Training machine learning models with differential privacy (DP) has received increasing interest in recent years. One of the most popular algorithms for training differentially private models is differentially private stochastic gradient descent (DPSGD) and its variants, where at each step gradients are clipped and combined with some noise. Given the increasing usage of DPSGD, we ask the question: is DPSGD alone sufficient to find a good minimizer for every dataset under privacy constraints? Towards answering this question, we show that even for the simple case of linear classification, unlike non-private optimization, (private) feature preprocessing is vital for differentially private optimization. In detail, we first show theoretically that there exists an example where without feature preprocessing, DPSGD incurs an optimality gap proportional to the maximum Euclidean norm of features over all samples. We then propose an algorithm called DPSGD-F, which combines DPSGD with feature preprocessing and prove that for classification tasks, it incurs an optimality gap proportional to the diameter of the features $\max_{x, x' \in D} \|x - x'\|_2$. We finally demonstrate the practicality of our algorithm on image classification benchmarks.
Transferability estimation aims to provide heuristics for quantifying how suitable a pre-trained model is for a specific downstream task, without fine-tuning them all. Prior studies have revealed that well-trained models exhibit the phenomenon of Neural Collapse. Based on a widely used neural collapse metric in existing literature, we observe a strong correlation between the neural collapse of pre-trained models and their corresponding fine-tuned models. Inspired by this observation, we propose a novel method termed Fair Collapse (FaCe) for transferability estimation by comprehensively measuring the degree of neural collapse in the pre-trained model. Typically, FaCe comprises two different terms: the variance collapse term, which assesses the class separation and within-class compactness, and the class fairness term, which quantifies the fairness of the pre-trained model towards each class. We investigate FaCe on a variety of pre-trained classification models across different network architectures, source datasets, and training loss functions. Results show that FaCe yields state-of-the-art performance on different tasks including image classification, semantic segmentation, and text classification, which demonstrate the effectiveness and generalization of our method.
We present a novel LLM-based pipeline for creating contextual descriptions of human body poses in images using only auxiliary attributes. This approach facilitates the creation of the MPII Pose Descriptions dataset, which includes natural language annotations for 17,367 images containing people engaged in 410 distinct activities. We demonstrate the effectiveness of our pose descriptions in enabling zero-shot human-centric classification using CLIP. Moreover, we introduce the FocusCLIP framework, which incorporates Subject-Focused Attention (SFA) in CLIP for improved text-to-image alignment. Our models were pretrained on the MPII Pose Descriptions dataset and their zero-shot performance was evaluated on five unseen datasets covering three tasks. FocusCLIP outperformed the baseline CLIP model, achieving an average accuracy increase of 8.61\% (33.65\% compared to CLIP's 25.04\%). Notably, our approach yielded improvements of 3.98\% in activity recognition, 14.78\% in age classification, and 7.06\% in emotion recognition. These results highlight the potential of integrating detailed pose descriptions and subject-level guidance into general pretraining frameworks for enhanced performance in downstream tasks.
신경망 모델 편집을 강화학습(RL) 문제로 공식화하여 에이전트가 보상 피드백을 통해 모델을 수정하는 탐색적 프레임워크를 제안한다. 텍스트 분류의 편향 완화 및 이미지 분류의 머신 언러닝 평가를 통해, 학습된 정책이 모델의 전반적인 성능을 유지하면서도 특정 수정 목표를 효과적으로 달성함을 입증했다.
본 논문에서는 정확한 군중 밀도 추정을 위한 최초의 완전한 CLIP 기반 모델인 CLIP-EBC를 제안합니다. CLIP 모델은 제로샷(zero-shot) 이미지 분류와 같은 인식 태스크에서 뛰어난 성과를 보여주었으나, 계수(counting)와 같은 회귀 문제를 인식 태스크로 변환하는 데 따르는 본질적인 어려움으로 인해 계수 태스크에서의 잠재력은 거의 탐색되지 않았습니다. 본 연구에서는 이미지로부터 군중 규모를 추정하는 태스크에 특히 초점을 맞추어 CLIP의 계수 능력을 탐구하고 개선합니다. 기존의 분류 기반 군중 계수 프레임워크는 계수 값을 인접한 실수형 빈(bin)으로 양자화하고 분류 오차에만 집중하는 등 심각한 한계를 가집니다. 이러한 방식은 인접 경계 근처에서의 레이블 모호성과 부정확한 계수 값 예측을 초래합니다. 따라서 기존 프레임워크에 CLIP을 직접 적용하면 차선의 성능을 보일 수 있습니다. 이러한 문제를 해결하기 위해, 먼저 개선된 블록별 분류(EBC, Enhanced Blockwise Classification) 프레임워크를 제안합니다. 이전 방법들과 달리 EBC는 정수형 빈을 사용하여 빈 경계 근처의 모호성을 효과적으로 줄입니다. 또한 밀도 맵(density map) 기반의 회귀 손실을 도입하여 계수 값 예측 정확도를 향상합니다. 백본에 구애받지 않는(backbone-agnostic) EBC 프레임워크 내에서, 해당 태스크에 CLIP의 인식 기능을 완전히 활용하는 CLIP-EBC를 소개합니다. 광범위한 실험을 통해 EBC의 효과와 CLIP-EBC의 경쟁력 있는 성능을 입증합니다. 구체적으로, 제안하는 EBC 프레임워크는 UCF-QNRF 데이터셋에서 기존 분류 기반 방법들의 성능을 최대 44.5% 향상하며, CLIP-EBC는 NWPU-Crowd 테스트 세트에서 MAE 58.2, RMSE 268.5를 기록하여 이전 최고 성능 방법인 STEERER 대비 각각 8.6% 및 13.3% 향상된 SOTA(state-of-the-art) 성능을 달성합니다. 코드와 가중치는 https://github.com/Yiming-M/CLIP-EBC 에서 확인할 수 있습니다.
Image forgery detection aims to detect and locate forged regions in an image. Most existing forgery detection algorithms formulate classification problems to classify pixels into forged or pristine. However, the definition of forged and pristine pixels is only relative within one single image, e.g., a forged region in image A is actually a pristine one in its source image B (splicing forgery). Such a relative definition has been severely overlooked by existing methods, which unnecessarily mix forged (pristine) regions across different images into the same category. To resolve this dilemma, we propose the FOrensic ContrAstive cLustering (FOCAL) method, a novel, simple yet very effective paradigm based on soft contrastive learning and unsupervised clustering for the image forgery detection. Specifically, FOCAL 1) designs a soft contrastive learning (SCL) to supervise the high-level forensic feature extraction in an image-by-image manner, explicitly reflecting the above relative definition; 2) employs an on-the-fly unsupervised clustering algorithm (instead of a trained one) to cluster the learned features into forged/pristine categories, further suppressing the cross-image influence from training data; and 3) allows to further boost the detection performance via simple feature-level concatenation without the need of retraining. Extensive experimental results over six public testing datasets demonstrate that our proposed FOCAL significantly outperforms the state-of-the-art competitors by big margins: +24.8% on Coverage, +18.9% on Columbia, +17.3% on FF++, +15.3% on MISD, +15.0% on CASIA and +10.5% on NIST in terms of IoU (see also Fig. 1). The paradigm of FOCAL could bring fresh insights and serve as a novel benchmark for the image forgery detection task. The code is available at https://github.com/HighwayWu/FOCAL.
As a prominent parameter-efficient fine-tuning technique in NLP, prompt tuning is being explored its potential in computer vision. Typical methods for visual prompt tuning follow the sequential modeling paradigm stemming from NLP, which represents an input image as a flattened sequence of token embeddings and then learns a set of unordered parameterized tokens prefixed to the sequence representation as the visual prompts for task adaptation of large vision models. While such sequential modeling paradigm of visual prompt has shown great promise, there are two potential limitations. First, the learned visual prompts cannot model the underlying spatial relations in the input image, which is crucial for image encoding. Second, since all prompt tokens play the same role of prompting for all image tokens without distinction, it lacks the fine-grained prompting capability, i.e., individual prompting for different image tokens. In this work, we propose the \mymodel model (\emph{SA$^2$VP}), which learns a two-dimensional prompt token map with equal (or scaled) size to the image token map, thereby being able to spatially align with the image map. Each prompt token is designated to prompt knowledge only for the spatially corresponding image tokens. As a result, our model can conduct individual prompting for different image tokens in a fine-grained manner. Moreover, benefiting from the capability of preserving the spatial structure by the learned prompt token map, our \emph{SA$^2$VP} is able to model the spatial relations in the input image, leading to more effective prompting. Extensive experiments on three challenging benchmarks for image classification demonstrate the superiority of our model over other state-of-the-art methods for visual prompt tuning. Code is available at \emph{https://github.com/tommy-xq/SA2VP}.
Modern deep neural networks can achieve high accuracy when the training distribution and test distribution are identically distributed, but this assumption is frequently violated in practice. When the train and test distributions are mismatched, accuracy can plummet. Currently there are few techniques that improve robustness to unforeseen data shifts encountered during deployment. In this work, we propose a technique to improve the robustness and uncertainty estimates of image classifiers. We propose AugMix, a data processing technique that is simple to implement, adds limited computational overhead, and helps models withstand unforeseen corruptions. AugMix significantly improves robustness and uncertainty measures on challenging image classification benchmarks, closing the gap between previous methods and the best possible performance in some cases by more than half.
Neural Radiance Fields (NeRFs) have emerged as a standard framework for representing 3D scenes and objects, introducing a novel data type for information exchange and storage. Concurrently, significant progress has been made in multimodal representation learning for text and image data. This paper explores a novel research direction that aims to connect the NeRF modality with other modalities, similar to established methodologies for images and text. To this end, we propose a simple framework that exploits pre-trained models for NeRF representations alongside multimodal models for text and image processing. Our framework learns a bidirectional mapping between NeRF embeddings and those obtained from corresponding images and text. This mapping unlocks several novel and useful applications, including NeRF zero-shot classification and NeRF retrieval from images or text.
With the ongoing rise of machine learning, the need for methods for explaining decisions made by artificial intelligence systems is becoming a more and more important topic. Especially for image classification tasks, many state-of-the-art tools to explain such classifiers rely on visual highlighting of important areas of the input data. Contrary, counterfactual explanation systems try to enable a counterfactual reasoning by modifying the input image in a way such that the classifier would have made a different prediction. By doing so, the users of counterfactual explanation systems are equipped with a completely different kind of explanatory information. However, methods for generating realistic counterfactual explanations for image classifiers are still rare. Especially in medical contexts, where relevant information often consists of textural and structural information, high-quality counterfactual images have the potential to give meaningful insights into decision processes. In this work, we present GANterfactual, an approach to generate such counterfactual image explanations based on adversarial image-to-image translation techniques. Additionally, we conduct a user study to evaluate our approach in an exemplary medical use case. Our results show that, in the chosen medical use-case, counterfactual explanations lead to significantly better results regarding mental models, explanation satisfaction, trust, emotions, and self-efficacy than two state-of-the-art systems that work with saliency maps, namely LIME and LRP.
결막 사진의 자동 등급 분류는 트라코마 유병률 조사의 비용과 변동성을 줄일 수 있지만, 공통 프로토콜 하에서 최신 사전학습 시각 표현, 경량 특징 적응, 학습 목표 설계의 상대적 가치는 아직 확립되지 않았다. 본 연구는 공개 UCSF/Lietman 컬렉션의 1,546장 이미지를 사용하여 Trachomatous Inflammation-Follicular(TF) 대 Normal의 이진 분류에 대한 통제된 평가를 제시한다. 이미지는 제로샷 tarsal-conjunctiva 분할, 정렬, 크롭, 표준화를 위해 OPTED 파이프라인으로 처리한다. 먼저 공통 분류 파이프라인으로 여섯 가지 사전학습 백본을 비교한 뒤, DINOv2 ViT-B/14에 네 가지 경량 적응 메커니즘을 평가한다. 계층화 5-fold 교차 검증 하에서, Efficient Channel Attention(ECA)과 focal-plus-center 손실을 적용한 DINOv2는 91.66 +/- 0.97% 정확도, 90.69 +/- 1.10% macro-F1, 96.06 +/- 0.71% AUC를 달성했다. ECA는 학습 가능 매개변수 다섯 개만 추가하면서 훨씬 큰 대안과 동등한 성능을 냈다. 목적 함수 ablation은 ECA가 손실 함수 전반에서 plain DINOv2를 일관되게 개선하지 않았음을 추가로 보여주었으며, 분산이 가장 낮은 91.66% 정확도는 교차 엔트로피와 center loss로 얻었다. 전반적으로, 파인튜닝된 DINOv2 표현이 예측 성능의 대부분을 제공했고, ECA는 학습 목표에 따라 효과가 달라지는 매우 매개변수 효율적인 정제를 제공했다. 결과 워크플로는 활성 트라코마 이미지 분류를 위한 재현 가능한 벤치마크를 제공한다.
Recent fast style transfer methods use a pre-trained convolutional neural network as a feature encoder and a perceptual loss network. Although the pre-trained network is used to generate responses of receptive fields effective for representing style and content of image, it is not optimized for image style transfer but rather for image classification. Furthermore, it also requires a time-consuming and correlation-considering feature alignment process for image style transfer because of its inter-channel correlation. In this paper, we propose an end-to-end learning method which optimizes an encoder/decoder network for the purpose of style transfer as well as relieves the feature alignment complexity from considering inter-channel correlation. We used uncorrelation loss, i.e., the total correlation coefficient between the responses of different encoder channels, with style and content losses for training style transfer network. This makes the encoder network to be trained to generate inter-channel uncorrelated features and to be optimized for the task of image style transfer which maintained the quality of image style only with a light-weighted and correlation-unaware feature alignment process. Moreover, our method drastically reduced redundant channels of the encoded feature and this resulted in the efficient size of structure of network and faster forward processing speed. Our method can also be applied to cascade network scheme for multiple scaled style transferring and allows user-control of style strength by using a content-style trade-off parameter.
Recently, deep learning has started to play an essential role in healthcare applications, including image search in digital pathology. Despite the recent progress in computer vision, significant issues remain for image searching in histopathology archives. A well-known problem is AI bias and lack of generalization. A more particular shortcoming of deep models is the ignorance toward search functionality. The former affects every model, the latter only search and matching. Due to the lack of ranking-based learning, researchers must train models based on the classification error and then use the resultant embedding for image search purposes. Moreover, deep models appear to be prone to internal bias even if using a large image repository of various hospitals. This paper proposes two novel ideas to improve image search performance. First, we use a ranking loss function to guide feature extraction toward the matching-oriented nature of the search. By forcing the model to learn the ranking of matched outputs, the representation learning is customized toward image search instead of learning a class label. Second, we introduce the concept of sequestering learning to enhance the generalization of feature extraction. By excluding the images of the input hospital from the matched outputs, i.e., sequestering the input domain, the institutional bias is reduced. The proposed ideas are implemented and validated through the largest public dataset of whole slide images. The experiments demonstrate superior results compare to the-state-of-art.
Reinforcement Learning (RL) is a research area that has blossomed tremendously in recent years and has shown remarkable potential for artificial intelligence based opponents in computer games. This success is primarily due to vast capabilities of Convolutional Neural Networks (ConvNet), enabling algorithms to extract useful information from noisy environments. Capsule Network (CapsNet) is a recent introduction to the Deep Learning algorithm group and has only barely begun to be explored. The network is an architecture for image classification, with superior performance for classification of the MNIST dataset. CapsNets have not been explored beyond image classification. This thesis introduces the use of CapsNet for Q-Learning based game algorithms. To successfully apply CapsNet in advanced game play, three main contributions follow. First, the introduction of four new game environments as frameworks for RL research with increasing complexity, namely Flash RL, Deep Line Wars, Deep RTS, and Deep Maze. These environments fill the gap between relatively simple and more complex game environments available for RL research and are in the thesis used to test and explore the CapsNet behavior. Second, the thesis introduces a generative modeling approach to produce artificial training data for use in Deep Learning models including CapsNets. We empirically show that conditional generative modeling can successfully generate game data of sufficient quality to train a Deep Q-Network well. Third, we show that CapsNet is a reliable architecture for Deep Q-Learning based algorithms for game AI. A capsule is a group of neurons that determine the presence of objects in the data and is in the literature shown to increase the robustness of training and predictions while lowering the amount training data needed. It should, therefore, be ideally suited for game plays.
This paper addresses the few-shot image classification problem, where the classification task is performed on unlabeled query samples given a small amount of labeled support samples only. One major challenge of the few-shot learning problem is the large variety of object visual appearances that prevents the support samples to represent that object comprehensively. This might result in a significant difference between support and query samples, therefore undermining the performance of few-shot algorithms. In this paper, we tackle the problem by proposing Few-shot Cosine Transformer (FS-CT), where the relational map between supports and queries is effectively obtained for the few-shot tasks. The FS-CT consists of two parts, a learnable prototypical embedding network to obtain categorical representations from support samples with hard cases, and a transformer encoder to effectively achieve the relational map from two different support and query samples. We introduce Cosine Attention, a more robust and stable attention module that enhances the transformer module significantly and therefore improves FS-CT performance from 5% to over 20% in accuracy compared to the default scaled dot-product mechanism. Our method performs competitive results in mini-ImageNet, CUB-200, and CIFAR-FS on 1-shot learning and 5-shot learning tasks across backbones and few-shot configurations. We also developed a custom few-shot dataset for Yoga pose recognition to demonstrate the potential of our algorithm for practical application. Our FS-CT with cosine attention is a lightweight, simple few-shot algorithm that can be applied for a wide range of applications, such as healthcare, medical, and security surveillance. The official implementation code of our Few-shot Cosine Transformer is available at https://github.com/vinuni-vishc/Few-Shot-Cosine-Transformer
Following the major successes of self-attention and Transformers for image analysis, we investigate the use of such attention mechanisms in the context of Image Quality Assessment (IQA) and propose a novel full-reference IQA method, Vision Transformer for Attention Modulated Image Quality (VTAMIQ). Our method achieves competitive or state-of-the-art performance on the existing IQA datasets and significantly outperforms previous metrics in cross-database evaluations. Most patch-wise IQA methods treat each patch independently; this partially discards global information and limits the ability to model long-distance interactions. We avoid this problem altogether by employing a transformer to encode a sequence of patches as a single global representation, which by design considers interdependencies between patches. We rely on various attention mechanisms -- first with self-attention within the Transformer, and second with channel attention within our difference modulation network -- specifically to reveal and enhance the more salient features throughout our architecture. With large-scale pre-training for both classification and IQA tasks, VTAMIQ generalizes well to unseen sets of images and distortions, further demonstrating the strength of transformer-based networks for vision modelling.
A continual learning (CL) model is desired for remote sensing image analysis because of varying camera parameters, spectral ranges, resolutions, etc. There exist some recent initiatives to develop CL techniques in this domain but they still depend on massive labelled samples which do not fully fit remote sensing applications because ground truths are often obtained via field-based surveys. This paper addresses this problem with a proposal of unsupervised flat-wide learning approach (UNISA) for unsupervised few-shot continual learning approaches of remote sensing image scene classifications which do not depend on any labelled samples for its model updates. UNISA is developed from the idea of prototype scattering and positive sampling for learning representations while the catastrophic forgetting problem is tackled with the flat-wide learning approach combined with a ball generator to address the data scarcity problem. Our numerical study with remote sensing image scene datasets and a hyperspectral dataset confirms the advantages of our solution. Source codes of UNISA are shared publicly in \url{https://github.com/anwarmaxsum/UNISA} to allow convenient future studies and reproductions of our numerical results.
안경저안구 사진(CFP)의 픽셀 수준 화질 평가를 위한 최초의 공개 벤치마크 FunPiQ를 소개하며, 300장의 이미지에 전문가 픽셀 단위 품질 주석을 제공한다. 또한 해부학적 가시성 기반 의사 레이블과 Non-Negative Positive-Unlabeled(nnPU) 학습을 결합한 설명 가능한 설계(EBD) 방법 EFIQA-CP를 제안하여 기존 방법 대비 우수한 성능을 달성하였다.
This paper considers the task of thorax disease classification on chest X-ray images. Existing methods generally use the global image as input for network learning. Such a strategy is limited in two aspects. 1) A thorax disease usually happens in (small) localized areas which are disease specific. Training CNNs using global image may be affected by the (excessive) irrelevant noisy areas. 2) Due to the poor alignment of some CXR images, the existence of irregular borders hinders the network performance. In this paper, we address the above problems by proposing a three-branch attention guided convolution neural network (AG-CNN). AG-CNN 1) learns from disease-specific regions to avoid noise and improve alignment, 2) also integrates a global branch to compensate the lost discriminative cues by local branch. Specifically, we first learn a global CNN branch using global images. Then, guided by the attention heat map generated from the global branch, we inference a mask to crop a discriminative region from the global image. The local region is used for training a local CNN branch. Lastly, we concatenate the last pooling layers of both the global and local branches for fine-tuning the fusion branch. The Comprehensive experiment is conducted on the ChestX-ray14 dataset. We first report a strong global baseline producing an average AUC of 0.841 with ResNet-50 as backbone. After combining the local cues with the global information, AG-CNN improves the average AUC to 0.868. While DenseNet-121 is used, the average AUC achieves 0.871, which is a new state of the art in the community.
Convolutional neural networks (CNNs) are similar to "ordinary" neural networks in the sense that they are made up of hidden layers consisting of neurons with "learnable" parameters. These neurons receive inputs, performs a dot product, and then follows it with a non-linearity. The whole network expresses the mapping between raw image pixels and their class scores. Conventionally, the Softmax function is the classifier used at the last layer of this network. However, there have been studies (Alalshekmubarak and Smith, 2013; Agarap, 2017; Tang, 2013) conducted to challenge this norm. The cited studies introduce the usage of linear support vector machine (SVM) in an artificial neural network architecture. This project is yet another take on the subject, and is inspired by (Tang, 2013). Empirical data has shown that the CNN-SVM model was able to achieve a test accuracy of ~99.04% using the MNIST dataset (LeCun, Cortes, and Burges, 2010). On the other hand, the CNN-Softmax was able to achieve a test accuracy of ~99.23% using the same dataset. Both models were also tested on the recently-published Fashion-MNIST dataset (Xiao, Rasul, and Vollgraf, 2017), which is suppose to be a more difficult image classification dataset than MNIST (Zalandoresearch, 2017). This proved to be the case as CNN-SVM reached a test accuracy of ~90.72%, while the CNN-Softmax reached a test accuracy of ~91.86%. The said results may be improved if data preprocessing techniques were employed on the datasets, and if the base CNN model was a relatively more sophisticated than the one used in this study.