Image classification is one of the most fundamental capabilities of machine vision intelligence. In this work, we revisit the image classification task using visually-grounded language models (VLMs) such as GPT-4V and LLaVA. We find that existing proprietary and public VLMs, despite often using CLIP as a vision encoder and having many more parameters, significantly underperform CLIP on standard image classification benchmarks like ImageNet. To understand the reason, we explore several hypotheses concerning the inference algorithms, training objectives, and data processing in VLMs. Our analysis reveals that the primary cause is data-related: critical information for image classification is encoded in the VLM's latent space but can only be effectively decoded with enough training data. Specifically, there is a strong correlation between the frequency of class exposure during VLM training and instruction-tuning and the VLM's performance in those classes; when trained with sufficient data, VLMs can match the accuracy of state-of-the-art classification models. Based on these findings, we enhance a VLM by integrating classification-focused datasets into its training, and demonstrate that the enhanced classification performance of the VLM transfers to its general capabilities, resulting in an improvement of 11.8% on the newly collected ImageWikiQA dataset.
Second-order optimizers, maintaining a matrix termed a preconditioner, are superior to first-order optimizers in both theory and practice. The states forming the preconditioner and its inverse root restrict the maximum size of models trained by second-order optimizers. To address this, compressing 32-bit optimizer states to lower bitwidths has shown promise in reducing memory usage. However, current approaches only pertain to first-order optimizers. In this paper, we propose the first 4-bit second-order optimizers, exemplified by 4-bit Shampoo, maintaining performance similar to that of 32-bit ones. We show that quantizing the eigenvector matrix of the preconditioner in 4-bit Shampoo is remarkably better than quantizing the preconditioner itself both theoretically and experimentally. By rectifying the orthogonality of the quantized eigenvector matrix, we enhance the approximation of the preconditioner's eigenvector matrix, which also benefits the computation of its inverse 4-th root. Besides, we find that linear square quantization slightly outperforms dynamic tree quantization when quantizing second-order optimizer states. Evaluation on various networks for image classification and natural language modeling demonstrates that our 4-bit Shampoo achieves comparable performance to its 32-bit counterpart while being more memory-efficient.
The joint-embedding predictive architecture (JEPA) recently has shown impressive results in extracting visual representations from unlabeled imagery under a masking strategy. However, we reveal its disadvantages, notably its insufficient understanding of local semantics. This deficiency originates from masked modeling in the embedding space, resulting in a reduction of discriminative power and can even lead to the neglect of critical local semantics. To bridge this gap, we introduce DMT-JEPA, a novel masked modeling objective rooted in JEPA, specifically designed to generate discriminative latent targets from neighboring information. Our key idea is simple: we consider a set of semantically similar neighboring patches as a target of a masked patch. To be specific, the proposed DMT-JEPA (a) computes feature similarities between each masked patch and its corresponding neighboring patches to select patches having semantically meaningful relations, and (b) employs lightweight cross-attention heads to aggregate features of neighboring patches as the masked targets. Consequently, DMT-JEPA demonstrates strong discriminative power, offering benefits across a diverse spectrum of downstream tasks. Through extensive experiments, we demonstrate our effectiveness across various visual benchmarks, including ImageNet-1K image classification, ADE20K semantic segmentation, and COCO object detection tasks. Code is available at: \url{https://github.com/DMTJEPA/DMTJEPA}.
This paper presents a novel model protection paradigm ModelLock that locks (destroys) the performance of a model on normal clean data so as to make it unusable or unextractable without the right key. Specifically, we proposed a diffusion-based framework dubbed ModelLock that explores text-guided image editing to transform the training data into unique styles or add new objects in the background. A model finetuned on this edited dataset will be locked and can only be unlocked by the key prompt, i.e., the text prompt used to transform the data. We conduct extensive experiments on both image classification and segmentation tasks, and show that 1) ModelLock can effectively lock the finetuned models without significantly reducing the expected performance, and more importantly, 2) the locked model cannot be easily unlocked without knowing both the key prompt and the diffusion model. Our work opens up a new direction for intellectual property protection of private models.
First-order optimization methods are currently the mainstream in training deep neural networks (DNNs). Optimizers like Adam incorporate limited curvature information by employing the diagonal matrix preconditioning of the stochastic gradient during the training. Despite their widespread, second-order optimization algorithms exhibit superior convergence properties compared to their first-order counterparts e.g. Adam and SGD. However, their practicality in training DNNs is still limited due to increased per-iteration computations compared to the first-order methods. We present \emph{AdaFisher}--an adaptive second-order optimizer that leverages a \emph{diagonal block-Kronecker} approximation of the Fisher information matrix for adaptive gradient preconditioning. AdaFisher aims to bridge the gap between enhanced \emph{convergence/generalization} capabilities and computational efficiency in second-order optimization framework for training DNNs. Despite the slow pace of second-order optimizers, we showcase that AdaFisher can be reliably adopted for image classification, language modeling and stands out for its stability and robustness in hyper-parameter tuning. We demonstrate that AdaFisher \textbf{outperforms the SOTA optimizers} in terms of both accuracy and convergence speed. Code is available from https://github.com/AtlasAnalyticsLab/AdaFisher.
Meta-learning has been proposed as a promising machine learning topic in recent years, with important applications to image classification, robotics, computer games, and control systems. In this paper, we study the problem of using meta-learning to deal with uncertainty and heterogeneity in ergodic linear quadratic regulators. We integrate the zeroth-order optimization technique with a typical meta-learning method, proposing an algorithm that omits the estimation of policy Hessian, which applies to tasks of learning a set of heterogeneous but similar linear dynamic systems. The induced meta-objective function inherits important properties of the original cost function when the set of linear dynamic systems are meta-learnable, allowing the algorithm to optimize over a learnable landscape without projection onto the feasible set. We provide a convergence result for the exact gradient descent process by analyzing the boundedness and smoothness of the gradient for the meta-objective, which justify the proposed algorithm with gradient estimation error being small. We also provide a numerical example to corroborate this perspective.
In-context learning based on attention models is examined for data with categorical outcomes, with inference in such models viewed from the perspective of functional gradient descent (GD). We develop a network composed of attention blocks, with each block employing a self-attention layer followed by a cross-attention layer, with associated skip connections. This model can exactly perform multi-step functional GD inference for in-context inference with categorical observations. We perform a theoretical analysis of this setup, generalizing many prior assumptions in this line of work, including the class of attention mechanisms for which it is appropriate. We demonstrate the framework empirically on synthetic data, image classification and language generation.
Deep neural networks face persistent challenges in defending against backdoor attacks, leading to an ongoing battle between attacks and defenses. While existing backdoor defense strategies have shown promising performance on reducing attack success rates, can we confidently claim that the backdoor threat has truly been eliminated from the model? To address it, we re-investigate the characteristics of the backdoored models after defense (denoted as defense models). Surprisingly, we find that the original backdoors still exist in defense models derived from existing post-training defense strategies, and the backdoor existence is measured by a novel metric called backdoor existence coefficient. It implies that the backdoors just lie dormant rather than being eliminated. To further verify this finding, we empirically show that these dormant backdoors can be easily re-activated during inference, by manipulating the original trigger with well-designed tiny perturbation using universal adversarial attack. More practically, we extend our backdoor reactivation to black-box scenario, where the defense model can only be queried by the adversary during inference, and develop two effective methods, i.e., query-based and transfer-based backdoor re-activation attacks. The effectiveness of the proposed methods are verified on both image classification and multimodal contrastive learning (i.e., CLIP) tasks. In conclusion, this work uncovers a critical vulnerability that has never been explored in existing defense strategies, emphasizing the urgency of designing more robust and advanced backdoor defense mechanisms in the future.
Domain Generalization (DG) is a challenging task in machine learning that requires a coherent ability to comprehend shifts across various domains through extraction of domain-invariant features. DG performance is typically evaluated by performing image classification in domains of various image styles. However, current methodology lacks quantitative understanding about shifts in stylistic domain, and relies on a vast amount of pre-training data, such as ImageNet1K, which are predominantly in photo-realistic style with weakly supervised class labels. Such a data-driven practice could potentially result in spurious correlation and inflated performance on DG benchmarks. In this paper, we introduce a new DG paradigm to address these risks. We first introduce two new quantitative measures ICV and IDD to describe domain shifts in terms of consistency of classes within one domain and similarity between two stylistic domains. We then present SuperMarioDomains (SMD), a novel synthetic multi-domain dataset sampled from video game scenes with more consistent classes and sufficient dissimilarity compared to ImageNet1K. We demonstrate our DG method SMOS. SMOS first uses SMD to train a precursor model, which is then used to ground the training on a DG benchmark. We observe that SMOS contributes to state-of-the-art performance across five DG benchmarks, gaining large improvements to performances on abstract domains along with on-par or slight improvements to those on photo-realistic domains. Our qualitative analysis suggests that these improvements can be attributed to reduced distributional divergence between originally distant domains. Our data are available at https://github.com/fpsluozi/SMD-SMOS .
Recent efforts in interpreting Convolutional Neural Networks (CNNs) focus on translating the activation of CNN filters into a stratified Answer Set Program (ASP) rule-sets. The CNN filters are known to capture high-level image concepts, thus the predicates in the rule-set are mapped to the concept that their corresponding filter represents. Hence, the rule-set exemplifies the decision-making process of the CNN w.r.t the concepts that it learns for any image classification task. These rule-sets help understand the biases in CNNs, although correcting the biases remains a challenge. We introduce a neurosymbolic framework called NeSyBiCor for bias correction in a trained CNN. Given symbolic concepts, as ASP constraints, that the CNN is biased towards, we convert the concepts to their corresponding vector representations. Then, the CNN is retrained using our novel semantic similarity loss that pushes the filters away from (or towards) learning the desired/undesired concepts. The final ASP rule-set obtained after retraining, satisfies the constraints to a high degree, thus showing the revision in the knowledge of the CNN. We demonstrate that our NeSyBiCor framework successfully corrects the biases of CNNs trained with subsets of classes from the "Places" dataset while sacrificing minimal accuracy and improving interpretability.
Training deep neural networks (DNNs) using traditional backpropagation (BP) presents challenges in terms of computational complexity and energy consumption, particularly for on-device learning where computational resources are limited. Various alternatives to BP, including random feedback alignment, forward-forward, and local classifiers, have been explored to address these challenges. These methods have their advantages, but they can encounter difficulties when dealing with intricate visual tasks or demand considerable computational resources. In this paper, we propose a novel Local Learning rule inspired by neural activity Synchronization phenomena (LLS) observed in the brain. LLS utilizes fixed periodic basis vectors to synchronize neuron activity within each layer, enabling efficient training without the need for additional trainable parameters. We demonstrate the effectiveness of LLS and its variations, LLS-M and LLS-MxM, on multiple image classification datasets, achieving accuracy comparable to BP with reduced computational complexity and minimal additional parameters. Specifically, LLS achieves comparable performance with up to $300 \times$ fewer multiply-accumulate (MAC) operations and half the memory requirements of BP. Furthermore, the performance of LLS on the Visual Wake Word (VWW) dataset highlights its suitability for on-device learning tasks, making it a promising candidate for edge hardware implementations.
Multi-label image classification datasets are often partially labeled where many labels are missing, posing a significant challenge to training accurate deep classifiers. However, the powerful Mixup sample-mixing data augmentation cannot be well utilized to address this challenge, as it cannot perform linear interpolation on the unknown labels to construct augmented samples. In this paper, we propose LogicMix, a Mixup variant designed for such partially labeled datasets. LogicMix mixes the sample labels by logical OR so that the unknown labels can be correctly mixed by utilizing OR's logical equivalences, including the domination and identity laws. Unlike Mixup, which mixes exactly two samples, LogicMix can mix multiple ($\geq2$) partially labeled samples, constructing visually more confused augmented samples to regularize training. LogicMix is more general and effective than other compared Mixup variants in the experiments on various partially labeled dataset scenarios. Moreover, it is plug-and-play and only requires minimal computation, hence it can be easily inserted into existing frameworks to collaborate with other methods to improve model performance with a negligible impact on training time, as demonstrated through extensive experiments. In particular, through the collaboration of LogicMix, RandAugment, Curriculum Labeling, and Category-wise Fine-Tuning, we attain state-of-the-art performance on MS-COCO, VG-200, and Pascal VOC 2007 benchmarking datasets. The remarkable generality, effectiveness, collaboration, and simplicity suggest that LogicMix promises to be a popular and vital data augmentation method.
Insufficient prior knowledge of a captured hyperspectral image (HSI) scene may lead the experts or the automatic labeling systems to offer incorrect labels or ambiguous labels (i.e., assigning each training sample to a group of candidate labels, among which only one of them is valid; this is also known as partial label learning) during the labeling process. Accordingly, how to learn from such data with ambiguous labels is a problem of great practical importance. In this paper, we propose a novel superpixelwise low-rank approximation (LRA)-based partial label learning method, namely SLAP, which is the first to take into account partial label learning in HSI classification. SLAP is mainly composed of two phases: disambiguating the training labels and acquiring the predictive model. Specifically, in the first phase, we propose a superpixelwise LRA-based model, preparing the affinity graph for the subsequent label propagation process while extracting the discriminative representation to enhance the following classification task of the second phase. Then to disambiguate the training labels, label propagation propagates the labeling information via the affinity graph of training pixels. In the second phase, we take advantage of the resulting disambiguated training labels and the discriminative representations to enhance the classification performance. The extensive experiments validate the advantage of the proposed SLAP method over state-of-the-art methods.
Mamba is an effective state space model with linear computation complexity. It has recently shown impressive efficiency in dealing with high-resolution inputs across various vision tasks. In this paper, we reveal that the powerful Mamba model shares surprising similarities with linear attention Transformer, which typically underperform conventional Transformer in practice. By exploring the similarities and disparities between the effective Mamba and subpar linear attention Transformer, we provide comprehensive analyses to demystify the key factors behind Mamba's success. Specifically, we reformulate the selective state space model and linear attention within a unified formulation, rephrasing Mamba as a variant of linear attention Transformer with six major distinctions: input gate, forget gate, shortcut, no attention normalization, single-head, and modified block design. For each design, we meticulously analyze its pros and cons, and empirically evaluate its impact on model performance in vision tasks. Interestingly, the results highlight the forget gate and block design as the core contributors to Mamba's success, while the other four designs are less crucial. Based on these findings, we propose a Mamba-Inspired Linear Attention (MILA) model by incorporating the merits of these two key designs into linear attention. The resulting model outperforms various vision Mamba models in both image classification and high-resolution dense prediction tasks, while enjoying parallelizable computation and fast inference speed. Code is available at https://github.com/LeapLabTHU/MLLA.
In current AI era, users may request AI companies to delete their data from the training dataset due to the privacy concerns. As a model owner, retraining a model will consume significant computational resources. Therefore, machine unlearning is a new emerged technology to allow model owner to delete requested training data or a class with little affecting on the model performance. However, for large-scaling complex data, such as image or text data, unlearning a class from a model leads to a inferior performance due to the difficulty to identify the link between classes and model. An inaccurate class deleting may lead to over or under unlearning. In this paper, to accurately defining the unlearning class of complex data, we apply the definition of Concept, rather than an image feature or a token of text data, to represent the semantic information of unlearning class. This new representation can cut the link between the model and the class, leading to a complete erasing of the impact of a class. To analyze the impact of the concept of complex data, we adopt a Post-hoc Concept Bottleneck Model, and Integrated Gradients to precisely identify concepts across different classes. Next, we take advantage of data poisoning with random and targeted labels to propose unlearning methods. We test our methods on both image classification models and large language models (LLMs). The results consistently show that the proposed methods can accurately erase targeted information from models and can largely maintain the performance of the models.
Federated learning (FL) is a machine learning approach where nodes collaboratively train a global model. As more nodes participate in a round of FL, the effectiveness of individual model updates by nodes also diminishes. In this study, we increase the effectiveness of client updates by dividing the network into smaller partitions, or cohorts. We introduce Cohort-Parallel Federated Learning (CPFL): a novel learning approach where each cohort independently trains a global model using FL, until convergence, and the produced models by each cohort are then unified using knowledge distillation. The insight behind CPFL is that smaller, isolated networks converge quicker than in a one-network setting where all nodes participate. Through exhaustive experiments involving realistic traces and non-IID data distributions on the CIFAR-10 and FEMNIST image classification tasks, we investigate the balance between the number of cohorts, model accuracy, training time, and compute resources. Compared to traditional FL, CPFL with four cohorts, non-IID data distribution, and CIFAR-10 yields a 1.9x reduction in train time and a 1.3x reduction in resource usage, with a minimal drop in test accuracy.
Online Lifelong Learning (OLL) addresses the challenge of learning from continuous and non-stationary data streams. Existing online lifelong learning methods based on image classification models often require preset conditions such as the total number of classes or maximum memory capacity, which hinders the realization of real never-ending learning and renders them impractical for real-world scenarios. In this work, we propose that vision-language models, such as Contrastive Language-Image Pretraining (CLIP), are more suitable candidates for online lifelong learning. We discover that maintaining symmetry between image and text is crucial during Parameter-Efficient Tuning (PET) for CLIP model in online lifelong learning. To this end, we introduce the Symmetric Image-Text (SIT) tuning strategy. We conduct extensive experiments on multiple lifelong learning benchmark datasets and elucidate the effectiveness of SIT through gradient analysis. Additionally, we assess the impact of lifelong learning on generalizability of CLIP and found that tuning the image encoder is beneficial for lifelong learning, while tuning the text encoder aids in zero-shot learning.
In deep learning, maintaining model robustness against distribution shifts is critical. This work explores a broad range of possibilities to adapt vision-language foundation models at test-time, with a particular emphasis on CLIP and its variants. The study systematically examines prompt-based techniques and existing test-time adaptation methods, aiming to improve the robustness under distribution shift in diverse real-world scenarios. Specifically, the investigation covers various prompt engineering strategies, including handcrafted prompts, prompt ensembles, and prompt learning techniques. Additionally, we introduce a vision-text-space ensemble that substantially enhances average performance compared to text-space-only ensembles. Since online test-time adaptation has shown to be effective to mitigate performance drops under distribution shift, the study extends its scope to evaluate the effectiveness of existing test-time adaptation methods that were originally designed for vision-only classification models. Through extensive experimental evaluations conducted across multiple datasets and diverse model architectures, the research demonstrates the effectiveness of these adaptation strategies. Code is available at: https://github.com/mariodoebler/test-time-adaptation
Large language models (LLMs) have been effectively used for many computer vision tasks, including image classification. In this paper, we present a simple yet effective approach for zero-shot image classification using multimodal LLMs. Using multimodal LLMs, we generate comprehensive textual representations from input images. These textual representations are then utilized to generate fixed-dimensional features in a cross-modal embedding space. Subsequently, these features are fused together to perform zero-shot classification using a linear classifier. Our method does not require prompt engineering for each dataset; instead, we use a single, straightforward set of prompts across all datasets. We evaluated our method on several datasets and our results demonstrate its remarkable effectiveness, surpassing benchmark accuracy on multiple datasets. On average, for ten benchmarks, our method achieved an accuracy gain of 6.2 percentage points, with an increase of 6.8 percentage points on the ImageNet dataset, compared to prior methods re-evaluated with the same setup. Our findings highlight the potential of multimodal LLMs to enhance computer vision tasks such as zero-shot image classification, offering a significant improvement over traditional methods.
Federated learning (FL) aims to collaboratively learn deep learning model parameters from decentralized data archives (i.e., clients) without accessing training data on clients. However, the training data across clients might be not independent and identically distributed (non-IID), which may result in difficulty in achieving optimal model convergence. In this work, we investigate the capability of state-of-the-art transformer architectures (which are MLP-Mixer, ConvMixer, PoolFormer) to address the challenges related to non-IID training data across various clients in the context of FL for multi-label classification (MLC) problems in remote sensing (RS). The considered transformer architectures are compared among themselves and with the ResNet-50 architecture in terms of their: 1) robustness to training data heterogeneity; 2) local training complexity; and 3) aggregation complexity under different non-IID levels. The experimental results obtained on the BigEarthNet-S2 benchmark archive demonstrate that the considered architectures increase the generalization ability with the cost of higher local training and aggregation complexities. On the basis of our analysis, some guidelines are derived for a proper selection of transformer architecture in the context of FL for RS MLC. The code of this work is publicly available at https://git.tu-berlin.de/rsim/FL-Transformer.