Contrastive pretraining of image-text foundation models, such as CLIP, demonstrated excellent zero-shot performance and improved robustness on a wide range of downstream tasks. However, these models utilize large transformer-based encoders with significant memory and latency overhead which pose challenges for deployment on mobile devices. In this work, we introduce MobileCLIP -- a new family of efficient image-text models optimized for runtime performance along with a novel and efficient training approach, namely multi-modal reinforced training. The proposed training approach leverages knowledge transfer from an image captioning model and an ensemble of strong CLIP encoders to improve the accuracy of efficient models. Our approach avoids train-time compute overhead by storing the additional knowledge in a reinforced dataset. MobileCLIP sets a new state-of-the-art latency-accuracy tradeoff for zero-shot classification and retrieval tasks on several datasets. Our MobileCLIP-S2 variant is 2.3$\times$ faster while more accurate compared to previous best CLIP model based on ViT-B/16. We further demonstrate the effectiveness of our multi-modal reinforced training by training a CLIP model based on ViT-B/16 image backbone and achieving +2.9% average performance improvement on 38 evaluation benchmarks compared to the previous best. Moreover, we show that the proposed approach achieves 10$\times$-1000$\times$ improved learning efficiency when compared with non-reinforced CLIP training. Code and models are available at https://github.com/apple/ml-mobileclip .
Deep learning models (DLMs) frequently achieve accurate segmentation and classification of tumors from medical images. However, DLMs lacking feedback on their image segmentation mechanisms, such as Dice coefficients and confidence in their performance, face challenges when processing previously unseen images in real-world clinical settings. Uncertainty estimates to identify DLM predictions at the cellular or single-pixel level that require clinician review can enhance trust. However, their deployment requires significant computational resources. This study reports two DLMs, one trained from scratch and another based on transfer learning, with Monte Carlo dropout or Bayes-by-backprop uncertainty estimations to segment lesions from the publicly available The International Skin Imaging Collaboration-19 dermoscopy image database with cancerous lesions. A novel approach to compute pixel-by-pixel uncertainty estimations of DLM segmentation performance in multiple clinical regions from a single dermoscopy image with corresponding Dice scores is reported for the first time. Image-level uncertainty maps demonstrated correspondence between imperfect DLM segmentation and high uncertainty levels in specific skin tissue regions, with or without lesions. Four new linear regression models that can predict the Dice performance of DLM segmentation using constants and uncertainty measures, either individually or in combination from lesions, tissue structures, and non-tissue pixel regions critical for clinical diagnosis and prognostication in skin images (Spearman's correlation, p < 0.05), are reported for the first time for low-compute uncertainty estimation workflows.
Capsule network is a type of neural network that uses the spatial relationship between features to classify images. By capturing the poses and relative positions between features, its ability to recognize affine transformation is improved, and it surpasses traditional convolutional neural networks (CNNs) when handling translation, rotation and scaling. The Stacked Capsule Autoencoder (SCAE) is the state-of-the-art capsule network. The SCAE encodes an image as capsules, each of which contains poses of features and their correlations. The encoded contents are then input into the downstream classifier to predict the categories of the images. Existing research mainly focuses on the security of capsule networks with dynamic routing or EM routing, and little attention has been given to the security and robustness of the SCAE. In this paper, we propose an evasion attack against the SCAE. After a perturbation is generated based on the output of the object capsules in the model, it is added to an image to reduce the contribution of the object capsules related to the original category of the image so that the perturbed image will be misclassified. We evaluate the attack using an image classification experiment, and the experimental results indicate that the attack can achieve high success rates and stealthiness. It confirms that the SCAE has a security vulnerability whereby it is possible to craft adversarial samples without changing the original structure of the image to fool the classifiers. We hope that our work will make the community aware of the threat of this attack and raise the attention given to the SCAE's security.
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.
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.
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.
Model explanation techniques play a critical role in understanding the source of a model's performance and making its decisions transparent. Here we investigate if explanation techniques can also be used as a mechanism for scientific discovery. We make three contributions: first, we propose a framework to convert predictions from explanation techniques to a mechanism of discovery. Second, we show how generative models in combination with black-box predictors can be used to generate hypotheses (without human priors) that can be critically examined. Third, with these techniques we study classification models for retinal images predicting Diabetic Macular Edema (DME), where recent work showed that a CNN trained on these images is likely learning novel features in the image. We demonstrate that the proposed framework is able to explain the underlying scientific mechanism, thus bridging the gap between the model's performance and human understanding.
In computer vision, models must be able to adapt to changes in image resolution to effectively carry out tasks such as image segmentation; This is known as scale-equivariance. Recent works have made progress in developing scale-equivariant convolutional neural networks, e.g., through weight-sharing and kernel resizing. However, these networks are not truly scale-equivariant in practice. Specifically, they do not consider anti-aliasing as they formulate the down-scaling operation in the continuous domain. To address this shortcoming, we directly formulate down-scaling in the discrete domain with consideration of anti-aliasing. We then propose a novel architecture based on Fourier layers to achieve truly scale-equivariant deep nets, i.e., absolute zero equivariance-error. Following prior works, we test this model on MNIST-scale and STL-10 datasets. Our proposed model achieves competitive classification performance while maintaining zero equivariance-error.
Neural networks are vulnerable to adversarial examples, which poses a threat to their application in security sensitive systems. We propose high-level representation guided denoiser (HGD) as a defense for image classification. Standard denoiser suffers from the error amplification effect, in which small residual adversarial noise is progressively amplified and leads to wrong classifications. HGD overcomes this problem by using a loss function defined as the difference between the target model's outputs activated by the clean image and denoised image. Compared with ensemble adversarial training which is the state-of-the-art defending method on large images, HGD has three advantages. First, with HGD as a defense, the target model is more robust to either white-box or black-box adversarial attacks. Second, HGD can be trained on a small subset of the images and generalizes well to other images and unseen classes. Third, HGD can be transferred to defend models other than the one guiding it. In NIPS competition on defense against adversarial attacks, our HGD solution won the first place and outperformed other models by a large margin.
In existing visual representation learning tasks, deep convolutional neural networks (CNNs) are often trained on images annotated with single tags, such as ImageNet. However, a single tag cannot describe all important contents of one image, and some useful visual information may be wasted during training. In this work, we propose to train CNNs from images annotated with multiple tags, to enhance the quality of visual representation of the trained CNN model. To this end, we build a large-scale multi-label image database with 18M images and 11K categories, dubbed Tencent ML-Images. We efficiently train the ResNet-101 model with multi-label outputs on Tencent ML-Images, taking 90 hours for 60 epochs, based on a large-scale distributed deep learning framework,i.e.,TFplus. The good quality of the visual representation of the Tencent ML-Images checkpoint is verified through three transfer learning tasks, including single-label image classification on ImageNet and Caltech-256, object detection on PASCAL VOC 2007, and semantic segmentation on PASCAL VOC 2012. The Tencent ML-Images database, the checkpoints of ResNet-101, and all the training codehave been released at https://github.com/Tencent/tencent-ml-images. It is expected to promote other vision tasks in the research and industry community.
Recently, there has been increasing interest in multimodal applications that integrate text with other modalities, such as images, audio and video, to facilitate natural language interactions with multimodal AI systems. While applications involving standard modalities have been extensively explored, there is still a lack of investigation into specific data modalities such as remote sensing (RS) data. Despite the numerous potential applications of RS data, including environmental protection, disaster monitoring and land planning, available solutions are predominantly focused on specific tasks like classification, captioning and retrieval. These solutions often overlook the unique characteristics of RS data, such as its capability to systematically provide information on the same geographical areas over time. This ability enables continuous monitoring of changes in the underlying landscape. To address this gap, we propose a novel foundation model for bi-temporal RS image pairs, in the context of change detection analysis, leveraging Contrastive Learning and the LEVIR-CC dataset for both captioning and text-image retrieval. By jointly training a contrastive encoder and captioning decoder, our model add text-image retrieval capabilities, in the context of bi-temporal change detection, while maintaining captioning performances that are comparable to the state of the art. We release the source code and pretrained weights at: https://github.com/rogerferrod/RSICRC.
This paper examines the mathematical foundations of transformer architectures, highlighting their limitations particularly in handling long sequences. We explore prerequisite models such as Mamba, Vision Mamba (ViM), and LV-ViT that pave the way for our proposed architecture, DeMansia. DeMansia integrates state space models with token labeling techniques to enhance performance in image classification tasks, efficiently addressing the computational challenges posed by traditional transformers. The architecture, benchmark, and comparisons with contemporary models demonstrate DeMansia's effectiveness. The implementation of this paper is available on GitHub at https://github.com/catalpaaa/DeMansia
In recent years, remarkable advancements have been achieved in the field of image generation, primarily driven by the escalating demand for high-quality outcomes across various image generation subtasks, such as inpainting, denoising, and super resolution. A major effort is devoted to exploring the application of super-resolution techniques to enhance the quality of low-resolution images. In this context, our method explores in depth the problem of ship image super resolution, which is crucial for coastal and port surveillance. We investigate the opportunity given by the growing interest in text-to-image diffusion models, taking advantage of the prior knowledge that such foundation models have already learned. In particular, we present a diffusion-model-based architecture that leverages text conditioning during training while being class-aware, to best preserve the crucial details of the ships during the generation of the super-resoluted image. Since the specificity of this task and the scarcity availability of off-the-shelf data, we also introduce a large labeled ship dataset scraped from online ship images, mostly from ShipSpotting\footnote{\url{www.shipspotting.com}} website. Our method achieves more robust results than other deep learning models previously employed for super resolution, as proven by the multiple experiments performed. Moreover, we investigate how this model can benefit downstream tasks, such as classification and object detection, thus emphasizing practical implementation in a real-world scenario. Experimental results show flexibility, reliability, and impressive performance of the proposed framework over state-of-the-art methods for different tasks. The code is available at: https://github.com/LuigiSigillo/ShipinSight .
Research of adversarial attacks is important for AI security because it shows the vulnerability of deep learning models and helps to build more robust models. Adversarial attacks on images are most widely studied, which include noise-based attacks, image editing-based attacks, and latent space-based attacks. However, the adversarial examples crafted by these methods often lack sufficient semantic information, making it challenging for humans to understand the failure modes of deep learning models under natural conditions. To address this limitation, we propose a natural language induced adversarial image attack method. The core idea is to leverage a text-to-image model to generate adversarial images given input prompts, which are maliciously constructed to lead to misclassification for a target model. To adopt commercial text-to-image models for synthesizing more natural adversarial images, we propose an adaptive genetic algorithm (GA) for optimizing discrete adversarial prompts without requiring gradients and an adaptive word space reduction method for improving query efficiency. We further used CLIP to maintain the semantic consistency of the generated images. In our experiments, we found that some high-frequency semantic information such as "foggy", "humid", "stretching", etc. can easily cause classifier errors. This adversarial semantic information exists not only in generated images but also in photos captured in the real world. We also found that some adversarial semantic information can be transferred to unknown classification tasks. Furthermore, our attack method can transfer to different text-to-image models (e.g., Midjourney, DALL-E 3, etc.) and image classifiers. Our code is available at: https://github.com/zxp555/Natural-Language-Induced-Adversarial-Images.
With the advent of large pre-trained vision-language models such as CLIP, prompt learning methods aim to enhance the transferability of the CLIP model. They learn the prompt given few samples from the downstream task given the specific class names as prior knowledge, which we term as semantic-aware classification. However, in many realistic scenarios, we only have access to few samples and knowledge of the class names (e.g., when considering instances of classes). This challenging scenario represents the semantic-agnostic discriminative case. Text-to-Image (T2I) personalization methods aim to adapt T2I models to unseen concepts by learning new tokens and endowing these tokens with the capability of generating the learned concepts. These methods do not require knowledge of class names as a semantic-aware prior. Therefore, in this paper, we first explore Textual Inversion and reveal that the new concept tokens possess both generation and classification capabilities by regarding each category as a single concept. However, learning classifiers from single-concept textual inversion is limited since the learned tokens are suboptimal for the discriminative tasks. To mitigate this issue, we propose Multi-Class textual inversion, which includes a discriminative regularization term for the token updating process. Using this technique, our method MC-TI achieves stronger Semantic-Agnostic Classification while preserving the generation capability of these modifier tokens given only few samples per category. In the experiments, we extensively evaluate MC-TI on 12 datasets covering various scenarios, which demonstrates that MC-TI achieves superior results in terms of both classification and generation outcomes.
Multi-label zero-shot learning (ZSL) is a more realistic counter-part of standard single-label ZSL since several objects can co-exist in a natural image. However, the occurrence of multiple objects complicates the reasoning and requires region-specific processing of visual features to preserve their contextual cues. We note that the best existing multi-label ZSL method takes a shared approach towards attending to region features with a common set of attention maps for all the classes. Such shared maps lead to diffused attention, which does not discriminatively focus on relevant locations when the number of classes are large. Moreover, mapping spatially-pooled visual features to the class semantics leads to inter-class feature entanglement, thus hampering the classification. Here, we propose an alternate approach towards region-based discriminability-preserving multi-label zero-shot classification. Our approach maintains the spatial resolution to preserve region-level characteristics and utilizes a bi-level attention module (BiAM) to enrich the features by incorporating both region and scene context information. The enriched region-level features are then mapped to the class semantics and only their class predictions are spatially pooled to obtain image-level predictions, thereby keeping the multi-class features disentangled. Our approach sets a new state of the art on two large-scale multi-label zero-shot benchmarks: NUS-WIDE and Open Images. On NUS-WIDE, our approach achieves an absolute gain of 6.9% mAP for ZSL, compared to the best published results.
Federated learning (FL) has been facilitating privacy-preserving deep learning in many walks of life such as medical image classification, network intrusion detection, and so forth. Whereas it necessitates a central parameter server for model aggregation, which brings about delayed model communication and vulnerability to adversarial attacks. A fully decentralized architecture like Swarm Learning allows peer-to-peer communication among distributed nodes, without the central server. One of the most challenging issues in decentralized deep learning is that data owned by each node are usually non-independent and identically distributed (non-IID), causing time-consuming convergence of model training. To this end, we propose a decentralized learning model called Homogeneous Learning (HL) for tackling non-IID data with a self-attention mechanism. In HL, training performs on each round's selected node, and the trained model of a node is sent to the next selected node at the end of each round. Notably, for the selection, the self-attention mechanism leverages reinforcement learning to observe a node's inner state and its surrounding environment's state, and find out which node should be selected to optimize the training. We evaluate our method with various scenarios for an image classification task. The result suggests that HL can produce a better performance compared with standalone learning and greatly reduce both the total training rounds by 50.8% and the communication cost by 74.6% compared with random policy-based decentralized learning for training on non-IID data.
Federated learning (FL) enables multiple clients to collaboratively train a global model without disclosing their data. Previous researches often require training the complete model parameters. However, the emergence of powerful pre-trained models makes it possible to achieve higher performance with fewer learnable parameters in FL. In this paper, we propose a federated adaptive prompt tuning algorithm, FedAPT, for multi-domain collaborative image classification with powerful foundation models, like CLIP. Compared with direct federated prompt tuning, our core idea is to adaptively unlock specific domain knowledge for each test sample in order to provide them with personalized prompts. To implement this idea, we design an adaptive prompt tuning module, which consists of a meta prompt, an adaptive network, and some keys. The server randomly generates a set of keys and assigns a unique key to each client. Then all clients cooperatively train the global adaptive network and meta prompt with the local datasets and the frozen keys. Ultimately, the global aggregation model can assign a personalized prompt to CLIP based on the domain features of each test sample. We perform extensive experiments on two multi-domain image classification datasets across two different settings -- supervised and unsupervised. The results show that FedAPT can achieve better performance with less than 10\% of the number of parameters of the fully trained model, and the global model can perform well in diverse client domains simultaneously. The source code is available at \url{https://github.com/leondada/FedAPT}.
Fairness has been a critical issue that affects the adoption of deep learning models in real practice. To improve model fairness, many existing methods have been proposed and evaluated to be effective in their own contexts. However, there is still no systematic evaluation among them for a comprehensive comparison under the same context, which makes it hard to understand the performance distinction among them, hindering the research progress and practical adoption of them. To fill this gap, this paper endeavours to conduct the first large-scale empirical study to comprehensively compare the performance of existing state-of-the-art fairness improving techniques. Specifically, we target the widely-used application scenario of image classification, and utilized three different datasets and five commonly-used performance metrics to assess in total 13 methods from diverse categories. Our findings reveal substantial variations in the performance of each method across different datasets and sensitive attributes, indicating over-fitting on specific datasets by many existing methods. Furthermore, different fairness evaluation metrics, due to their distinct focuses, yield significantly different assessment results. Overall, we observe that pre-processing methods and in-processing methods outperform post-processing methods, with pre-processing methods exhibiting the best performance. Our empirical study offers comprehensive recommendations for enhancing fairness in deep learning models. We approach the problem from multiple dimensions, aiming to provide a uniform evaluation platform and inspire researchers to explore more effective fairness solutions via a set of implications.