U-Net style networks are commonly utilized in unsupervised image registration to predict dense displacement fields, which for high-resolution volumetric image data is a resource-intensive and time-consuming task. To tackle this challenge, we first propose Fourier-Net, which replaces the costly U-Net style expansive path with a parameter-free model-driven decoder. Instead of directly predicting a full-resolution displacement field, our Fourier-Net learns a low-dimensional representation of the displacement field in the band-limited Fourier domain which our model-driven decoder converts to a full-resolution displacement field in the spatial domain. Expanding upon Fourier-Net, we then introduce Fourier-Net+, which additionally takes the band-limited spatial representation of the images as input and further reduces the number of convolutional layers in the U-Net style network's contracting path. Finally, to enhance the registration performance, we propose a cascaded version of Fourier-Net+. We evaluate our proposed methods on three datasets, on which our proposed Fourier-Net and its variants achieve comparable results with current state-of-the art methods, while exhibiting faster inference speeds, lower memory footprint, and fewer multiply-add operations. With such small computational cost, our Fourier-Net+ enables the efficient training of large-scale 3D registration on low-VRAM GPUs. Our code is publicly available at \url{https://github.com/xi-jia/Fourier-Net}.
Recently, U-shaped networks have dominated the field of medical image segmentation due to their simple and easily tuned structure. However, existing U-shaped segmentation networks: 1) mostly focus on designing complex self-attention modules to compensate for the lack of long-term dependence based on convolution operation, which increases the overall number of parameters and computational complexity of the network; 2) simply fuse the features of encoder and decoder, ignoring the connection between their spatial locations. In this paper, we rethink the above problem and build a lightweight medical image segmentation network, called SegNetr. Specifically, we introduce a novel SegNetr block that can perform local-global interactions dynamically at any stage and with only linear complexity. At the same time, we design a general information retention skip connection (IRSC) to preserve the spatial location information of encoder features and achieve accurate fusion with the decoder features. We validate the effectiveness of SegNetr on four mainstream medical image segmentation datasets, with 59\% and 76\% fewer parameters and GFLOPs than vanilla U-Net, while achieving segmentation performance comparable to state-of-the-art methods. Notably, the components proposed in this paper can be applied to other U-shaped networks to improve their segmentation performance.
The recent surge of large language models (LLMs) highlights their ability to perform in-context learning, i.e., "learning" to perform a task from a few demonstrations in the context without any parameter updates. However, their capabilities of in-context learning are limited by the model architecture: 1) the use of demonstrations is constrained by a maximum sentence length due to positional embeddings; 2) the quadratic complexity of attention hinders users from using more demonstrations efficiently; 3) LLMs are shown to be sensitive to the order of the demonstrations. In this work, we tackle these challenges by proposing a better architectural design for in-context learning. We propose SAICL (Structured Attention for In-Context Learning), which replaces the full-attention by a structured attention mechanism designed for in-context learning, and removes unnecessary dependencies between individual demonstrations, while making the model invariant to the permutation of demonstrations. We evaluate SAICL in a meta-training framework and show that SAICL achieves comparable or better performance than full attention while obtaining up to 3.4x inference speed-up. SAICL also consistently outperforms a strong Fusion-in-Decoder (FiD) baseline which processes each demonstration independently. Finally, thanks to its linear nature, we demonstrate that SAICL can easily scale to hundreds of demonstrations with continuous performance gains with scaling.
Although unsupervised domain adaptation (UDA) is a promising direction to alleviate domain shift, they fall short of their supervised counterparts. In this work, we investigate relatively less explored semi-supervised domain adaptation (SSDA) for medical image segmentation, where access to a few labeled target samples can improve the adaptation performance substantially. Specifically, we propose a two-stage training process. First, an encoder is pre-trained in a self-learning paradigm using a novel domain-content disentangled contrastive learning (CL) along with a pixel-level feature consistency constraint. The proposed CL enforces the encoder to learn discriminative content-specific but domain-invariant semantics on a global scale from the source and target images, whereas consistency regularization enforces the mining of local pixel-level information by maintaining spatial sensitivity. This pre-trained encoder, along with a decoder, is further fine-tuned for the downstream task, (i.e. pixel-level segmentation) using a semi-supervised setting. Furthermore, we experimentally validate that our proposed method can easily be extended for UDA settings, adding to the superiority of the proposed strategy. Upon evaluation on two domain adaptive image segmentation tasks, our proposed method outperforms the SoTA methods, both in SSDA and UDA settings. Code is available at https://github.com/hritam-98/GFDA-disentangled
Recently, the release of INSTRUCTEVAL has provided valuable insights into the performance of large language models (LLMs) that utilize encoder-decoder or decoder-only architecture. Interestingly, despite being introduced four years ago, T5-based LLMs, such as FLAN-T5, continue to outperform the latest decoder-based LLMs, such as LLAMA and VICUNA, on tasks that require general problem-solving skills. This performance discrepancy can be attributed to three key factors: (1) Pre-training data, (2) Backbone architecture, and (3) Instruction dataset. In this technical report, our main focus is on investigating the impact of the third factor by leveraging VICUNA, a large language model based on LLAMA, which has undergone fine-tuning on ChatGPT conversations. To achieve this objective, we fine-tuned VICUNA using a customized instruction dataset collection called FLANMINI. This collection includes a subset of the large-scale instruction dataset known as FLAN, as well as various code-related datasets and conversational datasets derived from ChatGPT/GPT-4. This dataset comprises a large number of tasks that demand problem-solving skills. Our experimental findings strongly indicate that the enhanced problem-solving abilities of our model, FLACUNA, are obtained through fine-tuning VICUNA on the FLAN dataset, leading to significant improvements across numerous benchmark datasets in INSTRUCTEVAL. FLACUNA is publicly available at https://huggingface.co/declare-lab/flacuna-13b-v1.0.
Protoform reconstruction is the task of inferring what morphemes or words appeared like in the ancestral languages of a set of daughter languages. Meloni et al. (2021) achieved the state-of-the-art on Latin protoform reconstruction with an RNN-based encoder-decoder with attention model. We update their model with the state-of-the-art seq2seq model: the Transformer. Our model outperforms their model on a suite of different metrics on two different datasets: their Romance data of 8,000 cognates spanning 5 languages and a Chinese dataset (Hou 2004) of 800+ cognates spanning 39 varieties. We also probe our model for potential phylogenetic signal contained in the model. Our code is publicly available at https://github.com/cmu-llab/acl-2023.
In this report, we describe the technical details of our approach for the Ego4D Long-Term Action Anticipation Challenge 2023. The aim of this task is to predict a sequence of future actions that will take place at an arbitrary time or later, given an input video. To accomplish this task, we introduce three improvements to the baseline model, which consists of an encoder that generates clip-level features from the video, an aggregator that integrates multiple clip-level features, and a decoder that outputs Z future actions. 1) Model ensemble of SlowFast and SlowFast-CLIP; 2) Label smoothing to relax order constraints for future actions; 3) Constraining the prediction of the action class (verb, noun) based on word co-occurrence. Our method outperformed the baseline performance and recorded as second place solution on the public leaderboard.
Limited by imaging systems, the reconstruction of Magnetic Resonance Imaging (MRI) images from partial measurement is essential to medical imaging research. Benefiting from the diverse and complementary information of multi-contrast MR images in different imaging modalities, multi-contrast Super-Resolution (SR) reconstruction is promising to yield SR images with higher quality. In the medical scenario, to fully visualize the lesion, radiologists are accustomed to zooming the MR images at arbitrary scales rather than using a fixed scale, as used by most MRI SR methods. In addition, existing multi-contrast MRI SR methods often require a fixed resolution for the reference image, which makes acquiring reference images difficult and imposes limitations on arbitrary scale SR tasks. To address these issues, we proposed an implicit neural representations based dual-arbitrary multi-contrast MRI super-resolution method, called Dual-ArbNet. First, we decouple the resolution of the target and reference images by a feature encoder, enabling the network to input target and reference images at arbitrary scales. Then, an implicit fusion decoder fuses the multi-contrast features and uses an Implicit Decoding Function~(IDF) to obtain the final MRI SR results. Furthermore, we introduce a curriculum learning strategy to train our network, which improves the generalization and performance of our Dual-ArbNet. Extensive experiments in two public MRI datasets demonstrate that our method outperforms state-of-the-art approaches under different scale factors and has great potential in clinical practice.
Recent transformer-based methods achieve notable gains in the Human-object Interaction Detection (HOID) task by leveraging the detection of DETR and the prior knowledge of Vision-Language Model (VLM). However, these methods suffer from extended training times and complex optimization due to the entanglement of object detection and HOI recognition during the decoding process. Especially, the query embeddings used to predict both labels and boxes suffer from ambiguous representations, and the gap between the prediction of HOI labels and verb labels is not considered. To address these challenges, we introduce SOV-STG-VLA with three key components: Subject-Object-Verb (SOV) decoding, Specific Target Guided (STG) denoising, and a Vision-Language Advisor (VLA). Our SOV decoders disentangle object detection and verb recognition with a novel interaction region representation. The STG denoising strategy learns label embeddings with ground-truth information to guide the training and inference. Our SOV-STG achieves a fast convergence speed and high accuracy and builds a foundation for the VLA to incorporate the prior knowledge of the VLM. We introduce a vision advisor decoder to fuse both the interaction region information and the VLM's vision knowledge and a Verb-HOI prediction bridge to promote interaction representation learning. Our VLA notably improves our SOV-STG and achieves SOTA performance with one-sixth of training epochs compared to recent SOTA. Code and models are available at https://github.com/cjw2021/SOV-STG-VLA
In computer-assisted orthodontics, three-dimensional tooth models are required for many medical treatments. Tooth segmentation from cone-beam computed tomography (CBCT) images is a crucial step in constructing the models. However, CBCT image quality problems such as metal artifacts and blurring caused by shooting equipment and patients' dental conditions make the segmentation difficult. In this paper, we propose ToothSegNet, a new framework which acquaints the segmentation model with generated degraded images during training. ToothSegNet merges the information of high and low quality images from the designed degradation simulation module using channel-wise cross fusion to reduce the semantic gap between encoder and decoder, and also refines the shape of tooth prediction through a structural constraint loss. Experimental results suggest that ToothSegNet produces more precise segmentation and outperforms the state-of-the-art medical image segmentation methods.
Sequence modeling approaches have shown promising results in robot imitation learning. Recently, diffusion models have been adopted for behavioral cloning in a sequence modeling fashion, benefiting from their exceptional capabilities in modeling complex data distributions. The standard diffusion-based policy iteratively generates action sequences from random noise conditioned on the input states. Nonetheless, the model for diffusion policy can be further improved in terms of visual representations. In this work, we propose Crossway Diffusion, a simple yet effective method to enhance diffusion-based visuomotor policy learning via a carefully designed state decoder and an auxiliary self-supervised learning (SSL) objective. The state decoder reconstructs raw image pixels and other state information from the intermediate representations of the reverse diffusion process. The whole model is jointly optimized by the SSL objective and the original diffusion loss. Our experiments demonstrate the effectiveness of Crossway Diffusion in various simulated and real-world robot tasks, confirming its consistent advantages over the standard diffusion-based policy and substantial improvements over the baselines.
LiDAR scanning for surveying applications acquire measurements over wide areas and long distances, which produces large-scale 3D point clouds with significant local density variations. While existing 3D semantic segmentation models conduct downsampling and upsampling to build robustness against varying point densities, they are less effective under the large local density variations characteristic of point clouds from surveying applications. To alleviate this weakness, we propose a novel architecture called HDVNet that contains a nested set of encoder-decoder pathways, each handling a specific point density range. Limiting the interconnections between the feature maps enables HDVNet to gauge the reliability of each feature based on the density of a point, e.g., downweighting high density features not existing in low density objects. By effectively handling input density variations, HDVNet outperforms state-of-the-art models in segmentation accuracy on real point clouds with inconsistent density, using just over half the weights.
The generative autoencoders, such as the variational autoencoders or the adversarial autoencoders, have achieved great success in lots of real-world applications, including image generation, and signal communication. However, little concern has been devoted to their robustness during practical deployment. Due to the probabilistic latent structure, variational autoencoders (VAEs) may confront problems such as a mismatch between the posterior distribution of the latent and real data manifold, or discontinuity in the posterior distribution of the latent. This leaves a back door for malicious attackers to collapse VAEs from the latent space, especially in scenarios where the encoder and decoder are used separately, such as communication and compressed sensing. In this work, we provide the first study on the adversarial robustness of generative autoencoders in the latent space. Specifically, we empirically demonstrate the latent vulnerability of popular generative autoencoders through attacks in the latent space. We also evaluate the difference between variational autoencoders and their deterministic variants and observe that the latter performs better in latent robustness. Meanwhile, we identify a potential trade-off between the adversarial robustness and the degree of the disentanglement of the latent codes. Additionally, we also verify the feasibility of improvement for the latent robustness of VAEs through adversarial training. In summary, we suggest concerning the adversarial latent robustness of the generative autoencoders, analyze several robustness-relative issues, and give some insights into a series of key challenges.
Recently, video summarization has been proposed as a method to help video exploration. However, traditional video summarization models only generate a fixed video summary which is usually independent of user-specific needs and hence limits the effectiveness of video exploration. Multi-modal video summarization is one of the approaches utilized to address this issue. Multi-modal video summarization has a video input and a text-based query input. Hence, effective modeling of the interaction between a video input and text-based query is essential to multi-modal video summarization. In this work, a new causality-based method named Causal Video Summarizer (CVS) is proposed to effectively capture the interactive information between the video and query to tackle the task of multi-modal video summarization. The proposed method consists of a probabilistic encoder and a probabilistic decoder. Based on the evaluation of the existing multi-modal video summarization dataset, experimental results show that the proposed approach is effective with the increase of +5.4% in accuracy and +4.92% increase of F 1- score, compared with the state-of-the-art method.
Unlocking the potential of deep learning in Peak-Hour Series Forecasting (PHSF) remains a critical yet underexplored task in various domains. While state-of-the-art deep learning models excel in regular Time Series Forecasting (TSF), they struggle to achieve comparable results in PHSF. This can be attributed to the challenges posed by the high degree of non-stationarity in peak-hour series, which makes direct forecasting more difficult than standard TSF. Additionally, manually extracting the maximum value from regular forecasting results leads to suboptimal performance due to models minimizing the mean deficit. To address these issues, this paper presents Seq2Peak, a novel framework designed specifically for PHSF tasks, bridging the performance gap observed in TSF models. Seq2Peak offers two key components: the CyclicNorm pipeline to mitigate the non-stationarity issue and a simple yet effective trainable-parameter-free peak-hour decoder with a hybrid loss function that utilizes both the original series and peak-hour series as supervised signals. Extensive experimentation on publicly available time series datasets demonstrates the effectiveness of the proposed framework, yielding a remarkable average relative improvement of 37.7% across four real-world datasets for both transformer- and non-transformer-based TSF models.
The effectiveness of ad creatives is greatly influenced by their visual appearance. Advertising platforms can generate ad creatives with different appearances by combining creative elements provided by advertisers. However, with the increasing number of ad creative elements, it becomes challenging to select a suitable combination from the countless possibilities. The industry's mainstream approach is to select individual creative elements independently, which often overlooks the importance of interaction between creative elements during the modeling process. In response, this paper proposes a Cross-Element Combinatorial Selection framework for multiple creative elements, termed CECS. In the encoder process, a cross-element interaction is adopted to dynamically adjust the expression of a single creative element based on the current candidate creatives. In the decoder process, the creative combination problem is transformed into a cascade selection problem of multiple creative elements. A pointer mechanism with a cascade design is used to model the associations among candidates. Comprehensive experiments on real-world datasets show that CECS achieved the SOTA score on offline metrics. Moreover, the CECS algorithm has been deployed in our industrial application, resulting in a significant 6.02% CTR and 10.37% GMV lift, which is beneficial to the business.
Visualization is a crucial step in exploratory data analysis. One possible approach is to train an autoencoder with low-dimensional latent space. Large network depth and width can help unfolding the data. However, such expressive networks can achieve low reconstruction error even when the latent representation is distorted. To avoid such misleading visualizations, we propose first a differential geometric perspective on the decoder, leading to insightful diagnostics for an embedding's distortion, and second a new regularizer mitigating such distortion. Our ``Geometric Autoencoder'' avoids stretching the embedding spuriously, so that the visualization captures the data structure more faithfully. It also flags areas where little distortion could not be achieved, thus guarding against misinterpretation.
Recent generative approaches for multi-hop question answering (QA) utilize the fusion-in-decoder method~\cite{izacard-grave-2021-leveraging} to generate a single sequence output which includes both a final answer and a reasoning path taken to arrive at that answer, such as passage titles and key facts from those passages. While such models can lead to better interpretability and high quantitative scores, they often have difficulty accurately identifying the passages corresponding to key entities in the context, resulting in incorrect passage hops and a lack of faithfulness in the reasoning path. To address this, we propose a single-sequence prediction method over a local reasoning graph (\model)\footnote{Code/Models will be released at \url{https://github.com/gowtham1997/SeqGraph}} that integrates a graph structure connecting key entities in each context passage to relevant subsequent passages for each question. We use a graph neural network to encode this graph structure and fuse the resulting representations into the entity representations of the model. Our experiments show significant improvements in answer exact-match/F1 scores and faithfulness of grounding in the reasoning path on the HotpotQA dataset and achieve state-of-the-art numbers on the Musique dataset with only up to a 4\% increase in model parameters.
Domain-generalized urban-scene semantic segmentation (USSS) aims to learn generalized semantic predictions across diverse urban-scene styles. Unlike domain gap challenges, USSS is unique in that the semantic categories are often similar in different urban scenes, while the styles can vary significantly due to changes in urban landscapes, weather conditions, lighting, and other factors. Existing approaches typically rely on convolutional neural networks (CNNs) to learn the content of urban scenes. In this paper, we propose a Content-enhanced Mask TransFormer (CMFormer) for domain-generalized USSS. The main idea is to enhance the focus of the fundamental component, the mask attention mechanism, in Transformer segmentation models on content information. To achieve this, we introduce a novel content-enhanced mask attention mechanism. It learns mask queries from both the image feature and its down-sampled counterpart, as lower-resolution image features usually contain more robust content information and are less sensitive to style variations. These features are fused into a Transformer decoder and integrated into a multi-resolution content-enhanced mask attention learning scheme. Extensive experiments conducted on various domain-generalized urban-scene segmentation datasets demonstrate that the proposed CMFormer significantly outperforms existing CNN-based methods for domain-generalized semantic segmentation, achieving improvements of up to 14.00\% in terms of mIoU (mean intersection over union). The source code is publicly available at \url{https://github.com/BiQiWHU/CMFormer}.
Filter pruning simultaneously accelerates the computation and reduces the memory overhead of CNNs, which can be effectively applied to edge devices and cloud services. In this paper, we propose a novel Knowledge-driven Differential Filter Sampler~(KDFS) with Masked Filter Modeling~(MFM) framework for filter pruning, which globally prunes the redundant filters based on the prior knowledge of a pre-trained model in a differential and non-alternative optimization. Specifically, we design a differential sampler with learnable sampling parameters to build a binary mask vector for each layer, determining whether the corresponding filters are redundant. To learn the mask, we introduce masked filter modeling to construct PCA-like knowledge by aligning the intermediate features from the pre-trained teacher model and the outputs of the student decoder taking sampling features as the input. The mask and sampler are directly optimized by the Gumbel-Softmax Straight-Through Gradient Estimator in an end-to-end manner in combination with global pruning constraint, MFM reconstruction error, and dark knowledge. Extensive experiments demonstrate the proposed KDFS's effectiveness in compressing the base models on various datasets. For instance, the pruned ResNet-50 on ImageNet achieves $55.36\%$ computation reduction, and $42.86\%$ parameter reduction, while only dropping $0.35\%$ Top-1 accuracy, significantly outperforming the state-of-the-art methods. The code is available at \url{https://github.com/Osilly/KDFS}.