Although transformers have become the neural architectures of choice for natural language processing, they require orders of magnitude more training data, GPU memory, and computations in order to compete with convolutional neural networks for computer vision. The attention mechanism of transformers scales quadratically with the length of the input sequence, and unrolled images have long sequence lengths. Plus, transformers lack an inductive bias that is appropriate for images. We tested three modifications to vision transformer (ViT) architectures that address these shortcomings. Firstly, we alleviate the quadratic bottleneck by using linear attention mechanisms, called X-formers (such that, X in {Performer, Linformer, Nystr\"omformer}), thereby creating Vision X-formers (ViXs). This resulted in up to a seven times reduction in the GPU memory requirement. We also compared their performance with FNet and multi-layer perceptron mixers, which further reduced the GPU memory requirement. Secondly, we introduced an inductive bias for images by replacing the initial linear embedding layer by convolutional layers in ViX, which significantly increased classification accuracy without increasing the model size. Thirdly, we replaced the learnable 1D position embeddings in ViT with Rotary Position Embedding (RoPE), which increases the classification accuracy for the same model size. We believe that incorporating such changes can democratize transformers by making them accessible to those with limited data and computing resources.
With the emergence of Artificial Intelligence (AI), new attention has been given to implement AI algorithms on resource constrained tiny devices to expand the application domain of IoT. Multimodal Learning has recently become very popular with the classification task due to its impressive performance for both image and audio event classification. This paper presents TinyM$^2$Net -- a flexible system algorithm co-designed multimodal learning framework for resource constrained tiny devices. The framework was designed to be evaluated on two different case-studies: COVID-19 detection from multimodal audio recordings and battle field object detection from multimodal images and audios. In order to compress the model to implement on tiny devices, substantial network architecture optimization and mixed precision quantization were performed (mixed 8-bit and 4-bit). TinyM$^2$Net shows that even a tiny multimodal learning model can improve the classification performance than that of any unimodal frameworks. The most compressed TinyM$^2$Net achieves 88.4% COVID-19 detection accuracy (14.5% improvement from unimodal base model) and 96.8% battle field object detection accuracy (3.9% improvement from unimodal base model). Finally, we test our TinyM$^2$Net models on a Raspberry Pi 4 to see how they perform when deployed to a resource constrained tiny device.
We investigate the role of projection heads, also known as projectors, within the encoder-projector framework (e.g., SimCLR) used in contrastive learning. We aim to demystify the observed phenomenon where representations learned before projectors outperform those learned after -- measured using the downstream linear classification accuracy, even when the projectors themselves are linear. In this paper, we make two significant contributions towards this aim. Firstly, through empirical and theoretical analysis, we identify two crucial effects -- expansion and shrinkage -- induced by the contrastive loss on the projectors. In essence, contrastive loss either expands or shrinks the signal direction in the representations learned by an encoder, depending on factors such as the augmentation strength, the temperature used in contrastive loss, etc. Secondly, drawing inspiration from the expansion and shrinkage phenomenon, we propose a family of linear transformations to accurately model the projector's behavior. This enables us to precisely characterize the downstream linear classification accuracy in the high-dimensional asymptotic limit. Our findings reveal that linear projectors operating in the shrinkage (or expansion) regime hinder (or improve) the downstream classification accuracy. This provides the first theoretical explanation as to why (linear) projectors impact the downstream performance of learned representations. Our theoretical findings are further corroborated by extensive experiments on both synthetic data and real image data.
데이터 기반 세계에서 객체 간 관계를 효율적으로 구성하고 매핑하는 것은 매우 중요하다. 그래프는 이러한 연결을 모델링하는 강력한 도구로, 소셜 네트워크, 통신, 생물학 등에 널리 사용된다. 그러나 그래프 기반 방법은 종종 높은 계산 비용, 특히 메모리 및 공간 사용 측면에서 부담을 겪는다. 이에 대응하기 위해 그래프 임베딩 기법, 즉 네트워크 표현 학습(Network Representation Learning)이라고도 불리는 방법은 구조적 측면을 보존하면서 그래프 정보를 저차원 표현으로 인코딩한다. 그러나 전통적인 방법은 해석 가능한 차원을 제공하지 못한다. RaDE(Rank Diffusion Embedding)는 순위 기반 정보를 활용하는 새로운 접근법을 제안하며, 핵심 단계로 대표적인 노드 부분집합을 선택하여 차원에 대한 해석 가능성을 부여하고 검색(retrieval) 작업을 개선한다. 잠재력에도 불구하고 RaDE의 원래 제안은 서로 다른 클래스에 걸친 대표 부분집합 선택의 효과를 충분히 탐구하지 않았고, 분류 및 클러스터링과 같은 작업에서 임베딩을 평가하지도 않았다. RaDE에서 영감을 받아, 본 연구는 GRaCE(Graph and Rank-based Contextual Embeddings)를 제안한다. 이는 대표 부분집합 선택 및 노드 임베딩을 위해 견고한 순위 기반 측도를 활용하여 해석 가능한 임베딩을 생성하는 완전 비지도 프레임워크이다. GRaCE는 텍스트 및 이미지 컬렉션을 포함한 다양한 데이터셋에서 RaDE와 원본 특징(Original Features)을 능가하며, 최신 트랜스포머 모델을 특징 서술자(feature descriptor)로, 분류 작업에서는 Graph Convolutional Networks 모델을 함께 고려할 때 검색, 분류, 클러스터링 작업에서 우수한 성능을 보인다.
Most image instance retrieval pipelines are based on comparison of vectors known as global image descriptors between a query image and the database images. Due to their success in large scale image classification, representations extracted from Convolutional Neural Networks (CNN) are quickly gaining ground on Fisher Vectors (FVs) as state-of-the-art global descriptors for image instance retrieval. While CNN-based descriptors are generally remarked for good retrieval performance at lower bitrates, they nevertheless present a number of drawbacks including the lack of robustness to common object transformations such as rotations compared with their interest point based FV counterparts. In this paper, we propose a method for computing invariant global descriptors from CNNs. Our method implements a recently proposed mathematical theory for invariance in a sensory cortex modeled as a feedforward neural network. The resulting global descriptors can be made invariant to multiple arbitrary transformation groups while retaining good discriminativeness. Based on a thorough empirical evaluation using several publicly available datasets, we show that our method is able to significantly and consistently improve retrieval results every time a new type of invariance is incorporated. We also show that our method which has few parameters is not prone to overfitting: improvements generalize well across datasets with different properties with regard to invariances. Finally, we show that our descriptors are able to compare favourably to other state-of-the-art compact descriptors in similar bitranges, exceeding the highest retrieval results reported in the literature on some datasets. A dedicated dimensionality reduction step --quantization or hashing-- may be able to further improve the competitiveness of the descriptors.
Visual Question Answering (VQA) within the surgical domain, utilizing Large Language Models (LLMs), offers a distinct opportunity to improve intra-operative decision-making and facilitate intuitive surgeon-AI interaction. However, the development of LLMs for surgical VQA is hindered by the scarcity of diverse and extensive datasets with complex reasoning tasks. Moreover, contextual fusion of the image and text modalities remains an open research challenge due to the inherent differences between these two types of information and the complexity involved in aligning them. This paper introduces PitVQA, a novel dataset specifically designed for VQA in endonasal pituitary surgery and PitVQA-Net, an adaptation of the GPT2 with a novel image-grounded text embedding for surgical VQA. PitVQA comprises 25 procedural videos and a rich collection of question-answer pairs spanning crucial surgical aspects such as phase and step recognition, context understanding, tool detection and localization, and tool-tissue interactions. PitVQA-Net consists of a novel image-grounded text embedding that projects image and text features into a shared embedding space and GPT2 Backbone with an excitation block classification head to generate contextually relevant answers within the complex domain of endonasal pituitary surgery. Our image-grounded text embedding leverages joint embedding, cross-attention and contextual representation to understand the contextual relationship between questions and surgical images. We demonstrate the effectiveness of PitVQA-Net on both the PitVQA and the publicly available EndoVis18-VQA dataset, achieving improvements in balanced accuracy of 8% and 9% over the most recent baselines, respectively. Our code and dataset is available at https://github.com/mobarakol/PitVQA.
표-이미지(tabular-to-image) 방법은 합성곱 신경망(convolutional neural network)과 비전 트랜스포머(vision transformer)의 높은 예측 성능을 활용하기 위한 새로운 접근으로 등장했다. 이들은 표 형식 데이터를 이미지 표현으로 변환하며, 각 특징을 차원 축소 방법(예: t-SNE, UMAP, PCA)에서 도출한 고정 픽셀 위치에 매핑한다. 그러나 이들은 각 특징의 주변(marginal) 값만 인코딩하고 특징 간 관계에 대한 정보는 버린다. 우리는 자기조직화 맵(Self-Organizing Map, SOM) 위에 구축된 표-이미지 인코딩인 TabSOM을 제안하며, 이는 (i) 충돌 없는 헝가리안 할당(Hungarian assignment)을 통해 구성요소 평면(component plane)에서 도출한 고정 캔버스 위치에 모든 입력 특징이 배치되는 공간 레이아웃과 (ii) SOM 구성요소 평면에서 도출한 특징 쌍(pairwise) 관계를 포착하는 그래프를 제공한다. 결과 이미지는 두 개의 다중 스케일 노드 채널을 쌓는다: 하나는 고정 스케일에서 특징 값을 인코딩하고, 다른 하나는 관련 특징 간 공간 연결로 특징 쌍 상호작용을 인코딩한다. SOM에서 파생된 두 가지 해석 가능성(interpretability) 접근을 도입한다: 프로토타입 영감 부분 의존 플롯(prototype-inspired partial dependence plot)과 클래스 분리(class-separation) 중요도 점수. 공개 이진 분류(binary-classification) 데이터셋에서 기존 12개 표-이미지 방법과 벤치마크한 결과, TabSOM은 모든 데이터셋에서 1위 또는 2위를 차지하며 평가된 방법 중 가장 낮은 분산을 달성했다. TabSOM으로 얻은 해석 가능성은 Random Forest, XGBoost, SHAP과 대비하여 검증되었으며, 클래스 분리 점수는 상위 순위 특징에서 기존 기준선과 합리적인 일치를 보이면서 입력 데이터로부터 보완적인 구조 정보를 포착한다. 이 결과는 TabSOM이 표 형식 데이터에 심층 학습 아키텍처를 적용하는 데 효과적이고 해석 가능한 접근을 제공하며, 이 영역에서 성능-해석 가능성 격차를 메운다는 것을 보여준다.
Masked image modeling (MIM) has shown great promise for self-supervised learning (SSL) yet been criticized for learning inefficiency. We believe the insufficient utilization of training signals should be responsible. To alleviate this issue, we introduce a conceptually simple yet learning-efficient MIM training scheme, termed Disjoint Masking with Joint Distillation (DMJD). For disjoint masking (DM), we sequentially sample multiple masked views per image in a mini-batch with the disjoint regulation to raise the usage of tokens for reconstruction in each image while keeping the masking rate of each view. For joint distillation (JD), we adopt a dual branch architecture to respectively predict invisible (masked) and visible (unmasked) tokens with superior learning targets. Rooting in orthogonal perspectives for training efficiency improvement, DM and JD cooperatively accelerate the training convergence yet not sacrificing the model generalization ability. Concretely, DM can train ViT with half of the effective training epochs (3.7 times less time-consuming) to report competitive performance. With JD, our DMJD clearly improves the linear probing classification accuracy over ConvMAE by 5.8%. On fine-grained downstream tasks like semantic segmentation, object detection, etc., our DMJD also presents superior generalization compared with state-of-the-art SSL methods. The code and model will be made public at https://github.com/mx-mark/DMJD.
Medical image synthesis has emerged as a promising solution to address the limited availability of annotated medical data needed for training machine learning algorithms in the context of image-based Clinical Decision Support (CDS) systems. To this end, Generative Adversarial Networks (GANs) have been mainly applied to support the algorithm training process by generating synthetic images for data augmentation. However, in the field of Wireless Capsule Endoscopy (WCE), the limited content diversity and size of existing publicly available annotated datasets, adversely affect both the training stability and synthesis performance of GANs. Aiming to a viable solution for WCE image synthesis, a novel Variational Autoencoder architecture is proposed, namely "This Intestine Does not Exist" (TIDE). The proposed architecture comprises multiscale feature extraction convolutional blocks and residual connections, which enable the generation of high-quality and diverse datasets even with a limited number of training images. Contrary to the current approaches, which are oriented towards the augmentation of the available datasets, this study demonstrates that using TIDE, real WCE datasets can be fully substituted by artificially generated ones, without compromising classification performance. Furthermore, qualitative and user evaluation studies by experienced WCE specialists, validate from a medical viewpoint that both the normal and abnormal WCE images synthesized by TIDE are sufficiently realistic.
Neural image classifiers can often learn to make predictions by overly relying on non-predictive features that are spuriously correlated with the class labels in the training data. This leads to poor performance in real-world atypical scenarios where such features are absent. This paper presents ASPIRE (Language-guided Data Augmentation for SPurIous correlation REmoval), a simple yet effective solution for supplementing the training dataset with images without spurious features, for robust learning against spurious correlations via better generalization. ASPIRE, guided by language at various steps, can generate non-spurious images without requiring any group labeling or existing non-spurious images in the training set. Precisely, we employ LLMs to first extract foreground and background features from textual descriptions of an image, followed by advanced language-guided image editing to discover the features that are spuriously correlated with the class label. Finally, we personalize a text-to-image generation model using the edited images to generate diverse in-domain images without spurious features. ASPIRE is complementary to all prior robust training methods in literature, and we demonstrate its effectiveness across 4 datasets and 9 baselines and show that ASPIRE improves the worst-group classification accuracy of prior methods by 1% - 38%. We also contribute a novel test set for the challenging Hard ImageNet dataset.
The accuracy and robustness of machine learning models against adversarial attacks are significantly influenced by factors such as training data quality, model architecture, the training process, and the deployment environment. In recent years, duplicated data in training sets, especially in language models, has attracted considerable attention. It has been shown that deduplication enhances both training performance and model accuracy in language models. While the importance of data quality in training image classifier Deep Neural Networks (DNNs) is widely recognized, the impact of duplicated images in the training set on model generalization and performance has received little attention. In this paper, we address this gap and provide a comprehensive study on the effect of duplicates in image classification. Our analysis indicates that the presence of duplicated images in the training set not only negatively affects the efficiency of model training but also may result in lower accuracy of the image classifier. This negative impact of duplication on accuracy is particularly evident when duplicated data is non-uniform across classes or when duplication, whether uniform or non-uniform, occurs in the training set of an adversarially trained model. Even when duplicated samples are selected in a uniform way, increasing the amount of duplication does not lead to a significant improvement in accuracy.
Remote sensing increasingly captures the Earth's surface through complementary sensors such as SAR and optical, yet ground-truth annotations remain scarce, limiting supervised approaches. Self-supervised learning (SSL) addresses the label scarcity, and cross-modal SSL further exploits the complementarity between co-located sensors, but existing cross-modal frameworks align modalities mostly through global image-level representations or treat them as largely redundant, leaving their fine-grained complementary structure underexploited. We propose CroCoMAE, a cross-modal SSL framework for SAR-optical remote sensing that unifies contrastive learning and masked image modeling within a ViT-based architecture, operating at two complementary levels. At the image level, an intra-modality contrastive loss on the class token yields globally discriminative, augmentation-invariant representations. At the patch level, masked image modeling drives fine-grained per-modality representations, while a cross-modal contrastive objective aligns a shared subspace across SAR and optical patch tokens, leaving the complementary subspace unconstrained to retain modality-specific information. On two SAR-optical benchmarks covering scene classification and semantic segmentation, CroCoMAE improves over uni-modal SSL baselines and outperforms state-of-the-art cross-modal frameworks in most configurations, with gains in both single-modality and multi-modal evaluation across varying label regimes and the largest improvements on semantic segmentation.
Driven by successes in deep learning, computer vision research has begun to move beyond object detection and image classification to more sophisticated tasks like image captioning or visual question answering. Motivating such endeavors is the desire for models to capture not only objects present in an image, but more fine-grained aspects of a scene such as relationships between objects and their attributes. Scene graphs provide a formal construct for capturing these aspects of an image. Despite this, there have been only a few recent efforts to generate scene graphs from imagery. Previous works limit themselves to settings where bounding box information is available at train time and do not attempt to generate scene graphs with attributes. In this paper we propose a method, based on recent advancements in Generative Adversarial Networks, to overcome these deficiencies. We take the approach of first generating small subgraphs, each describing a single statement about a scene from a specific region of the input image chosen using an attention mechanism. By doing so, our method is able to produce portions of the scene graphs with attribute information without the need for bounding box labels. Then, the complete scene graph is constructed from these subgraphs. We show that our model improves upon prior work in scene graph generation on state-of-the-art data sets and accepted metrics. Further, we demonstrate that our model is capable of handling a larger vocabulary size than prior work has attempted.
Reducing the data footprint of visual content via image compression is essential to reduce storage requirements, but also to reduce the bandwidth and latency requirements for transmission. In particular, the use of compressed images allows for faster transfer of data, and faster response times for visual recognition in edge devices that rely on cloud-based services. In this paper, we first analyze the impact of image compression using traditional codecs, as well as recent state-of-the-art neural compression approaches, on three visual recognition tasks: image classification, object detection, and semantic segmentation. We consider a wide range of compression levels, ranging from 0.1 to 2 bits-per-pixel (bpp). We find that for all three tasks, the recognition ability is significantly impacted when using strong compression. For example, for segmentation mIoU is reduced from 44.5 to 30.5 mIoU when compressing to 0.1 bpp using the best compression model we evaluated. Second, we test to what extent this performance drop can be ascribed to a loss of relevant information in the compressed image, or to a lack of generalization of visual recognition models to images with compression artefacts. We find that to a large extent the performance loss is due to the latter: by finetuning the recognition models on compressed training images, most of the performance loss is recovered. For example, bringing segmentation accuracy back up to 42 mIoU, i.e. recovering 82% of the original drop in accuracy.
Fully convolutional neural networks (F-CNNs) have set the state-of-the-art in image segmentation for a plethora of applications. Architectural innovations within F-CNNs have mainly focused on improving spatial encoding or network connectivity to aid gradient flow. In this paper, we explore an alternate direction of recalibrating the feature maps adaptively, to boost meaningful features, while suppressing weak ones. We draw inspiration from the recently proposed squeeze & excitation (SE) module for channel recalibration of feature maps for image classification. Towards this end, we introduce three variants of SE modules for image segmentation, (i) squeezing spatially and exciting channel-wise (cSE), (ii) squeezing channel-wise and exciting spatially (sSE) and (iii) concurrent spatial and channel squeeze & excitation (scSE). We effectively incorporate these SE modules within three different state-of-the-art F-CNNs (DenseNet, SD-Net, U-Net) and observe consistent improvement of performance across all architectures, while minimally effecting model complexity. Evaluations are performed on two challenging applications: whole brain segmentation on MRI scans (Multi-Atlas Labelling Challenge Dataset) and organ segmentation on whole body contrast enhanced CT scans (Visceral Dataset).
In this work we propose a neural network based image descriptor suitable for image patch matching, which is an important task in many computer vision applications. Our approach is influenced by recent success of deep convolutional neural networks (CNNs) in object detection and classification tasks. We develop a model which maps the raw input patch to a low dimensional feature vector so that the distance between representations is small for similar patches and large otherwise. As a distance metric we utilize L2 norm, i.e. Euclidean distance, which is fast to evaluate and used in most popular hand-crafted descriptors, such as SIFT. According to the results, our approach outperforms state-of-the-art L2-based descriptors and can be considered as a direct replacement of SIFT. In addition, we conducted experiments with batch normalization and histogram equalization as a preprocessing method of the input data. The results confirm that these techniques further improve the performance of the proposed descriptor. Finally, we show promising preliminary results by appending our CNNs with recently proposed spatial transformer networks and provide a visualisation and interpretation of their impact.
The task of aesthetic quality assessment is complicated due to its subjectivity. In recent years, the target representation of image aesthetic quality has changed from a one-dimensional binary classification label or numerical score to a multi-dimensional score distribution. According to current methods, the ground truth score distributions are straightforwardly regressed. However, the subjectivity of aesthetics is not taken into account, that is to say, the psychological processes of human beings are not taken into consideration, which limits the performance of the task. In this paper, we propose a Deep Drift-Diffusion (DDD) model inspired by psychologists to predict aesthetic score distribution from images. The DDD model can describe the psychological process of aesthetic perception instead of traditional modeling of the results of assessment. We use deep convolution neural networks to regress the parameters of the drift-diffusion model. The experimental results in large scale aesthetic image datasets reveal that our novel DDD model is simple but efficient, which outperforms the state-of-the-art methods in aesthetic score distribution prediction. Besides, different psychological processes can also be predicted by our model.
Elliptically-contoured distributions (ECD) play a significant role, in computer vision, image processing, radar, and biomedical signal processing. Maximum likelihood. estimation (MLE) of ECD leads to a system of non-linear equations, most-often addressed using fixed-point (FP) methods. Unfortunately, the computation time required for these methods is unacceptably long, for large-scale or high-dimensional datasets. To overcome this difficulty, the present work introduces a Riemannian optimisation method, the information stochastic gradient (ISG). The ISG is an online (recursive) method, which achieves the same performance as MLE, for large-scale datasets, while requiring modest memory and time resources. To develop the ISG method, the Riemannian information gradient is derived taking into account the product manifold associated to the underlying parameter space of the ECD. From this information gradient definition, we define also, the information deterministic gradient (IDG), an offline (batch) method, which is an alternative, for moderate-sized datasets. The present work formulates these two methods, and demonstrates their performance through numerical simulations. Two applications, to image re-colorization, and to texture classification, are also worked out.
Machine learning models are known to be vulnerable to adversarial evasion attacks as illustrated by image classification models. Thoroughly understanding such attacks is critical in order to ensure the safety and robustness of critical AI tasks. However, most evasion attacks are difficult to deploy against a majority of AI systems because they have focused on image domain with only few constraints. An image is composed of homogeneous, numerical, continuous, and independent features, unlike many other input types to AI systems used in practice. Furthermore, some input types include additional semantic and functional constraints that must be observed to generate realistic adversarial inputs. In this work, we propose a new framework to enable the generation of adversarial inputs irrespective of the input type and task domain. Given an input and a set of pre-defined input transformations, our framework discovers a sequence of transformations that result in a semantically correct and functional adversarial input. We demonstrate the generality of our approach on several diverse machine learning tasks with various input representations. We also show the importance of generating adversarial examples as they enable the deployment of mitigation techniques.
Patient-level diagnosis of severity in ulcerative colitis (UC) is common in real clinical settings, where the most severe score in a patient is recorded. However, previous UC classification methods (i.e., image-level estimation) mainly assumed the input was a single image. Thus, these methods can not utilize severity labels recorded in real clinical settings. In this paper, we propose a patient-level severity estimation method by a transformer with selective aggregator tokens, where a severity label is estimated from multiple images taken from a patient, similar to a clinical setting. Our method can effectively aggregate features of severe parts from a set of images captured in each patient, and it facilitates improving the discriminative ability between adjacent severity classes. Experiments demonstrate the effectiveness of the proposed method on two datasets compared with the state-of-the-art MIL methods. Moreover, we evaluated our method in real clinical settings and confirmed that our method outperformed the previous image-level methods. The code is publicly available at https://github.com/Shiku-Kaito/Ordinal-Multiple-instance-Learning-for-Ulcerative-Colitis-Severity-Estimation.