Google, Inc
Abstract:In this paper, we tackle the critical challenge of compressing large language models (LLMs) to facilitate their practical deployment and broader adoption. We introduce a novel post-training compression paradigm that focuses on low-rank decomposition of LLM weights. Our analysis identifies two main challenges in this task: the variability in LLM activation distributions and handling unseen activations from different datasets and models. To address these challenges, we propose a nested activation-aware framework (NSVD) for LLMs, a training-free approach designed to enhance the accuracy of low-rank decompositions by managing activation outliers through transforming the weight matrix based on activation distribution and the original weight matrix. This method allows for the absorption of outliers into the transformed weight matrix, improving decomposition accuracy. Our comprehensive evaluation across eight datasets and six models from three distinct LLM families demonstrates the superiority of NSVD over current state-of-the-art methods, especially at medium to large compression ratios or in multilingual and multitask settings.
Abstract:Diffusion models for super-resolution (SR) produce high-quality visual results but require expensive computational costs. Despite the development of several methods to accelerate diffusion-based SR models, some (e.g., SinSR) fail to produce realistic perceptual details, while others (e.g., OSEDiff) may hallucinate non-existent structures. To overcome these issues, we present RSD, a new distillation method for ResShift, one of the top diffusion-based SR models. Our method is based on training the student network to produce such images that a new fake ResShift model trained on them will coincide with the teacher model. RSD achieves single-step restoration and outperforms the teacher by a large margin. We show that our distillation method can surpass the other distillation-based method for ResShift - SinSR - making it on par with state-of-the-art diffusion-based SR distillation methods. Compared to SR methods based on pre-trained text-to-image models, RSD produces competitive perceptual quality, provides images with better alignment to degraded input images, and requires fewer parameters and GPU memory. We provide experimental results on various real-world and synthetic datasets, including RealSR, RealSet65, DRealSR, ImageNet, and DIV2K.
Abstract:Learning diffusion bridge models is easy; making them fast and practical is an art. Diffusion bridge models (DBMs) are a promising extension of diffusion models for applications in image-to-image translation. However, like many modern diffusion and flow models, DBMs suffer from the problem of slow inference. To address it, we propose a novel distillation technique based on the inverse bridge matching formulation and derive the tractable objective to solve it in practice. Unlike previously developed DBM distillation techniques, the proposed method can distill both conditional and unconditional types of DBMs, distill models in a one-step generator, and use only the corrupted images for training. We evaluate our approach for both conditional and unconditional types of bridge matching on a wide set of setups, including super-resolution, JPEG restoration, sketch-to-image, and other tasks, and show that our distillation technique allows us to accelerate the inference of DBMs from 4x to 100x and even provide better generation quality than used teacher model depending on particular setup.
Abstract:Recent advances in deep learning have shown that learning robust feature representations is critical for the success of many computer vision tasks, including medical image segmentation. In particular, both transformer and convolutional-based architectures have benefit from leveraging pretext tasks for pretraining. However, the adoption of pretext tasks in 3D medical imaging has been less explored and remains a challenge, especially in the context of learning generalizable feature representations. We propose a novel pretraining strategy using diffusion models with anatomical guidance, tailored to the intricacies of 3D medical image data. We introduce an auxiliary diffusion process to pretrain a model that produce generalizable feature representations, useful for a variety of downstream segmentation tasks. We employ an additional model that predicts 3D universal body-part coordinates, providing guidance during the diffusion process and improving spatial awareness in generated representations. This approach not only aids in resolving localization inaccuracies but also enriches the model's ability to understand complex anatomical structures. Empirical validation on a 13-class organ segmentation task demonstrate the effectiveness of our pretraining technique. It surpasses existing restorative pretraining methods in 3D medical image segmentation by $7.5\%$, and is competitive with the state-of-the-art contrastive pretraining approach, achieving an average Dice coefficient of 67.8 in a non-linear evaluation scenario.
Abstract:This paper presents a novel approach to binary classification using dynamic logistic ensemble models. The proposed method addresses the challenges posed by datasets containing inherent internal clusters that lack explicit feature-based separations. By extending traditional logistic regression, we develop an algorithm that automatically partitions the dataset into multiple subsets, constructing an ensemble of logistic models to enhance classification accuracy. A key innovation in this work is the recursive probability calculation, derived through algebraic manipulation and mathematical induction, which enables scalable and efficient model construction. Compared to traditional ensemble methods such as Bagging and Boosting, our approach maintains interpretability while offering competitive performance. Furthermore, we systematically employ maximum likelihood and cost functions to facilitate the analytical derivation of recursive gradients as functions of ensemble depth. The effectiveness of the proposed approach is validated on a custom dataset created by introducing noise and shifting data to simulate group structures, resulting in significant performance improvements with layers. Implemented in Python, this work balances computational efficiency with theoretical rigor, providing a robust and interpretable solution for complex classification tasks with broad implications for machine learning applications. Code at https://github.com/ensemble-art/Dynamic-Logistic-Ensembles
Abstract:The Iterative Markovian Fitting (IMF) procedure based on iterative reciprocal and Markovian projections has recently been proposed as a powerful method for solving the Schr\"odinger Bridge problem. However, it has been observed that for the practical implementation of this procedure, it is crucial to alternate between fitting a forward and backward time diffusion at each iteration. Such implementation is thought to be a practical heuristic, which is required to stabilize training and obtain good results in applications such as unpaired domain translation. In our work, we show that this heuristic closely connects with the pioneer approaches for the Schr\"odinger Bridge based on the Iterative Proportional Fitting (IPF) procedure. Namely, we find that the practical implementation of IMF is, in fact, a combination of IMF and IPF procedures, and we call this combination the Iterative Proportional Markovian Fitting (IPMF) procedure. We show both theoretically and practically that this combined IPMF procedure can converge under more general settings, thus, showing that the IPMF procedure opens a door towards developing a unified framework for solving Schr\"odinger Bridge problems.
Abstract:This paper presents an approach to improve text embedding models through contrastive fine-tuning on small datasets augmented with expert scores. It focuses on enhancing semantic textual similarity tasks and addressing text retrieval problems. The proposed method uses soft labels derived from expert-augmented scores to fine-tune embedding models, preserving their versatility and ensuring retrieval capability is improved. The paper evaluates the method using a Q\&A dataset from an online shopping website and eight expert models. Results show improved performance over a benchmark model across multiple metrics on various retrieval tasks from the massive text embedding benchmark (MTEB). The method is cost-effective and practical for real-world applications, especially when labeled data is scarce.
Abstract:Medical procedures are an essential part of healthcare delivery, and the acquisition of procedural skills is a critical component of medical education. Unfortunately, procedural skill is not evenly distributed among medical providers. Skills may vary within departments or institutions, and across geographic regions, depending on the provider's training and ongoing experience. We present a mixed reality real-time communication system to increase access to procedural skill training and to improve remote emergency assistance. Our system allows a remote expert to guide a local operator through a medical procedure. RGBD cameras capture a volumetric view of the local scene including the patient, the operator, and the medical equipment. The volumetric capture is augmented onto the remote expert's view to allow the expert to spatially guide the local operator using visual and verbal instructions. We evaluated our mixed reality communication system in a study in which experts teach the ultrasound-guided placement of a central venous catheter (CVC) to students in a simulation setting. The study compares state-of-the-art video communication against our system. The results indicate that our system enhances and offers new possibilities for visual communication compared to video teleconference-based training.
Abstract:Recently, several approaches have emerged for generating neural representations with multiple levels of detail (LODs). LODs can improve the rendering by using lower resolutions and smaller model sizes when appropriate. However, existing methods generally focus on a few discrete LODs which suffer from aliasing and flicker artifacts as details are changed and limit their granularity for adapting to resource limitations. In this paper, we propose a method to encode light field networks with continuous LODs, allowing for finely tuned adaptations to rendering conditions. Our training procedure uses summed-area table filtering allowing efficient and continuous filtering at various LODs. Furthermore, we use saliency-based importance sampling which enables our light field networks to distribute their capacity, particularly limited at lower LODs, towards representing the details viewers are most likely to focus on. Incorporating continuous LODs into neural representations enables progressive streaming of neural representations, decreasing the latency and resource utilization for rendering.
Abstract:Neural representations have shown great promise in their ability to represent radiance and light fields while being very compact compared to the image set representation. However, current representations are not well suited for streaming as decoding can only be done at a single level of detail and requires downloading the entire neural network model. Furthermore, high-resolution light field networks can exhibit flickering and aliasing as neural networks are sampled without appropriate filtering. To resolve these issues, we present a progressive multi-scale light field network that encodes a light field with multiple levels of detail. Lower levels of detail are encoded using fewer neural network weights enabling progressive streaming and reducing rendering time. Our progressive multi-scale light field network addresses aliasing by encoding smaller anti-aliased representations at its lower levels of detail. Additionally, per-pixel level of detail enables our representation to support dithered transitions and foveated rendering.