Image collections, if critical aspects of image content are exposed, can spur research and practical applications in many domains. Supervised machine learning may be the only feasible way to annotate very large collections, but leading approaches rely on large samples of completely and accurately annotated images. In the case of a large forensic collection, we are aiming to annotate, neither the complete annotation nor the large training samples can be feasibly produced. We, therefore, investigate ways to assist manual annotation efforts done by forensic experts. We present a method that can propose both images and areas within an image likely to contain desired classes. Evaluation of the method with human annotators showed highly accurate classification that was strongly helped by transfer learning. The segmentation precision (mAP) was improved by adding a separate class capturing background, but that did not affect the recall (mAR). Further work is needed to both increase the accuracy of segmentation and enhances prediction with additional covariates affecting decomposition. We hope this effort to be of help in other domains that require weak segmentation and have limited availability of qualified annotators.
Within the last years, the classification of variable stars with Machine Learning has become a mainstream area of research. Recently, visualization of time series is attracting more attention in data science as a tool to visually help scientists to recognize significant patterns in complex dynamics. Within the Machine Learning literature, dictionary-based methods have been widely used to encode relevant parts of image data. These methods intrinsically assign a degree of importance to patches in pictures, according to their contribution in the image reconstruction. Inspired by dictionary-based techniques, we present an approach that naturally provides the visualization of salient parts in astronomical light curves, making the analogy between image patches and relevant pieces in time series. Our approach encodes the most meaningful patterns such that we can approximately reconstruct light curves by just using the encoded information. We test our method in light curves from the OGLE-III and StarLight databases. Our results show that the proposed model delivers an automatic and intuitive visualization of relevant light curve parts, such as local peaks and drops in magnitude.
Contrastive Language-Image Pre-training (CLIP) has made a remarkable breakthrough in open-vocabulary zero-shot image recognition. Many recent studies leverage the pre-trained CLIP models for image-level classification and manipulation. In this paper, we wish examine the intrinsic potential of CLIP for pixel-level dense prediction, specifically in semantic segmentation. To this end, with minimal modification, we show that MaskCLIP yields compelling segmentation results on open concepts across various datasets in the absence of annotations and fine-tuning. By adding pseudo labeling and self-training, MaskCLIP+ surpasses SOTA transductive zero-shot semantic segmentation methods by large margins, e.g., mIoUs of unseen classes on PASCAL VOC/PASCAL Context/COCO Stuff are improved from 35.6/20.7/30.3 to 86.1/66.7/54.7. We also test the robustness of MaskCLIP under input corruption and evaluate its capability in discriminating fine-grained objects and novel concepts. Our finding suggests that MaskCLIP can serve as a new reliable source of supervision for dense prediction tasks to achieve annotation-free segmentation. Source code is available at https://github.com/chongzhou96/MaskCLIP.
Imitation learning often needs a large demonstration set in order to handle the full range of situations that an agent might find itself in during deployment. However, collecting expert demonstrations can be expensive. Recent work in vision, reinforcement learning, and NLP has shown that auxiliary representation learning objectives can reduce the need for large amounts of expensive, task-specific data. Our Empirical Investigation of Representation Learning for Imitation (EIRLI) investigates whether similar benefits apply to imitation learning. We propose a modular framework for constructing representation learning algorithms, then use our framework to evaluate the utility of representation learning for imitation across several environment suites. In the settings we evaluate, we find that existing algorithms for image-based representation learning provide limited value relative to a well-tuned baseline with image augmentations. To explain this result, we investigate differences between imitation learning and other settings where representation learning has provided significant benefit, such as image classification. Finally, we release a well-documented codebase which both replicates our findings and provides a modular framework for creating new representation learning algorithms out of reusable components.
Existing vision-text contrastive learning like CLIP aims to match the paired image and caption embeddings while pushing others apart, which improves representation transferability and supports zero-shot prediction. However, medical image-text datasets are orders of magnitude below the general images and captions from the internet. Moreover, previous methods encounter many false negatives, i.e., images and reports from separate patients probably carry the same semantics but are wrongly treated as negatives. In this paper, we decouple images and texts for multimodal contrastive learning thus scaling the usable training data in a combinatorial magnitude with low cost. We also propose to replace the InfoNCE loss with semantic matching loss based on medical knowledge to eliminate false negatives in contrastive learning. We prove that MedCLIP is a simple yet effective framework: it outperforms state-of-the-art methods on zero-shot prediction, supervised classification, and image-text retrieval. Surprisingly, we observe that with only 20K pre-training data, MedCLIP wins over the state-of-the-art method (using around 200K data). Our code is available at https://github.com/RyanWangZf/MedCLIP.
Statistical analysis and modeling is becoming increasingly popular for the world's leading organizations, especially for professional NBA teams. Sophisticated methods and models of sport talent evaluation have been created for this purpose. In this research, we present a different perspective from the dominant tactic of statistical data analysis. Based on a strategy that NBA teams have followed in the past, hiring human professionals, we deploy image analysis and Convolutional Neural Networks in an attempt to predict the career trajectory of newly drafted players from each draft class. We created a database consisting of about 1500 image data from players from every draft since 1990. We then divided the players into five different quality classes based on their expected NBA career. Next, we trained popular pre-trained image classification models in our data and conducted a series of tests in an attempt to create models that give reliable predictions of the rookie players' careers. The results of this study suggest that there is a potential correlation between facial characteristics and athletic talent, worth of further investigation.
Feature-Imitating-Networks (FINs) are neural networks that are first trained to approximate closed-form statistical features (e.g. Entropy), and then embedded into other networks to enhance their performance. In this work, we perform the first evaluation of FINs for biomedical image processing tasks. We begin by training a set of FINs to imitate six common radiomics features, and then compare the performance of larger networks (with and without embedding the FINs) for three experimental tasks: COVID-19 detection from CT scans, brain tumor classification from MRI scans, and brain-tumor segmentation from MRI scans. We found that models embedded with FINs provided enhanced performance for all three tasks when compared to baseline networks without FINs, even when those baseline networks had more parameters. Additionally, we found that models embedded with FINs converged faster and more consistently compared to baseline networks with similar or greater representational capacity. The results of our experiments provide evidence that FINs may offer state-of-the-art performance for a variety of other biomedical image processing tasks.
Backdoor attacks have severely threatened deep neural network (DNN) models in the past several years. These attacks can occur in almost every stage of the deep learning pipeline. Although the attacked model behaves normally on benign samples, it makes wrong predictions for samples containing triggers. However, most existing attacks use visible patterns (e.g., a patch or image transformations) as triggers, which are vulnerable to human inspection. In this paper, we propose a novel backdoor attack, making imperceptible changes. Concretely, our attack first utilizes the pre-trained victim model to extract low-level and high-level semantic features from clean images and generates trigger pattern associated with high-level features based on channel attention. Then, the encoder model generates poisoned images based on the trigger and extracted low-level semantic features without causing noticeable feature loss. We evaluate our attack on three prominent image classification DNN across three standard datasets. The results demonstrate that our attack achieves high attack success rates while maintaining robustness against backdoor defenses. Furthermore, we conduct extensive image similarity experiments to emphasize the stealthiness of our attack strategy.
Most multi-modal tasks can be formulated into problems of either generation or embedding. Existing models usually tackle these two types of problems by decoupling language modules into a text decoder for generation, and a text encoder for embedding. To explore the minimalism of multi-modal paradigms, we attempt to achieve only one model per modality in this work. We propose a Multi-Modal Generative Embedding Model (MM-GEM), whereby the generative and embedding objectives are encapsulated in one Large Language Model. We also propose a PoolAggregator to boost efficiency and enable the ability of fine-grained embedding and generation. A surprising finding is that these two objectives do not significantly conflict with each other. For example, MM-GEM instantiated from ViT-Large and TinyLlama shows competitive performance on benchmarks for multimodal embedding models such as cross-modal retrieval and zero-shot classification, while has good ability of image captioning. Additionally, MM-GEM can seamlessly execute region-level image caption generation and retrieval tasks. Besides, the advanced text model in MM-GEM brings over 5% improvement in Recall@1 for long text and image retrieval.
With the rapid advancement of generative AI, AI-generated images have become increasingly realistic, raising concerns about creativity, misinformation, and content authenticity. Detecting such images and identifying their source models has become a critical challenge in ensuring the integrity of digital media. This paper tackles the detection of AI-generated images and identifying their source models using CNN and CLIP-ViT classifiers. For the CNN-based classifier, we leverage EfficientNet-B0 as the backbone and feed with RGB channels, frequency features, and reconstruction errors, while for CLIP-ViT, we adopt a pretrained CLIP image encoder to extract image features and SVM to perform classification. Evaluated on the Defactify 4 dataset, our methods demonstrate strong performance in both tasks, with CLIP-ViT showing superior robustness to image perturbations. Compared to baselines like AEROBLADE and OCC-CLIP, our approach achieves competitive results. Notably, our method ranked Top-3 overall in the Defactify 4 competition, highlighting its effectiveness and generalizability. All of our implementations can be found in https://github.com/uuugaga/Defactify_4
현대 비전-언어 모델(VLM)은 재난 대응, OSINT 검증, 위치 프라이버시에 매우 중요한 과제인 이미지 지리적 위치 추정에서 인간 기준선을 크게 상회하는 성능을 보인다. 그러나 이 과제에서 AI 행동을 연구하려는 대부분의 노력은 정적 이미지 기반 검색, 분류, 예측에 국한되어 있다. 우리는 과제를 충실히 재현하려면 다중모달 에이전트가 예측을 제출하기 전에 주변을 자율적으로 탐색하여 관측을 수집하는 체화 내비게이션이 포함되어야 한다고 주장한다. 이를 위해 에이전트가 Street View 환경을 탐색하며 순차적 추론을 통해 지리적 위치 추정을 정교화해야 하는 에이전트형 환경 기반 벤치마크 GeoAgent를 제안한다. 분석 결과, 현대 VLM은 국가 및 대륙 수준 예측에는 성공하지만 지역적 패턴을 구별하는 데 어려움을 겪는다. 정적 이미지 기반 기준선과 비교하면 에이전트형 내비게이션은 확립된 지표 전반에서 정확도를 크게 향상시킨다. 또한 최신 모델 아키텍처 전반에서 선진국/개발도상국 지역 맥락의 심각한 바이어스와 잘못된 사전 정보 하에서의 낮은 자기 개선 능력을 확인한다. 전반적으로 본 연구는 체화 내비게이션과 지리공간 추론의 난제를 정립한다. 코드와 GeoAgent 환경을 공개한다: https://geoagent-benchmark.github.io
This paper presents an approach for answering fill-in-the-blank multiple choice questions from the Visual Madlibs dataset. Instead of generic and commonly used representations trained on the ImageNet classification task, our approach employs a combination of networks trained for specialized tasks such as scene recognition, person activity classification, and attribute prediction. We also present a method for localizing phrases from candidate answers in order to provide spatial support for feature extraction. We map each of these features, together with candidate answers, to a joint embedding space through normalized canonical correlation analysis (nCCA). Finally, we solve an optimization problem to learn to combine scores from nCCA models trained on multiple cues to select the best answer. Extensive experimental results show a significant improvement over the previous state of the art and confirm that answering questions from a wide range of types benefits from examining a variety of image cues and carefully choosing the spatial support for feature extraction.
In the presence of large sets of labeled data, Deep Learning (DL) has accomplished extraordinary triumphs in the avenue of computer vision, particularly in object classification and recognition tasks. However, DL cannot always perform well when the training and testing images come from different distributions or in the presence of domain shift between training and testing images. They also suffer in the absence of labeled input data. Domain adaptation (DA) methods have been proposed to make up the poor performance due to domain shift. In this paper, we present a new unsupervised deep domain adaptation method based on the alignment of second order statistics (covariances) as well as maximum mean discrepancy of the source and target data with a two stream Convolutional Neural Network (CNN). We demonstrate the ability of the proposed approach to achieve state-of the-art performance for image classification on three benchmark domain adaptation datasets: Office-31 [27], Office-Home [37] and Office-Caltech [8].
Discriminatively trained neural classifiers can be trusted, only when the input data comes from the training distribution (in-distribution). Therefore, detecting out-of-distribution (OOD) samples is very important to avoid classification errors. In the context of OOD detection for image classification, one of the recent approaches proposes training a classifier called "confident-classifier" by minimizing the standard cross-entropy loss on in-distribution samples and minimizing the KL divergence between the predictive distribution of OOD samples in the low-density regions of in-distribution and the uniform distribution (maximizing the entropy of the outputs). Thus, the samples could be detected as OOD if they have low confidence or high entropy. In this paper, we analyze this setting both theoretically and experimentally. We conclude that the resulting confident-classifier still yields arbitrarily high confidence for OOD samples far away from the in-distribution. We instead suggest training a classifier by adding an explicit "reject" class for OOD samples.
Stochastic variational inference for Bayesian deep neural network (DNN) requires specifying priors and approximate posterior distributions over neural network weights. Specifying meaningful weight priors is a challenging problem, particularly for scaling variational inference to deeper architectures involving high dimensional weight space. We propose MOdel Priors with Empirical Bayes using DNN (MOPED) method to choose informed weight priors in Bayesian neural networks. We formulate a two-stage hierarchical modeling, first find the maximum likelihood estimates of weights with DNN, and then set the weight priors using empirical Bayes approach to infer the posterior with variational inference. We empirically evaluate the proposed approach on real-world tasks including image classification, video activity recognition and audio classification with varying complex neural network architectures. We also evaluate our proposed approach on diabetic retinopathy diagnosis task and benchmark with the state-of-the-art Bayesian deep learning techniques. We demonstrate MOPED method enables scalable variational inference and provides reliable uncertainty quantification.
The renaissance of neural architecture search (NAS) has seen classical methods such as genetic algorithms (GA) and genetic programming (GP) being exploited for convolutional neural network (CNN) architectures. While recent work have achieved promising performance on visual perception tasks, the direct encoding scheme of both GA and GP has functional complexity deficiency and does not scale well on large architectures like CNN. To address this, we present a new generative encoding scheme -- $symbolic\ linear\ generative\ encoding$ (SLGE) -- simple, yet powerful scheme which embeds local graph transformations in chromosomes of linear fixed-length string to develop CNN architectures of variant shapes and sizes via evolutionary process of gene expression programming. In experiments, the effectiveness of SLGE is shown in discovering architectures that improve the performance of the state-of-the-art handcrafted CNN architectures on CIFAR-10 and CIFAR-100 image classification tasks; and achieves a competitive classification error rate with the existing NAS methods using less GPU resources.
In many classification tasks, the set of target classes can be organized into a hierarchy. This structure induces a semantic distance between classes, and can be summarised under the form of a cost matrix, which defines a finite metric on the class set. In this paper, we propose to model the hierarchical class structure by integrating this metric in the supervision of a prototypical network. Our method relies on jointly learning a feature-extracting network and a set of class prototypes whose relative arrangement in the embedding space follows an hierarchical metric. We show that this approach allows for a consistent improvement of the error rate weighted by the cost matrix when compared to traditional methods and other prototype-based strategies. Furthermore, when the induced metric contains insight on the data structure, our method improves the overall precision as well. Experiments on four different public datasets - from agricultural time series classification to depth image semantic segmentation - validate our approach.
The existing deep learning models suffer from out-of-distribution (o.o.d.) performance drop in computer vision tasks. In comparison, humans have a remarkable ability to interpret images, even if the scenes in the images are rare, thanks to the systematicity of acquired knowledge. This work focuses on 1) the acquisition of systematic knowledge of 2D transformations, and 2) architectural components that can leverage the learned knowledge in image classification tasks in an o.o.d. setting. With a new training methodology based on synthetic datasets that are constructed under the causal framework, the deep neural networks acquire knowledge from semantically different domains (e.g. even from noise), and exhibit certain level of systematicity in parameter estimation experiments. Based on this, a novel architecture is devised consisting of a classifier, an estimator and an identifier (abbreviated as "CED"). By emulating the "hypothesis-verification" process in human visual perception, CED improves the classification accuracy significantly on test sets under covariate shift.
We introduce a new benchmark dataset, Placenta, for node classification in an underexplored domain: predicting microanatomical tissue structures from cell graphs in placenta histology whole slide images. This problem is uniquely challenging for graph learning for a few reasons. Cell graphs are large (>1 million nodes per image), node features are varied (64-dimensions of 11 types of cells), class labels are imbalanced (9 classes ranging from 0.21% of the data to 40.0%), and cellular communities cluster into heterogeneously distributed tissues of widely varying sizes (from 11 nodes to 44,671 nodes for a single structure). Here, we release a dataset consisting of two cell graphs from two placenta histology images totalling 2,395,747 nodes, 799,745 of which have ground truth labels. We present inductive benchmark results for 7 scalable models and show how the unique qualities of cell graphs can help drive the development of novel graph neural network architectures.
Deep learning has proven to be successful in various domains and for different tasks. However, when it comes to private data several restrictions are making it difficult to use deep learning approaches in these application fields. Recent approaches try to generate data privately instead of applying a privacy-preserving mechanism directly, on top of the classifier. The solution is to create public data from private data in a manner that preserves the privacy of the data. In this work, two very prominent GAN-based architectures were evaluated in the context of private time series classification. In contrast to previous work, mostly limited to the image domain, the scope of this benchmark was the time series domain. The experiments show that especially GSWGAN performs well across a variety of public datasets outperforming the competitor DPWGAN. An analysis of the generated datasets further validates the superiority of GSWGAN in the context of time series generation.