The fast growing capabilities of large-scale deep learning models, such as Bert, GPT and ViT, are revolutionizing the landscape of NLP, CV and many other domains. Training such models, however, poses an unprecedented demand for computing power, which incurs exponentially increasing energy cost and carbon dioxide emissions. It is thus critical to develop efficient training solutions to reduce the training costs. Motivated by a set of key observations of inter- and intra-layer similarities among feature maps and attentions that can be identified from typical training processes, we propose a multi-level framework for training acceleration. Specifically, the framework is based on three basic operators, Coalescing, De-coalescing and Interpolation, which can be orchestrated to build a multi-level training framework. The framework consists of a V-cycle training process, which progressively down- and up-scales the model size and projects the parameters between adjacent levels of models via coalescing and de-coalescing. The key idea is that a smaller model that can be trained for fast convergence and the trained parameters provides high-qualities intermediate solutions for the next level larger network. The interpolation operator is designed to break the symmetry of neurons incurred by de-coalescing for better convergence performance. Our experiments on transformer-based language models (e.g. Bert, GPT) as well as a vision model (e.g. DeiT) prove that the proposed framework reduces the computational cost by about 20% on training BERT/GPT-Base models and up to 51.6% on training the BERT-Large model while preserving the performance.
Generating radiology reports automatically reduces the workload of radiologists and helps the diagnoses of specific diseases. Many existing methods take this task as modality transfer process. However, since the key information related to disease accounts for a small proportion in both image and report, it is hard for the model to learn the latent relation between the radiology image and its report, thus failing to generate fluent and accurate radiology reports. To tackle this problem, we propose a memory-based cross-modal semantic alignment model (MCSAM) following an encoder-decoder paradigm. MCSAM includes a well initialized long-term clinical memory bank to learn disease-related representations as well as prior knowledge for different modalities to retrieve and use the retrieved memory to perform feature consolidation. To ensure the semantic consistency of the retrieved cross modal prior knowledge, a cross-modal semantic alignment module (SAM) is proposed. SAM is also able to generate semantic visual feature embeddings which can be added to the decoder and benefits report generation. More importantly, to memorize the state and additional information while generating reports with the decoder, we use learnable memory tokens which can be seen as prompts. Extensive experiments demonstrate the promising performance of our proposed method which generates state-of-the-art performance on the MIMIC-CXR dataset.
Simultaneous functional PET/MR (sf-PET/MR) presents a cutting-edge multimodal neuroimaging technique. It provides an unprecedented opportunity for concurrently monitoring and integrating multifaceted brain networks built by spatiotemporally covaried metabolic activity, neural activity, and cerebral blood flow (perfusion). Albeit high scientific/clinical values, short in hardware accessibility of PET/MR hinders its applications, let alone modern AI-based PET/MR fusion models. Our objective is to develop a clinically feasible AI-based disease diagnosis model trained on comprehensive sf-PET/MR data with the power of, during inferencing, allowing single modality input (e.g., PET only) as well as enforcing multimodal-based accuracy. To this end, we propose MX-ARM, a multimodal MiXture-of-experts Alignment and Reconstruction Model. It is modality detachable and exchangeable, allocating different multi-layer perceptrons dynamically ("mixture of experts") through learnable weights to learn respective representations from different modalities. Such design will not sacrifice model performance in uni-modal situation. To fully exploit the inherent complex and nonlinear relation among modalities while producing fine-grained representations for uni-modal inference, we subsequently add a modal alignment module to line up a dominant modality (e.g., PET) with representations of auxiliary modalities (MR). We further adopt multimodal reconstruction to promote the quality of learned features. Experiments on precious multimodal sf-PET/MR data for Mild Cognitive Impairment diagnosis showcase the efficacy of our model toward clinically feasible precision medicine.
Spiking Transformers, which integrate Spiking Neural Networks (SNNs) with Transformer architectures, have attracted significant attention due to their potential for energy efficiency and high performance. However, existing models in this domain still suffer from suboptimal performance. We introduce several innovations to improve the performance: i) We propose a novel spike-form Q-K attention mechanism, tailored for SNNs, which efficiently models the importance of token or channel dimensions through binary vectors with linear complexity. ii) We incorporate the hierarchical structure, which significantly benefits the performance of both the brain and artificial neural networks, into spiking transformers to obtain multi-scale spiking representation. iii) We design a versatile and powerful patch embedding module with a deformed shortcut specifically for spiking transformers. Together, we develop QKFormer, a hierarchical spiking transformer based on Q-K attention with direct training. QKFormer shows significantly superior performance over existing state-of-the-art SNN models on various mainstream datasets. Notably, with comparable size to Spikformer (66.34 M, 74.81%), QKFormer (64.96 M) achieves a groundbreaking top-1 accuracy of 85.65% on ImageNet-1k, substantially outperforming Spikformer by 10.84%. To our best knowledge, this is the first time that directly training SNNs have exceeded 85% accuracy on ImageNet-1K. The code and models are publicly available at https://github.com/zhouchenlin2096/QKFormer
Automated X-ray image segmentation would accelerate research and development in diagnostic and interventional precision medicine. Prior efforts have contributed task-specific models capable of solving specific image analysis problems, but the utility of these models is restricted to their particular task domain, and expanding to broader use requires additional data, labels, and retraining efforts. Recently, foundation models (FMs) -- machine learning models trained on large amounts of highly variable data thus enabling broad applicability -- have emerged as promising tools for automated image analysis. Existing FMs for medical image analysis focus on scenarios and modalities where objects are clearly defined by visually apparent boundaries, such as surgical tool segmentation in endoscopy. X-ray imaging, by contrast, does not generally offer such clearly delineated boundaries or structure priors. During X-ray image formation, complex 3D structures are projected in transmission onto the imaging plane, resulting in overlapping features of varying opacity and shape. To pave the way toward an FM for comprehensive and automated analysis of arbitrary medical X-ray images, we develop FluoroSAM, a language-aligned variant of the Segment-Anything Model, trained from scratch on 1.6M synthetic X-ray images. FluoroSAM is trained on data including masks for 128 organ types and 464 non-anatomical objects, such as tools and implants. In real X-ray images of cadaveric specimens, FluoroSAM is able to segment bony anatomical structures based on text-only prompting with 0.51 and 0.79 DICE with point-based refinement, outperforming competing SAM variants for all structures. FluoroSAM is also capable of zero-shot generalization to segmenting classes beyond the training set thanks to its language alignment, which we demonstrate for full lung segmentation on real chest X-rays.
Multiple cameras can provide multi-view video coverage of a person. It is necessary to fuse multi-view data, e.g., for subsequent behavioral analysis, while such fusion often relies on calibration of cameras in traditional solutions. However, it is non-trivial to calibrate multiple cameras. In this work, we propose a method to reconstruct 3D human body from multiple uncalibrated camera views. First, we adopt a pre-trained human body encoder to process each individual camera view, such that human body models and parameters can be reconstructed for each view. Next, instead of simply averaging models across views, we train a network to determine the weights of individual views for their fusion, based on the parameters estimated for joints and hands of human body as well as camera positions. Further, we turn to the mesh surface of human body for dynamic fusion, such that facial expression can be seamlessly integrated into the model of human body. Our method has demonstrated superior performance in reconstructing human body upon two public datasets. More importantly, our method can flexibly support ad-hoc deployment of an arbitrary number of cameras, which has significant potential in related applications. We will release source code upon acceptance of the paper.
This paper investigates a wireless-powered Internet of Things (IoT) network comprising a hybrid access point (HAP) and two devices. The HAP facilitates downlink wireless energy transfer (WET) for device charging and uplink wireless information transfer (WIT) to collect status updates from the devices. To keep the information fresh, concurrent WET and WIT are allowed, and orthogonal multiple access (OMA) and non-orthogonal multiple access (NOMA) are adaptively scheduled for WIT. Consequently, we formulate an expected weighted sum age of information (EWSAoI) minimization problem to adaptively schedule the transmission scheme, choosing from WET, OMA, NOMA, and WET+OMA, and to allocate transmit power. To address this, we reformulate the problem as a Markov decision process (MDP) and develop an optimal policy based on instantaneous AoI and remaining battery power to determine scheme selection and transmit power allocation. Extensive results demonstrate the effectiveness of the proposed policy, and the optimal policy has a distinct decision boundary-switching property, providing valuable insights for practical system design.
Reinforcement Learning from Human Feedback (RLHF) is commonly utilized to improve the alignment of Large Language Models (LLMs) with human preferences. Given the evolving nature of human preferences, continual alignment becomes more crucial and practical in comparison to traditional static alignment. Nevertheless, making RLHF compatible with Continual Learning (CL) is challenging due to its complex process. Meanwhile, directly learning new human preferences may lead to Catastrophic Forgetting (CF) of historical preferences, resulting in helpless or harmful outputs. To overcome these challenges, we propose the Continual Optimal Policy Regularization (COPR) method, which draws inspiration from the optimal policy theory. COPR utilizes a sampling distribution as a demonstration and regularization constraints for CL. It adopts the Lagrangian Duality (LD) method to dynamically regularize the current policy based on the historically optimal policy, which prevents CF and avoids over-emphasizing unbalanced objectives. We also provide formal proof for the learnability of COPR. The experimental results show that COPR outperforms strong CL baselines on our proposed benchmark, in terms of reward-based, GPT-4 evaluations and human assessment. Furthermore, we validate the robustness of COPR under various CL settings, including different backbones, replay memory sizes, and learning orders.
Image segmentation plays a crucial role in extracting important objects of interest from images, enabling various applications. While existing methods have shown success in segmenting clean images, they often struggle to produce accurate segmentation results when dealing with degraded images, such as those containing noise or occlusions. To address this challenge, interactive segmentation has emerged as a promising approach, allowing users to provide meaningful input to guide the segmentation process. However, an important problem in interactive segmentation lies in determining how to incorporate minimal yet meaningful user guidance into the segmentation model. In this paper, we propose the quasi-conformal interactive segmentation (QIS) model, which incorporates user input in the form of positive and negative clicks. Users mark a few pixels belonging to the object region as positive clicks, indicating that the segmentation model should include a region around these clicks. Conversely, negative clicks are provided on pixels belonging to the background, instructing the model to exclude the region near these clicks from the segmentation mask. Additionally, the segmentation mask is obtained by deforming a template mask with the same topology as the object of interest using an orientation-preserving quasiconformal mapping. This approach helps to avoid topological errors in the segmentation results. We provide a thorough analysis of the proposed model, including theoretical support for the ability of QIS to include or exclude regions of interest or disinterest based on the user's indication. To evaluate the performance of QIS, we conduct experiments on synthesized images, medical images, natural images and noisy natural images. The results demonstrate the efficacy of our proposed method.