Recent development in the field of Deep Learning have exposed the underlying vulnerability of Deep Neural Network (DNN) against adversarial examples. In image classification, an adversarial example is a carefully modified image that is visually imperceptible to the original image but can cause DNN model to misclassify it. Training the network with Gaussian noise is an effective technique to perform model regularization, thus improving model robustness against input variation. Inspired by this classical method, we explore to utilize the regularization characteristic of noise injection to improve DNN's robustness against adversarial attack. In this work, we propose Parametric-Noise-Injection (PNI) which involves trainable Gaussian noise injection at each layer on either activation or weights through solving the min-max optimization problem, embedded with adversarial training. These parameters are trained explicitly to achieve improved robustness. To the best of our knowledge, this is the first work that uses trainable noise injection to improve network robustness against adversarial attacks, rather than manually configuring the injected noise level through cross-validation. The extensive results show that our proposed PNI technique effectively improves the robustness against a variety of powerful white-box and black-box attacks such as PGD, C & W, FGSM, transferable attack and ZOO attack. Last but not the least, PNI method improves both clean- and perturbed-data accuracy in comparison to the state-of-the-art defense methods, which outperforms current unbroken PGD defense by 1.1 % and 6.8 % on clean test data and perturbed test data respectively using Resnet-20 architecture.
Deep object recognition models have been very successful over benchmark datasets such as ImageNet. How accurate and robust are they to distribution shifts arising from natural and synthetic variations in datasets? Prior research on this problem has primarily focused on ImageNet variations (e.g., ImageNetV2, ImageNet-A). To avoid potential inherited biases in these studies, we take a different approach. Specifically, we reanalyze the ObjectNet dataset recently proposed by Barbu et al. containing objects in daily life situations. They showed a dramatic performance drop of the state of the art object recognition models on this dataset. Due to the importance and implications of their results regarding the generalization ability of deep models, we take a second look at their analysis. We find that applying deep models to the isolated objects, rather than the entire scene as is done in the original paper, results in around 20-30% performance improvement. Relative to the numbers reported in Barbu et al., around 10-15% of the performance loss is recovered, without any test time data augmentation. Despite this gain, however, we conclude that deep models still suffer drastically on the ObjectNet dataset. We also investigate the robustness of models against synthetic image perturbations such as geometric transformations (e.g., scale, rotation, translation), natural image distortions (e.g., impulse noise, blur) as well as adversarial attacks (e.g., FGSM and PGD-5). Our results indicate that limiting the object area as much as possible (i.e., from the entire image to the bounding box to the segmentation mask) leads to consistent improvement in accuracy and robustness.
As two fundamental representation modalities of 3D objects, 3D point clouds and multi-view 2D images record shape information from different domains of geometric structures and visual appearances. In the current deep learning era, remarkable progress in processing such two data modalities has been achieved through respectively customizing compatible 3D and 2D network architectures. However, unlike multi-view image-based 2D visual modeling paradigms, which have shown leading performance in several common 3D shape recognition benchmarks, point cloud-based 3D geometric modeling paradigms are still highly limited by insufficient learning capacity, due to the difficulty of extracting discriminative features from irregular geometric signals. In this paper, we explore the possibility of boosting deep 3D point cloud encoders by transferring visual knowledge extracted from deep 2D image encoders under a standard teacher-student distillation workflow. Generally, we propose PointMCD, a unified multi-view cross-modal distillation architecture, including a pretrained deep image encoder as the teacher and a deep point encoder as the student. To perform heterogeneous feature alignment between 2D visual and 3D geometric domains, we further investigate visibility-aware feature projection (VAFP), by which point-wise embeddings are reasonably aggregated into view-specific geometric descriptors. By pair-wisely aligning multi-view visual and geometric descriptors, we can obtain more powerful deep point encoders without exhausting and complicated network modification. Experiments on 3D shape classification, part segmentation, and unsupervised learning strongly validate the effectiveness of our method. The code and data will be publicly available at https://github.com/keeganhk/PointMCD.
Mamba, a special case of the State Space Model, is gaining popularity as an alternative to template-based deep learning approaches in medical image analysis. While transformers are powerful architectures, they have drawbacks, including quadratic computational complexity and an inability to address long-range dependencies efficiently. This limitation affects the analysis of large and complex datasets in medical imaging, where there are many spatial and temporal relationships. In contrast, Mamba offers benefits that make it well-suited for medical image analysis. It has linear time complexity, which is a significant improvement over transformers. Mamba processes longer sequences without attention mechanisms, enabling faster inference and requiring less memory. Mamba also demonstrates strong performance in merging multimodal data, improving diagnosis accuracy and patient outcomes. The organization of this paper allows readers to appreciate the capabilities of Mamba in medical imaging step by step. We begin by defining core concepts of SSMs and models, including S4, S5, and S6, followed by an exploration of Mamba architectures such as pure Mamba, U-Net variants, and hybrid models with convolutional neural networks, transformers, and Graph Neural Networks. We also cover Mamba optimizations, techniques and adaptations, scanning, datasets, applications, experimental results, and conclude with its challenges and future directions in medical imaging. This review aims to demonstrate the transformative potential of Mamba in overcoming existing barriers within medical imaging while paving the way for innovative advancements in the field. A comprehensive list of Mamba architectures applied in the medical field, reviewed in this work, is available at Github.
Residual neural networks (ResNets) are a promising class of deep neural networks that have shown excellent performance for a number of learning tasks, e.g., image classification and recognition. Mathematically, ResNet architectures can be interpreted as forward Euler discretizations of a nonlinear initial value problem whose time-dependent control variables represent the weights of the neural network. Hence, training a ResNet can be cast as an optimal control problem of the associated dynamical system. For similar time-dependent optimal control problems arising in engineering applications, parallel-in-time methods have shown notable improvements in scalability. This paper demonstrates the use of those techniques for efficient and effective training of ResNets. The proposed algorithms replace the classical (sequential) forward and backward propagation through the network layers by a parallel nonlinear multigrid iteration applied to the layer domain. This adds a new dimension of parallelism across layers that is attractive when training very deep networks. From this basic idea, we derive multiple layer-parallel methods. The most efficient version employs a simultaneous optimization approach where updates to the network parameters are based on inexact gradient information in order to speed up the training process. Using numerical examples from supervised classification, we demonstrate that the new approach achieves similar training performance to traditional methods, but enables layer-parallelism and thus provides speedup over layer-serial methods through greater concurrency.
Segmentation has emerged as a fundamental field of computer vision and natural language processing, which assigns a label to every pixel/feature to extract regions of interest from an image/text. To evaluate the performance of segmentation, the Dice and IoU metrics are used to measure the degree of overlap between the ground truth and the predicted segmentation. In this paper, we establish a theoretical foundation of segmentation with respect to the Dice/IoU metrics, including the Bayes rule and Dice-/IoU-calibration, analogous to classification-calibration or Fisher consistency in classification. We prove that the existing thresholding-based framework with most operating losses are not consistent with respect to the Dice/IoU metrics, and thus may lead to a suboptimal solution. To address this pitfall, we propose a novel consistent ranking-based framework, namely RankDice/RankIoU, inspired by plug-in rules of the Bayes segmentation rule. Three numerical algorithms with GPU parallel execution are developed to implement the proposed framework in large-scale and high-dimensional segmentation. We study statistical properties of the proposed framework. We show it is Dice-/IoU-calibrated, and its excess risk bounds and the rate of convergence are also provided. The numerical effectiveness of RankDice/mRankDice is demonstrated in various simulated examples and Fine-annotated CityScapes, Pascal VOC and Kvasir-SEG datasets with state-of-the-art deep learning architectures.
In this technical report, we briefly introduce our solution for the Zero/Few-shot Track of the Visual Anomaly and Novelty Detection (VAND) 2023 Challenge. For industrial visual inspection, building a single model that can be rapidly adapted to numerous categories without or with only a few normal reference images is a promising research direction. This is primarily because of the vast variety of the product types. For the zero-shot track, we propose a solution based on the CLIP model by adding extra linear layers. These layers are used to map the image features to the joint embedding space, so that they can compare with the text features to generate the anomaly maps. Besides, when the reference images are available, we utilize multiple memory banks to store their features and compare them with the features of the test images during the testing phase. In this challenge, our method achieved first place in the zero-shot track, especially excelling in segmentation with an impressive F1 score improvement of 0.0489 over the second-ranked participant. Furthermore, in the few-shot track, we secured the fourth position overall, with our classification F1 score of 0.8687 ranking first among all participating teams.
Convolutional neural network (CNN) architectures utilize downsampling layers, which restrict the subsequent layers to learn spatially invariant features while reducing computational costs. However, such a downsampling operation makes it impossible to use the full spectrum of input features. Motivated by this observation, we propose a novel layer called parallel grid pooling (PGP) which is applicable to various CNN models. PGP performs downsampling without discarding any intermediate feature. It works as data augmentation and is complementary to commonly used data augmentation techniques. Furthermore, we demonstrate that a dilated convolution can naturally be represented using PGP operations, which suggests that the dilated convolution can also be regarded as a type of data augmentation technique. Experimental results based on popular image classification benchmarks demonstrate the effectiveness of the proposed method. Code is available at: https://github.com/akitotakeki
We present a complete classification of all minimal problems for generic arrangements of points and lines completely observed by calibrated perspective cameras. We show that there are only 30 minimal problems in total, no problems exist for more than 6 cameras, for more than 5 points, and for more than 6 lines. We present a sequence of tests for detecting minimality starting with counting degrees of freedom and ending with full symbolic and numeric verification of representative examples. For all minimal problems discovered, we present their algebraic degrees, i.e. the number of solutions, which measure their intrinsic difficulty. It shows how exactly the difficulty of problems grows with the number of views. Importantly, several new minimal problems have small degrees that might be practical in image matching and 3D reconstruction.
Neural architecture search (NAS) searches architectures automatically for given tasks, e.g., image classification and language modeling. Improving the search efficiency and effectiveness have attracted increasing attention in recent years. However, few efforts have been devoted to understanding the generated architectures. In this paper, we first reveal that existing NAS algorithms (e.g., DARTS, ENAS) tend to favor architectures with wide and shallow cell structures. These favorable architectures consistently achieve fast convergence and are consequently selected by NAS algorithms. Our empirical and theoretical study further confirms that their fast convergence derives from their smooth loss landscape and accurate gradient information. Nonetheless, these architectures may not necessarily lead to better generalization performance compared with other candidate architectures in the same search space, and therefore further improvement is possible by revising existing NAS algorithms.
We developed a rich dataset of Chest X-Ray (CXR) images to assist investigators in artificial intelligence. The data were collected using an eye tracking system while a radiologist reviewed and reported on 1,083 CXR images. The dataset contains the following aligned data: CXR image, transcribed radiology report text, radiologist's dictation audio and eye gaze coordinates data. We hope this dataset can contribute to various areas of research particularly towards explainable and multimodal deep learning / machine learning methods. Furthermore, investigators in disease classification and localization, automated radiology report generation, and human-machine interaction can benefit from these data. We report deep learning experiments that utilize the attention maps produced by eye gaze dataset to show the potential utility of this data.
We present a dynamic model in which the weights are conditioned on an input sample x and are learned to match those that would be obtained by finetuning a base model on x and its label y. This mapping between an input sample and network weights is approximated by a denoising diffusion model. The diffusion model we employ focuses on modifying a single layer of the base model and is conditioned on the input, activations, and output of this layer. Since the diffusion model is stochastic in nature, multiple initializations generate different networks, forming an ensemble, which leads to further improvements. Our experiments demonstrate the wide applicability of the method for image classification, 3D reconstruction, tabular data, speech separation, and natural language processing. Our code is available at https://github.com/ShaharLutatiPersonal/OCD
The increasing popularity of facial manipulation (Deepfakes) and synthetic face creation raises the need to develop robust forgery detection solutions. Crucially, most work in this domain assume that the Deepfakes in the test set come from the same Deepfake algorithms that were used for training the network. This is not how things work in practice. Instead, we consider the case where the network is trained on one Deepfake algorithm, and tested on Deepfakes generated by another algorithm. Typically, supervised techniques follow a pipeline of visual feature extraction from a deep backbone, followed by a binary classification head. Instead, our algorithm aggregates features extracted across all layers of one backbone network to detect a fake. We evaluate our approach on two domains of interest - Deepfake detection and Synthetic image detection, and find that we achieve SOTA results.
Subsampling layers play a crucial role in deep nets by discarding a portion of an activation map to reduce its spatial dimensions. This encourages the deep net to learn higher-level representations. Contrary to this motivation, we hypothesize that the discarded activations are useful and can be incorporated on the fly to improve models' prediction. To validate our hypothesis, we propose a search and aggregate method to find useful activation maps to be used at test time. We applied our approach to the task of image classification and semantic segmentation. Extensive experiments over nine different architectures on multiple datasets show that our method consistently improves model test-time performance, complementing existing test-time augmentation techniques. Our code is available at https://github.com/ca-joe-yang/discard-in-subsampling.
We introduce \textbf{Knowledge Swapping}, a novel task designed to selectively regulate knowledge of a pretrained model by enabling the forgetting of user\-specified information, retaining essential knowledge, and acquiring new knowledge simultaneously. By delving into the analysis of knock-on feature hierarchy, we find that incremental learning typically progresses from low\-level representations to higher\-level semantics, whereas forgetting tends to occur in the opposite direction\-starting from high-level semantics and moving down to low-level features. Building upon this, we propose to benchmark the knowledge swapping task with the strategy of \textit{Learning Before Forgetting}. Comprehensive experiments on various tasks like image classification, object detection, and semantic segmentation validate the effectiveness of the proposed strategy. The source code is available at \href{https://github.com/xingmingyu123456/KnowledgeSwapping}{https://github.com/xingmingyu123456/KnowledgeSwapping}.
Vision Transformers (ViTs) have become prominent models for solving various vision tasks. However, the interpretability of ViTs has not kept pace with their promising performance. While there has been a surge of interest in developing {\it post hoc} solutions to explain ViTs' outputs, these methods do not generalize to different downstream tasks and various transformer architectures. Furthermore, if ViTs are not properly trained with the given data and do not prioritize the region of interest, the {\it post hoc} methods would be less effective. Instead of developing another {\it post hoc} approach, we introduce a novel training procedure that inherently enhances model interpretability. Our interpretability-aware ViT (IA-ViT) draws inspiration from a fresh insight: both the class patch and image patches consistently generate predicted distributions and attention maps. IA-ViT is composed of a feature extractor, a predictor, and an interpreter, which are trained jointly with an interpretability-aware training objective. Consequently, the interpreter simulates the behavior of the predictor and provides a faithful explanation through its single-head self-attention mechanism. Our comprehensive experimental results demonstrate the effectiveness of IA-ViT in several image classification tasks, with both qualitative and quantitative evaluations of model performance and interpretability. Source code is available from: https://github.com/qiangyao1988/IA-ViT.
Self-supervised learning (SSL) is one strategy for addressing the paucity of labelled data in medical imaging by learning representations from unlabelled images. Contrastive and non-contrastive SSL methods produce learned representations that are similar for pairs of related images. Such pairs are commonly constructed by randomly distorting the same image twice. The videographic nature of ultrasound offers flexibility for defining the similarity relationship between pairs of images. In this study, we investigated the effect of utilizing proximal, distinct images from the same B-mode ultrasound video as pairs for SSL. Additionally, we introduced a sample weighting scheme that increases the weight of closer image pairs and demonstrated how it can be integrated into SSL objectives. Named Intra-Video Positive Pairs (IVPP), the method surpassed previous ultrasound-specific contrastive learning methods' average test accuracy on COVID-19 classification with the POCUS dataset by $\ge 1.3\%$. Detailed investigations of IVPP's hyperparameters revealed that some combinations of IVPP hyperparameters can lead to improved or worsened performance, depending on the downstream task. Guidelines for practitioners were synthesized based on the results, such as the merit of IVPP with task-specific hyperparameters, and the improved performance of contrastive methods for ultrasound compared to non-contrastive counterparts.
Harvesting is a critical task in the tree fruit industry, demanding extensive manual labor and substantial costs, and exposing workers to potential hazards. Recent advances in automated harvesting offer a promising solution by enabling efficient, cost-effective, and ergonomic fruit picking within tight harvesting windows. However, existing harvesting technologies often indiscriminately harvest all visible and accessible fruits, including those that are unripe or undersized. This study introduces a novel foundation model-based framework for efficient apple ripeness and size estimation. Specifically, we curated two public RGBD-based Fuji apple image datasets, integrating expanded annotations for ripeness ("Ripe" vs. "Unripe") based on fruit color and image capture dates. The resulting comprehensive dataset, Fuji-Ripeness-Size Dataset, includes 4,027 images and 16,257 annotated apples with ripeness and size labels. Using Grounding-DINO, a language-model-based object detector, we achieved robust apple detection and ripeness classification, outperforming other state-of-the-art models. Additionally, we developed and evaluated six size estimation algorithms, selecting the one with the lowest error and variation for optimal performance. The Fuji-Ripeness-Size Dataset and the apple detection and size estimation algorithms are made publicly available, which provides valuable benchmarks for future studies in automated and selective harvesting.
Recent advances have increasingly applied large language models (LLMs) to electrocardiogram (ECG) interpretation, giving rise to Electrocardiogram-Language Models (ELMs). Conditioned on an ECG and a textual query, an ELM autoregressively generates a free-form textual response. Unlike traditional classification-based systems, ELMs emulate expert cardiac electrophysiologists by issuing diagnoses, analyzing waveform morphology, identifying contributing factors, and proposing patient-specific action plans. To realize this potential, researchers are curating instruction-tuning datasets that pair ECGs with textual dialogues and are training ELMs on these resources. Yet before scaling ELMs further, there is a fundamental question yet to be explored: What is the most effective ECG input representation? In recent works, three candidate representations have emerged-raw time-series signals, rendered images, and discretized symbolic sequences. We present the first comprehensive benchmark of these modalities across 6 public datasets and 5 evaluation metrics. We find symbolic representations achieve the greatest number of statistically significant wins over both signal and image inputs. We further ablate the LLM backbone, ECG duration, and token budget, and we evaluate robustness to signal perturbations. We hope that our findings offer clear guidance for selecting input representations when developing the next generation of ELMs.
LDCT has drawn major attention in the medical imaging field due to the potential health risks of CT-associated X-ray radiation to patients. Reducing the radiation dose, however, decreases the quality of the reconstructed images, which consequently compromises the diagnostic performance. Various deep learning techniques have been introduced to improve the image quality of LDCT images through denoising. GANs-based denoising methods usually leverage an additional classification network, i.e. discriminator, to learn the most discriminate difference between the denoised and normal-dose images and, hence, regularize the denoising model accordingly; it often focuses either on the global structure or local details. To better regularize the LDCT denoising model, this paper proposes a novel method, termed DU-GAN, which leverages U-Net based discriminators in the GANs framework to learn both global and local difference between the denoised and normal-dose images in both image and gradient domains. The merit of such a U-Net based discriminator is that it can not only provide the per-pixel feedback to the denoising network through the outputs of the U-Net but also focus on the global structure in a semantic level through the middle layer of the U-Net. In addition to the adversarial training in the image domain, we also apply another U-Net based discriminator in the image gradient domain to alleviate the artifacts caused by photon starvation and enhance the edge of the denoised CT images. Furthermore, the CutMix technique enables the per-pixel outputs of the U-Net based discriminator to provide radiologists with a confidence map to visualize the uncertainty of the denoised results, facilitating the LDCT-based screening and diagnosis. Extensive experiments on the simulated and real-world datasets demonstrate superior performance over recently published methods both qualitatively and quantitatively.