In this paper, we question if well pre-trained vision transformer (ViT) models could be used as teachers that exhibit scalable properties to advance cross architecture knowledge distillation (KD) research, in the context of using large-scale datasets for evaluation. To make this possible, our analysis underlines the importance of seeking effective strategies to align (1) feature computing paradigm differences, (2) model scale differences, and (3) knowledge density differences. By combining three coupled components namely cross attention projector, dual-view feature mimicking and teacher parameter perception tailored to address the above problems, we present a simple and effective KD method, called ScaleKD. Our method can train student backbones that span across a variety of convolutional neural network (CNN), multi-layer perceptron (MLP), and ViT architectures on image classification datasets, achieving state-of-the-art distillation performance. For instance, taking a well pre-trained Swin-L as the teacher model, our method gets 75.15%|82.03%|84.16%|78.63%|81.96%|83.93%|83.80%|85.53% top-1 accuracies for MobileNet-V1|ResNet-50|ConvNeXt-T|Mixer-S/16|Mixer-B/16|ViT-S/16|Swin-T|ViT-B/16 models trained on ImageNet-1K dataset from scratch, showing 3.05%|3.39%|2.02%|4.61%|5.52%|4.03%|2.62%|3.73% absolute gains to the individually trained counterparts. Intriguingly, when scaling up the size of teacher models or their pre-training datasets, our method showcases the desired scalable properties, bringing increasingly larger gains to student models. The student backbones trained by our method transfer well on downstream MS-COCO and ADE20K datasets. More importantly, our method could be used as a more efficient alternative to the time-intensive pre-training paradigm for any target student model if a strong pre-trained ViT is available, reducing the amount of viewed training samples up to 195x.
Deep learning (DL) models are popular across various domains due to their remarkable performance and efficiency. However, their effectiveness relies heavily on large amounts of labeled data, which are often time-consuming and labor-intensive to generate manually. To overcome this challenge, it is essential to develop strategies that reduce reliance on extensive labeled data while preserving model performance. In this paper, we propose FisherMask, a Fisher information-based active learning (AL) approach that identifies key network parameters by masking them based on their Fisher information values. FisherMask enhances batch AL by using Fisher information to select the most critical parameters, allowing the identification of the most impactful samples during AL training. Moreover, Fisher information possesses favorable statistical properties, offering valuable insights into model behavior and providing a better understanding of the performance characteristics within the AL pipeline. Our extensive experiments demonstrate that FisherMask significantly outperforms state-of-the-art methods on diverse datasets, including CIFAR-10 and FashionMNIST, especially under imbalanced settings. These improvements lead to substantial gains in labeling efficiency. Hence serving as an effective tool to measure the sensitivity of model parameters to data samples. Our code is available on \url{https://github.com/sgchr273/FisherMask}.
Convolutional Neural Networks (CNNs) have seen significant performance improvements in recent years. However, due to their size and complexity, they function as black-boxes, leading to transparency concerns. State-of-the-art saliency methods generate local explanations that highlight the area in the input image where a class is identified but cannot explain how a concept of interest contributes to the prediction, which is essential for bias mitigation. On the other hand, concept-based methods, such as TCAV (Testing with Concept Activation Vectors), provide insights into how sensitive is the network to a concept, but cannot compute its attribution in a specific prediction nor show its location within the input image. This paper introduces a novel post-hoc explainability framework, Visual-TCAV, which aims to bridge the gap between these methods by providing both local and global explanations for CNN-based image classification. Visual-TCAV uses Concept Activation Vectors (CAVs) to generate saliency maps that show where concepts are recognized by the network. Moreover, it can estimate the attribution of these concepts to the output of any class using a generalization of Integrated Gradients. This framework is evaluated on popular CNN architectures, with its validity further confirmed via experiments where ground truth for explanations is known, and a comparison with TCAV. Our code will be made available soon.
The Platonic Representation Hypothesis claims that recent foundation models are converging to a shared representation space as a function of their downstream task performance, irrespective of the objectives and data modalities used to train these models. Representational similarity is generally measured for individual datasets and is not necessarily consistent across datasets. Thus, one may wonder whether this convergence of model representations is confounded by the datasets commonly used in machine learning. Here, we propose a systematic way to measure how representational similarity between models varies with the set of stimuli used to construct the representations. We find that the objective function is the most crucial factor in determining the consistency of representational similarities across datasets. Specifically, self-supervised vision models learn representations whose relative pairwise similarities generalize better from one dataset to another compared to those of image classification or image-text models. Moreover, the correspondence between representational similarities and the models' task behavior is dataset-dependent, being most strongly pronounced for single-domain datasets. Our work provides a framework for systematically measuring similarities of model representations across datasets and linking those similarities to differences in task behavior.
Spiking Neural Networks (SNNs) are biologically-inspired deep neural networks that efficiently extract temporal information while offering promising gains in terms of energy efficiency and latency when deployed on neuromorphic devices. However, SNN model parameters are sensitive to temporal resolution, leading to significant performance drops when the temporal resolution of target data at the edge is not the same with that of the pre-deployment source data used for training, especially when fine-tuning is not possible at the edge. To address this challenge, we propose three novel domain adaptation methods for adapting neuron parameters to account for the change in time resolution without re-training on target time-resolution. The proposed methods are based on a mapping between neuron dynamics in SNNs and State Space Models (SSMs); and are applicable to general neuron models. We evaluate the proposed methods under spatio-temporal data tasks, namely the audio keyword spotting datasets SHD and MSWC as well as the image classification NMINST dataset. Our methods provide an alternative to - and in majority of the cases significantly outperform - the existing reference method that simply scales the time constant. Moreover, our results show that high accuracy on high temporal resolution data can be obtained by time efficient training on lower temporal resolution data and model adaptation.
It has been observed that the input space of deep neural network classifiers can exhibit `fragmentation', where the model function rapidly changes class as the input space is traversed. The severity of this fragmentation tends to follow the double descent curve, achieving a maximum at the interpolation regime. We study this phenomenon in the context of image classification and ask whether fragmentation could be predictive of generalization performance. Using a fragmentation-based complexity measure, we show this to be possible by achieving good performance on the PGDL (Predicting Generalization in Deep Learning) benchmark. In addition, we report on new observations related to fragmentation, namely (i) fragmentation is not limited to the input space but occurs in the hidden representations as well, (ii) fragmentation follows the trends in the validation error throughout training, and (iii) fragmentation is not a direct result of increased weight norms. Together, this indicates that fragmentation is a phenomenon worth investigating further when studying the generalization ability of deep neural networks.
While large language models (LLMs) have advanced the field of natural language processing (NLP), their "black box" nature obscures their decision-making processes. To address this, researchers developed structured approaches using higher order tensors. These are able to model linguistic relations, but stall when training on classical computers due to their excessive size. Tensors are natural inhabitants of quantum systems and training on quantum computers provides a solution by translating text to variational quantum circuits. In this paper, we develop MultiQ-NLP: a framework for structure-aware data processing with multimodal text+image data. Here, "structure" refers to syntactic and grammatical relationships in language, as well as the hierarchical organization of visual elements in images. We enrich the translation with new types and type homomorphisms and develop novel architectures to represent structure. When tested on a main stream image classification task (SVO Probes), our best model showed a par performance with the state of the art classical models; moreover the best model was fully structured.
Deep learning models for image classification have become standard tools in recent years. A well known vulnerability of these models is their susceptibility to adversarial examples. These are generated by slightly altering an image of a certain class in a way that is imperceptible to humans but causes the model to classify it wrongly as another class. Many algorithms have been proposed to address this problem, falling generally into one of two categories: (i) building robust classifiers (ii) directly detecting attacked images. Despite the good performance of these detectors, we argue that in a white-box setting, where the attacker knows the configuration and weights of the network and the detector, they can overcome the detector by running many examples on a local copy, and sending only those that were not detected to the actual model. This problem is common in security applications where even a very good model is not sufficient to ensure safety. In this paper we propose to overcome this inherent limitation of any static defence with randomization. To do so, one must generate a very large family of detectors with consistent performance, and select one or more of them randomly for each input. For the individual detectors, we suggest the method of neural fingerprints. In the training phase, for each class we repeatedly sample a tiny random subset of neurons from certain layers of the network, and if their average is sufficiently different between clean and attacked images of the focal class they are considered a fingerprint and added to the detector bank. During test time, we sample fingerprints from the bank associated with the label predicted by the model, and detect attacks using a likelihood ratio test. We evaluate our detectors on ImageNet with different attack methods and model architectures, and show near-perfect detection with low rates of false detection.
Despite recent research advancements in adversarial attack methods, current approaches against XAI monitors are still discoverable and slower. In this paper, we present an adaptive framework for attention mask generation to enable stealthy, explainable and efficient PGD image classification adversarial attack under XAI monitors. Specifically, we utilize mutation XAI mixture and multitask self-supervised X-UNet for attention mask generation to guide PGD attack. Experiments on MNIST (MLP), CIFAR-10 (AlexNet) have shown that our system can outperform benchmark PGD, Sparsefool and SOTA SINIFGSM in balancing among stealth, efficiency and explainability which is crucial for effectively fooling SOTA defense protected classifiers.
Our proposed framework attempts to break the trade-off between performance and explainability by introducing an explainable-by-design convolutional neural network (CNN) based on the lateral inhibition mechanism. The ExplaiNet model consists of the predictor, that is a high-accuracy CNN with residual or dense skip connections, and the explainer probabilistic graph that expresses the spatial interactions of the network neurons. The value on each graph node is a local discrete feature (LDF) vector, a patch descriptor that represents the indices of antagonistic neurons ordered by the strength of their activations, which are learned with gradient descent. Using LDFs as sequences we can increase the conciseness of explanations by repurposing EXTREME, an EM-based sequence motif discovery method that is typically used in molecular biology. Having a discrete feature motif matrix for each one of intermediate image representations, instead of a continuous activation tensor, allows us to leverage the inherent explainability of Bayesian networks. By collecting observations and directly calculating probabilities, we can explain causal relationships between motifs of adjacent levels and attribute the model's output to global motifs. Moreover, experiments on various tiny image benchmark datasets confirm that our predictor ensures the same level of performance as the baseline architecture for a given count of parameters and/or layers. Our novel method shows promise to exceed this performance while providing an additional stream of explanations. In the solved MNIST classification task, it reaches a comparable to the state-of-the-art performance for single models, using standard training setup and 0.75 million parameters.
Machine learning models are prone to adversarial attacks, where inputs can be manipulated in order to cause misclassifications. While previous research has focused on techniques like Generative Adversarial Networks (GANs), there's limited exploration of GANs and Synthetic Minority Oversampling Technique (SMOTE) in text and image classification models to perform adversarial attacks. Our study addresses this gap by training various machine learning models and using GANs and SMOTE to generate additional data points aimed at attacking text classification models. Furthermore, we extend our investigation to face recognition models, training a Convolutional Neural Network(CNN) and subjecting it to adversarial attacks with fast gradient sign perturbations on key features identified by GradCAM, a technique used to highlight key image characteristics CNNs use in classification. Our experiments reveal a significant vulnerability in classification models. Specifically, we observe a 20 % decrease in accuracy for the top-performing text classification models post-attack, along with a 30 % decrease in facial recognition accuracy. This highlights the susceptibility of these models to manipulation of input data. Adversarial attacks not only compromise the security but also undermine the reliability of machine learning systems. By showcasing the impact of adversarial attacks on both text classification and face recognition models, our study underscores the urgent need for develop robust defenses against such vulnerabilities.
Data augmentation has become a pivotal tool in enhancing the performance of computer vision tasks, with the KeepOriginalAugment method emerging as a standout technique for its intelligent incorporation of salient regions within less prominent areas, enabling augmentation in both regions. Despite its success in image classification, its potential in addressing biases remains unexplored. In this study, we introduce an extension of the KeepOriginalAugment method, termed FaceKeepOriginalAugment, which explores various debiasing aspects-geographical, gender, and stereotypical biases-in computer vision models. By maintaining a delicate balance between data diversity and information preservation, our approach empowers models to exploit both diverse salient and non-salient regions, thereby fostering increased diversity and debiasing effects. We investigate multiple strategies for determining the placement of the salient region and swapping perspectives to decide which part undergoes augmentation. Leveraging the Image Similarity Score (ISS), we quantify dataset diversity across a range of datasets, including Flickr Faces HQ (FFHQ), WIKI, IMDB, Labelled Faces in the Wild (LFW), UTK Faces, and Diverse Dataset. We evaluate the effectiveness of FaceKeepOriginalAugment in mitigating gender bias across CEO, Engineer, Nurse, and School Teacher datasets, utilizing the Image-Image Association Score (IIAS) in convolutional neural networks (CNNs) and vision transformers (ViTs). Our findings shows the efficacy of FaceKeepOriginalAugment in promoting fairness and inclusivity within computer vision models, demonstrated by reduced gender bias and enhanced overall fairness. Additionally, we introduce a novel metric, Saliency-Based Diversity and Fairness Metric, which quantifies both diversity and fairness while handling data imbalance across various datasets.
Recent studies have shown that deep learning models are very vulnerable to poisoning attacks. Many defense methods have been proposed to address this issue. However, traditional poisoning attacks are not as threatening as commonly believed. This is because they often cause differences in how the model performs on the training set compared to the validation set. Such inconsistency can alert defenders that their data has been poisoned, allowing them to take the necessary defensive actions. In this paper, we introduce a more threatening type of poisoning attack called the Deferred Poisoning Attack. This new attack allows the model to function normally during the training and validation phases but makes it very sensitive to evasion attacks or even natural noise. We achieve this by ensuring the poisoned model's loss function has a similar value as a normally trained model at each input sample but with a large local curvature. A similar model loss ensures that there is no obvious inconsistency between the training and validation accuracy, demonstrating high stealthiness. On the other hand, the large curvature implies that a small perturbation may cause a significant increase in model loss, leading to substantial performance degradation, which reflects a worse robustness. We fulfill this purpose by making the model have singular Hessian information at the optimal point via our proposed Singularization Regularization term. We have conducted both theoretical and empirical analyses of the proposed method and validated its effectiveness through experiments on image classification tasks. Furthermore, we have confirmed the hazards of this form of poisoning attack under more general scenarios using natural noise, offering a new perspective for research in the field of security.
Fine-tuned vision-language models (VLMs) often capture spurious correlations between image features and textual attributes, resulting in degraded zero-shot performance at test time. Existing approaches for addressing spurious correlations (i) primarily operate at the global image-level rather than intervening directly on fine-grained image features and (ii) are predominantly designed for unimodal settings. In this work, we present RaVL, which takes a fine-grained perspective on VLM robustness by discovering and mitigating spurious correlations using local image features rather than operating at the global image level. Given a fine-tuned VLM, RaVL first discovers spurious correlations by leveraging a region-level clustering approach to identify precise image features contributing to zero-shot classification errors. Then, RaVL mitigates the identified spurious correlation with a novel region-aware loss function that enables the VLM to focus on relevant regions and ignore spurious relationships during fine-tuning. We evaluate RaVL on 654 VLMs with various model architectures, data domains, and learned spurious correlations. Our results show that RaVL accurately discovers (191% improvement over the closest baseline) and mitigates (8.2% improvement on worst-group image classification accuracy) spurious correlations. Qualitative evaluations on general-domain and medical-domain VLMs confirm our findings.
Federated learning enables multiple actors to collaboratively train models without sharing private data. This unlocks the potential for scaling machine learning to diverse applications. Existing algorithms for this task are well-justified when clients and the intended target domain share the same distribution of features and labels, but this assumption is often violated in real-world scenarios. One common violation is label shift, where the label distributions differ across clients or between clients and the target domain, which can significantly degrade model performance. To address this problem, we propose FedPALS, a novel model aggregation scheme that adapts to label shifts by leveraging knowledge of the target label distribution at the central server. Our approach ensures unbiased updates under stochastic gradient descent, ensuring robust generalization across clients with diverse, label-shifted data. Extensive experiments on image classification demonstrate that FedPALS consistently outperforms standard baselines by aligning model aggregation with the target domain. Our findings reveal that conventional federated learning methods suffer severely in cases of extreme client sparsity, highlighting the critical need for target-aware aggregation. FedPALS offers a principled and practical solution to mitigate label distribution mismatch, ensuring models trained in federated settings can generalize effectively to label-shifted target domains.
Adam is one of the most popular optimization algorithms in deep learning. However, it is known that Adam does not converge in theory unless choosing a hyperparameter, i.e., $\beta_2$, in a problem-dependent manner. There have been many attempts to fix the non-convergence (e.g., AMSGrad), but they require an impractical assumption that the gradient noise is uniformly bounded. In this paper, we propose a new adaptive gradient method named ADOPT, which achieves the optimal convergence rate of $\mathcal{O} ( 1 / \sqrt{T} )$ with any choice of $\beta_2$ without depending on the bounded noise assumption. ADOPT addresses the non-convergence issue of Adam by removing the current gradient from the second moment estimate and changing the order of the momentum update and the normalization by the second moment estimate. We also conduct intensive numerical experiments, and verify that our ADOPT achieves superior results compared to Adam and its variants across a wide range of tasks, including image classification, generative modeling, natural language processing, and deep reinforcement learning. The implementation is available at https://github.com/iShohei220/adopt.
In order to improve model accuracy, generalization, and class imbalance issues, this work offers a strong methodology for classifying endoscopic images. We suggest a hybrid feature extraction method that combines convolutional neural networks (CNNs), multi-layer perceptrons (MLPs), and radiomics. Rich, multi-scale feature extraction is made possible by this combination, which captures both deep and handmade representations. These features are then used by a classification head to classify diseases, producing a model with higher generalization and accuracy. In this framework we have achieved a validation accuracy of 76.2% in the capsule endoscopy video frame classification task.
Contrastive Language-Image Pretraining (CLIP) models maximize the mutual information between text and visual modalities to learn representations. This makes the nature of the training data a significant factor in the efficacy of CLIP for downstream tasks. However, the lack of compositional diversity in contemporary image-text datasets limits the compositional reasoning ability of CLIP. We show that generating ``hard'' negative captions via in-context learning and synthesizing corresponding negative images with text-to-image generators offers a solution. We introduce a novel contrastive pre-training strategy that leverages these hard negative captions and images in an alternating fashion to train CLIP. We demonstrate that our method, named TripletCLIP, when applied to existing datasets such as CC3M and CC12M, enhances the compositional capabilities of CLIP, resulting in an absolute improvement of over 9% on the SugarCrepe benchmark on an equal computational budget, as well as improvements in zero-shot image classification and image retrieval. Our code, models, and data are available at: https://tripletclip.github.io
Semi-supervised learning (SSL) is a popular solution to alleviate the high annotation cost in medical image classification. As a main branch of SSL, consistency regularization engages in imposing consensus between the predictions of a single sample from different views, termed as Absolute Location consistency (AL-c). However, only AL-c may be insufficient. Just like when diagnosing a case in practice, besides the case itself, the doctor usually refers to certain related trustworthy cases to make more reliable decisions.Therefore, we argue that solely relying on AL-c may ignore the relative differences across samples, which we interpret as relative locations, and only exploit limited information from one perspective. To address this issue, we propose a Sample Consistency Mean Teacher (SCMT) which not only incorporates AL c but also additionally enforces consistency between the samples' relative similarities to its related samples, called Relative Location consistency (RL c). AL c and RL c conduct consistency regularization from two different perspectives, jointly extracting more diverse semantic information for classification. On the other hand, due to the highly similar structures in medical images, the sample distribution could be overly dense in feature space, making their relative locations susceptible to noise. To tackle this problem, we further develop a Sample Scatter Mean Teacher (SSMT) by utilizing contrastive learning to sparsify the sample distribution and obtain robust and effective relative locations. Extensive experiments on different datasets demonstrate the superiority of our method.
Open-set single-source domain generalization aims to use a single-source domain to learn a robust model that can be generalized to unknown target domains with both domain shifts and label shifts. The scarcity of the source domain and the unknown data distribution of the target domain pose a great challenge for domain-invariant feature learning and unknown class recognition. In this paper, we propose a novel learning approach based on domain expansion and boundary growth to expand the scarce source samples and enlarge the boundaries across the known classes that indirectly broaden the boundary between the known and unknown classes. Specifically, we achieve domain expansion by employing both background suppression and style augmentation on the source data to synthesize new samples. Then we force the model to distill consistent knowledge from the synthesized samples so that the model can learn domain-invariant information. Furthermore, we realize boundary growth across classes by using edge maps as an additional modality of samples when training multi-binary classifiers. In this way, it enlarges the boundary between the inliers and outliers, and consequently improves the unknown class recognition during open-set generalization. Extensive experiments show that our approach can achieve significant improvements and reach state-of-the-art performance on several cross-domain image classification datasets.