in parameter-efficient audio-visual learning without audio pre-training. However, few studies have investigated effective methods for aligning multimodal features in parameter-efficient audio-visual transformers. In this paper, we propose MA-AVT, a new parameter-efficient audio-visual transformer employing deep modality alignment for corresponding multimodal semantic features. Specifically, we introduce joint unimodal and multimodal token learning for aligning the two modalities with a frozen modality-shared transformer. This allows the model to learn separate representations for each modality, while also attending to the cross-modal relationships between them. In addition, unlike prior work that only aligns coarse features from the output of unimodal encoders, we introduce blockwise contrastive learning to align coarse-to-fine-grain hierarchical features throughout the encoding phase. Furthermore, to suppress the background features in each modality from foreground matched audio-visual features, we introduce a robust discriminative foreground mining scheme. Through extensive experiments on benchmark AVE, VGGSound, and CREMA-D datasets, we achieve considerable performance improvements over SOTA methods.

Tanvir Mahmud, Chun-Hao Liu, Burhaneddin Yaman et al.

Despite significant progress in semi-supervised learning for image object detection, several key issues are yet to be addressed for video object detection: (1) Achieving good performance for supervised video object detection greatly depends on the availability of annotated frames. (2) Despite having large inter-frame correlations in a video, collecting annotations for a large number of frames per video is expensive, time-consuming, and often redundant. (3) Existing semi-supervised techniques on static images can hardly exploit the temporal motion dynamics inherently present in videos. In this paper, we introduce SSVOD, an end-to-end semi-supervised video object detection framework that exploits motion dynamics of videos to utilize large-scale unlabeled frames with sparse annotations. To selectively assemble robust pseudo-labels across groups of frames, we introduce \textit{flow-warped predictions} from nearby frames for temporal-consistency estimation. In particular, we introduce cross-IoU and cross-divergence based selection methods over a set of estimated predictions to include robust pseudo-labels for bounding boxes and class labels, respectively. To strike a balance between confirmation bias and uncertainty noise in pseudo-labels, we propose confidence threshold based combination of hard and soft pseudo-labels. Our method achieves significant performance improvements over existing methods on ImageNet-VID, Epic-KITCHENS, and YouTube-VIS datasets. Code and pre-trained models will be released.

Ting-Wu Chin, Diana Marculescu, Ari S. Morcos

Optimizing the channel counts for different layers of a CNN has shown great promise in improving the efficiency of CNNs at test-time. However, these methods often introduce large computational overhead (e.g., an additional 2x FLOPs of standard training). Minimizing this overhead could therefore significantly speed up training. In this work, we propose width transfer, a technique that harnesses the assumptions that the optimized widths (or channel counts) are regular across sizes and depths. We show that width transfer works well across various width optimization algorithms and networks. Specifically, we can achieve up to 320x reduction in width optimization overhead without compromising the top-1 accuracy on ImageNet, making the additional cost of width optimization negligible relative to initial training. Our findings not only suggest an efficient way to conduct width optimization but also highlight that the widths that lead to better accuracy are invariant to various aspects of network architectures and training data.

Ahmet Inci, Mehmet Meric Isgenc, Diana Marculescu

Non-volatile memory (NVM) technologies such as spin-transfer torque magnetic random access memory (STT-MRAM) and spin-orbit torque magnetic random access memory (SOT-MRAM) have significant advantages compared to conventional SRAM due to their non-volatility, higher cell density, and scalability features. While previous work has investigated several architectural implications of NVM for generic applications, in this work we present DeepNVM++, a framework to characterize, model, and analyze NVM-based caches in GPU architectures for deep learning (DL) applications by combining technology-specific circuit-level models and the actual memory behavior of various DL workloads. We present both iso-capacity and iso-area performance and energy analysis for systems whose last-level caches rely on conventional SRAM and emerging STT-MRAM and SOT-MRAM technologies. In the iso-capacity case, STT-MRAM and SOT-MRAM provide up to 3.8x and 4.7x energy-delay product (EDP) reduction and 2.4x and 2.8x area reduction compared to conventional SRAM, respectively. Under iso-area assumptions, STT-MRAM and SOT-MRAM provide up to 2x and 2.3x EDP reduction and accommodate 2.3x and 3.3x cache capacity when compared to SRAM, respectively. We also perform a scalability analysis and show that STT-MRAM and SOT-MRAM achieve orders of magnitude EDP reduction when compared to SRAM for large cache capacities. Our comprehensive cross-layer framework is demonstrated on STT-/SOT-MRAM technologies and can be used for the characterization, modeling, and analysis of any NVM technology for last-level caches in GPUs for DL applications.

Ahmet Inci, Evgeny Bolotin, Yaosheng Fu et al.

With deep reinforcement learning (RL) methods achieving results that exceed human capabilities in games, robotics, and simulated environments, continued scaling of RL training is crucial to its deployment in solving complex real-world problems. However, improving the performance scalability and power efficiency of RL training through understanding the architectural implications of CPU-GPU systems remains an open problem. In this work we investigate and improve the performance and power efficiency of distributed RL training on CPU-GPU systems by approaching the problem not solely from the GPU microarchitecture perspective but following a holistic system-level analysis approach. We quantify the overall hardware utilization on a state-of-the-art distributed RL training framework and empirically identify the bottlenecks caused by GPU microarchitectural, algorithmic, and system-level design choices. We show that the GPU microarchitecture itself is well-balanced for state-of-the-art RL frameworks, but further investigation reveals that the number of actors running the environment interactions and the amount of hardware resources available to them are the primary performance and power efficiency limiters. To this end, we introduce a new system design metric, CPU/GPU ratio, and show how to find the optimal balance between CPU and GPU resources when designing scalable and efficient CPU-GPU systems for RL training.

Ting-Wu Chin, Pierce I-Jen Chuang, Vikas Chandra et al.

Weight quantization for deep ConvNets has shown promising results for applications such as image classification and semantic segmentation and is especially important for applications where memory storage is limited. However, when aiming for quantization without accuracy degradation, different tasks may end up with different bitwidths. This creates complexity for software and hardware support and the complexity accumulates when one considers mixed-precision quantization, in which case each layer's weights use a different bitwidth. Our key insight is that optimizing for the least bitwidth subject to no accuracy degradation is not necessarily an optimal strategy. This is because one cannot decide optimality between two bitwidths if one has a smaller model size while the other has better accuracy. In this work, we take the first step to understand if some weight bitwidth is better than others by aligning all to the same model size using a width-multiplier. Under this setting, somewhat surprisingly, we show that using a single bitwidth for the whole network can achieve better accuracy compared to mixed-precision quantization targeting zero accuracy degradation when both have the same model size. In particular, our results suggest that when the number of channels becomes a target hyperparameter, a single weight bitwidth throughout the network shows superior results for model compression.