Despite the empirical success of Low-Rank Adaptation (LoRA) in fine-tuning pre-trained models, there is little theoretical understanding of how first-order methods with carefully crafted initialization adapt models to new tasks. In this work, we take the first step towards bridging this gap by theoretically analyzing the learning dynamics of LoRA for matrix factorization (MF) under gradient flow (GF), emphasizing the crucial role of initialization. For small initialization, we theoretically show that GF converges to a neighborhood of the optimal solution, with smaller initialization leading to lower final error. Our analysis shows that the final error is affected by the misalignment between the singular spaces of the pre-trained model and the target matrix, and reducing the initialization scale improves alignment. To address this misalignment, we propose a spectral initialization for LoRA in MF and theoretically prove that GF with small spectral initialization converges to the fine-tuning task with arbitrary precision. Numerical experiments from MF and image classification validate our findings.
Model merging enables efficient multi-task models by combining task-specific fine-tuned checkpoints. However, storing multiple task-specific checkpoints requires significant memory, limiting scalability and restricting model merging to larger models and diverse tasks. In this paper, we propose quantizing task vectors (i.e., the difference between pre-trained and fine-tuned checkpoints) instead of quantizing fine-tuned checkpoints. We observe that task vectors exhibit a narrow weight range, enabling low precision quantization (up to 4 bit) within existing task vector merging frameworks. To further mitigate quantization errors within ultra-low bit precision (e.g., 2 bit), we introduce Residual Task Vector Quantization, which decomposes the task vector into a base vector and offset component. We allocate bits based on quantization sensitivity, ensuring precision while minimizing error within a memory budget. Experiments on image classification and dense prediction show our method maintains or improves model merging performance while using only 8% of the memory required for full-precision checkpoints.
The increasing demand for continual learning in sequential data processing has led to progressively complex training methodologies and larger recurrent network architectures. Consequently, this has widened the knowledge gap between continual learning with recurrent neural networks (RNNs) and their ability to operate on devices with limited memory and compute. To address this challenge, we investigate the effectiveness of simplifying RNN architectures, particularly gated recurrent unit (GRU), and its impact on both single-task and multitask sequential learning. We propose a new variant of GRU, namely the minion recurrent unit (MiRU). MiRU replaces conventional gating mechanisms with scaling coefficients to regulate dynamic updates of hidden states and historical context, reducing computational costs and memory requirements. Despite its simplified architecture, MiRU maintains performance comparable to the standard GRU while achieving 2.90x faster training and reducing parameter usage by 2.88x, as demonstrated through evaluations on sequential image classification and natural language processing benchmarks. The impact of model simplification on its learning capacity is also investigated by performing continual learning tasks with a rehearsal-based strategy and global inhibition. We find that MiRU demonstrates stable performance in multitask learning even when using only rehearsal, unlike the standard GRU and its variants. These features position MiRU as a promising candidate for edge-device applications.
Federated learning (FL) enables decentralized clients to collaboratively train a global model under the orchestration of a central server without exposing their individual data. However, the iterative exchange of model parameters between the server and clients imposes heavy communication burdens, risks potential privacy leakage, and even precludes collaboration among heterogeneous clients. Distillation-based FL tackles these challenges by exchanging low-dimensional model outputs rather than model parameters, yet it highly relies on a task-relevant auxiliary dataset that is often not available in practice. Data-free FL attempts to overcome this limitation by training a server-side generator to directly synthesize task-specific data samples for knowledge transfer. However, the update rule of the generator requires clients to share on-device models for white-box access, which greatly compromises the advantages of distillation-based FL. This motivates us to explore a data-free and black-box FL framework via Zeroth-order Gradient Estimation (FedZGE), which estimates the gradients after flowing through on-device models in a black-box optimization manner to complete the training of the generator in terms of fidelity, transferability, diversity, and equilibrium, without involving any auxiliary data or sharing any model parameters, thus combining the advantages of both distillation-based FL and data-free FL. Experiments on large-scale image classification datasets and network architectures demonstrate the superiority of FedZGE in terms of data heterogeneity, model heterogeneity, communication efficiency, and privacy protection.
Kolmogorov-Arnold Networks (KANs) have inspired numerous works exploring their applications across a wide range of scientific problems, with the potential to replace Multilayer Perceptrons (MLPs). While many KANs are designed using basis and polynomial functions, such as B-splines, ReLU-KAN utilizes a combination of ReLU functions to mimic the structure of B-splines and take advantage of ReLU's speed. However, ReLU-KAN is not built for multiple inputs, and its limitations stem from ReLU's handling of negative values, which can restrict feature extraction. To address these issues, we introduce Activation Function-Based Kolmogorov-Arnold Networks (AF-KAN), expanding ReLU-KAN with various activations and their function combinations. This novel KAN also incorporates parameter reduction methods, primarily attention mechanisms and data normalization, to enhance performance on image classification datasets. We explore different activation functions, function combinations, grid sizes, and spline orders to validate the effectiveness of AF-KAN and determine its optimal configuration. In the experiments, AF-KAN significantly outperforms MLP, ReLU-KAN, and other KANs with the same parameter count. It also remains competitive even when using fewer than 6 to 10 times the parameters while maintaining the same network structure. However, AF-KAN requires a longer training time and consumes more FLOPs. The repository for this work is available at https://github.com/hoangthangta/All-KAN.
In recent years, the random vector functional link (RVFL) network has gained significant popularity in hyperspectral image (HSI) classification due to its simplicity, speed, and strong generalization performance. However, despite these advantages, RVFL models face several limitations, particularly in handling non-linear relationships and complex data structures. The random initialization of input-to-hidden weights can lead to instability, and the model struggles with determining the optimal number of hidden nodes, affecting its performance on more challenging datasets. To address these issues, we propose a novel randomized based restricted kernel machine ($R^2KM$) model that combines the strehyperngths of RVFL and restricted kernel machines (RKM). $R^2KM$ introduces a layered structure that represents kernel methods using both visible and hidden variables, analogous to the energy function in restricted Boltzmann machines (RBM). This structure enables $R^2KM$ to capture complex data interactions and non-linear relationships more effectively, improving both interpretability and model robustness. A key contribution of $R^2KM$ is the introduction of a novel conjugate feature duality based on the Fenchel-Young inequality, which expresses the problem in terms of conjugate dual variables and provides an upper bound on the objective function. This duality enhances the model's flexibility and scalability, offering a more efficient and flexible solution for complex data analysis tasks. Extensive experiments on hyperspectral image datasets and real-world data from the UCI and KEEL repositories show that $R^2KM$ outperforms baseline models, demonstrating its effectiveness in classification and regression tasks.
Compared to single view medical image classification, using multiple views can significantly enhance predictive accuracy as it can account for the complementarity of each view while leveraging correlations between views. Existing multi-view approaches typically employ separate convolutional or transformer branches combined with simplistic feature fusion strategies. However, these approaches inadvertently disregard essential cross-view correlations, leading to suboptimal classification performance, and suffer from challenges with limited receptive field (CNNs) or quadratic computational complexity (transformers). Inspired by state space sequence models, we propose XFMamba, a pure Mamba-based cross-fusion architecture to address the challenge of multi-view medical image classification. XFMamba introduces a novel two-stage fusion strategy, facilitating the learning of single-view features and their cross-view disparity. This mechanism captures spatially long-range dependencies in each view while enhancing seamless information transfer between views. Results on three public datasets, MURA, CheXpert and DDSM, illustrate the effectiveness of our approach across diverse multi-view medical image classification tasks, showing that it outperforms existing convolution-based and transformer-based multi-view methods. Code is available at https://github.com/XZheng0427/XFMamba.
This study investigates the classification of aerial images depicting transmission towers, forests, farmland, and mountains. To complete the classification job, features are extracted from input photos using a Convolutional Neural Network (CNN) architecture. Then, the images are classified using Softmax. To test the model, we ran it for ten epochs using a batch size of 90, the Adam optimizer, and a learning rate of 0.001. Both training and assessment are conducted using a dataset that blends self-collected pictures from Google satellite imagery with the MLRNet dataset. The comprehensive dataset comprises 10,400 images. Our study shows that transfer learning models and MobileNetV2 in particular, work well for landscape categorization. These models are good options for practical use because they strike a good mix between precision and efficiency; our approach achieves results with an overall accuracy of 87% on the built CNN model. Furthermore, we reach even higher accuracies by utilizing the pretrained VGG16 and MobileNetV2 models as a starting point for transfer learning. Specifically, VGG16 achieves an accuracy of 90% and a test loss of 0.298, while MobileNetV2 outperforms both models with an accuracy of 96% and a test loss of 0.119; the results demonstrate the effectiveness of employing transfer learning with MobileNetV2 for classifying transmission towers, forests, farmland, and mountains.
This paper investigates the impact of randomly initialized unique encoding of classification label position on the visual masked self-supervised pre-trained model when fine-tuning downstream classification tasks. Our findings indicate that different random initializations lead to significant variations in fine-tuned results, even when using the same allocation strategy for classification datasets. The accuracy gap between these results suggests that the visual masked self-supervised pre-trained model has an inherent preference for classification label positions. To investigate this, we compare it with the non-self-supervised visual pre-trained model and hypothesize that the masked self-supervised model exhibits a self-aware bias toward certain label positions. To mitigate the instability caused by random encoding, we propose a classification label position ranking algorithm, Label Ranker. It is based on 1-D dimensionality reduction of feature maps using Linear Discriminant Analysis and position-rank encoding of them by unsupervised feature clustering using the similarity property of Euclidean distance. This algorithm ensures that label position encoding align with the model’s inherent preference. Extensive ablation experiments using ImageMAE and VideoMAE models on the CIFAR-100, UCF101, and HMDB51 classification datasets validate our approach. Results demonstrate that our method effectively stabilizes classification label position encoding, improving fine-tuned performance for visual masked self-supervised models.
Sharpness-Aware Minimization (SAM) has emerged as a powerful method for improving generalization in machine learning models by minimizing the sharpness of the loss landscape. However, despite its success, several important questions regarding the convergence properties of SAM in non-convex settings are still open, including the benefits of using normalization in the update rule, the dependence of the analysis on the restrictive bounded variance assumption, and the convergence guarantees under different sampling strategies. To address these questions, in this paper, we provide a unified analysis of SAM and its unnormalized variant (USAM) under one single flexible update rule (Unified SAM), and we present convergence results of the new algorithm under a relaxed and more natural assumption on the stochastic noise. Our analysis provides convergence guarantees for SAM under different step size selections for non-convex problems and functions that satisfy the Polyak-Lojasiewicz (PL) condition (a non-convex generalization of strongly convex functions). The proposed theory holds under the arbitrary sampling paradigm, which includes importance sampling as special case, allowing us to analyze variants of SAM that were never explicitly considered in the literature. Experiments validate the theoretical findings and further demonstrate the practical effectiveness of Unified SAM in training deep neural networks for image classification tasks.
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 stability and convergence guarantees for the exact gradient descent process by analyzing the boundedness and local smoothness of the gradient for the meta-objective, which justify the proposed algorithm with gradient estimation error being small. We provide the sample complexity conditions for these theoretical guarantees, as well as a numerical example at the end to corroborate this perspective.
Deep learning scaling laws predict how performance improves with increased model and dataset size. Here we identify measurement noise in data as another performance scaling axis, governed by a distinct logarithmic law. We focus on representation learning models of biological single cell genomic data, where a dominant source of measurement noise is due to molecular undersampling. We introduce an information-theoretic metric for cellular representation model quality, and find that it scales with sampling depth. A single quantitative relationship holds across several model types and across several datasets. We show that the analytical form of this relationship can be derived from a simple Gaussian noise model, which in turn provides an intuitive interpretation for the scaling law. Finally, we show that the same relationship emerges in image classification models with respect to two types of imaging noise, suggesting that measurement noise scaling may be a general phenomenon. Scaling with noise can serve as a guide in generating and curating data for deep learning models, particularly in fields where measurement quality can vary dramatically between datasets.
In recent years, vision transformers (ViTs) have emerged as powerful and promising techniques for computer vision tasks such as image classification, object detection, and segmentation. Unlike convolutional neural networks (CNNs), which rely on hierarchical feature extraction, ViTs treat images as sequences of patches and leverage self-attention mechanisms. However, their high computational complexity and memory demands pose significant challenges for deployment on resource-constrained edge devices. To address these limitations, extensive research has focused on model compression techniques and hardware-aware acceleration strategies. Nonetheless, a comprehensive review that systematically categorizes these techniques and their trade-offs in accuracy, efficiency, and hardware adaptability for edge deployment remains lacking. This survey bridges this gap by providing a structured analysis of model compression techniques, software tools for inference on edge, and hardware acceleration strategies for ViTs. We discuss their impact on accuracy, efficiency, and hardware adaptability, highlighting key challenges and emerging research directions to advance ViT deployment on edge platforms, including graphics processing units (GPUs), application-specific integrated circuit (ASICs), and field-programmable gate arrays (FPGAs). The goal is to inspire further research with a contemporary guide on optimizing ViTs for efficient deployment on edge devices.
Deep neural networks (DNNs) have remarkably succeeded in various image processing tasks. However, their large size and computational complexity present significant challenges for deploying them in resource-constrained environments. This paper presents an innovative approach for integrating Mamba Architecture within a Progressive Knowledge Distillation (PKD) process to address the challenge of reducing model complexity while maintaining accuracy in image classification tasks. The proposed framework distills a large teacher model into progressively smaller student models, designed using Mamba blocks. Each student model is trained using Selective-State-Space Models (S-SSM) within the Mamba blocks, focusing on important input aspects while reducing computational complexity. The work's preliminary experiments use MNIST and CIFAR-10 as datasets to demonstrate the effectiveness of this approach. For MNIST, the teacher model achieves 98% accuracy. A set of seven student models as a group retained 63% of the teacher's FLOPs, approximating the teacher's performance with 98% accuracy. The weak student used only 1% of the teacher's FLOPs and maintained 72% accuracy. Similarly, for CIFAR-10, the students achieved 1% less accuracy compared to the teacher, with the small student retaining 5% of the teacher's FLOPs to achieve 50% accuracy. These results confirm the flexibility and scalability of Mamba Architecture, which can be integrated into PKD, succeeding in the process of finding students as weak learners. The framework provides a solution for deploying complex neural networks in real-time applications with a reduction in computational cost.
Machine unlearning, where users can request the deletion of a forget dataset, is becoming increasingly important because of numerous privacy regulations. Initial works on ``exact'' unlearning (e.g., retraining) incur large computational overheads. However, while computationally inexpensive, ``approximate'' methods have fallen short of reaching the effectiveness of exact unlearning: models produced fail to obtain comparable accuracy and prediction confidence on both the forget and test (i.e., unseen) dataset. Exploiting this observation, we propose a new unlearning method, Adversarial Machine UNlearning (AMUN), that outperforms prior state-of-the-art (SOTA) methods for image classification. AMUN lowers the confidence of the model on the forget samples by fine-tuning the model on their corresponding adversarial examples. Adversarial examples naturally belong to the distribution imposed by the model on the input space; fine-tuning the model on the adversarial examples closest to the corresponding forget samples (a) localizes the changes to the decision boundary of the model around each forget sample and (b) avoids drastic changes to the global behavior of the model, thereby preserving the model's accuracy on test samples. Using AMUN for unlearning a random $10\%$ of CIFAR-10 samples, we observe that even SOTA membership inference attacks cannot do better than random guessing.
Reinforcement Fine-Tuning (RFT) in Large Reasoning Models like OpenAI o1 learns from feedback on its answers, which is especially useful in applications when fine-tuning data is scarce. Recent open-source work like DeepSeek-R1 demonstrates that reinforcement learning with verifiable reward is one key direction in reproducing o1. While the R1-style model has demonstrated success in language models, its application in multi-modal domains remains under-explored. This work introduces Visual Reinforcement Fine-Tuning (Visual-RFT), which further extends the application areas of RFT on visual tasks. Specifically, Visual-RFT first uses Large Vision-Language Models (LVLMs) to generate multiple responses containing reasoning tokens and final answers for each input, and then uses our proposed visual perception verifiable reward functions to update the model via the policy optimization algorithm such as Group Relative Policy Optimization (GRPO). We design different verifiable reward functions for different perception tasks, such as the Intersection over Union (IoU) reward for object detection. Experimental results on fine-grained image classification, few-shot object detection, reasoning grounding, as well as open-vocabulary object detection benchmarks show the competitive performance and advanced generalization ability of Visual-RFT compared with Supervised Fine-tuning (SFT). For example, Visual-RFT improves accuracy by $24.3\%$ over the baseline in one-shot fine-grained image classification with around 100 samples. In few-shot object detection, Visual-RFT also exceeds the baseline by $21.9$ on COCO's two-shot setting and $15.4$ on LVIS. Our Visual-RFT represents a paradigm shift in fine-tuning LVLMs, offering a data-efficient, reward-driven approach that enhances reasoning and adaptability for domain-specific tasks.
Transformer-based deep learning models have achieved state-of-the-art performance across numerous language and vision tasks. While the self-attention mechanism, a core component of transformers, has proven capable of handling complex data patterns, it has been observed that the representational capacity of the attention matrix degrades significantly across transformer layers, thereby hurting its overall performance. In this work, we leverage the connection between self-attention computations and low-pass non-local means (NLM) smoothing filters and propose the Twicing Attention, a novel attention mechanism that uses kernel twicing procedure in nonparametric regression to alleviate the low-pass behavior of associated NLM smoothing with compelling theoretical guarantees and enhanced adversarial robustness. This approach enables the extraction and reuse of meaningful information retained in the residuals following the imperfect smoothing operation at each layer. Our proposed method offers two key advantages over standard self-attention: 1) a provably slower decay of representational capacity and 2) improved robustness and accuracy across various data modalities and tasks. We empirically demonstrate the performance gains of our model over baseline transformers on multiple tasks and benchmarks, including image classification and language modeling, on both clean and corrupted data.
Foundation model approaches such as masked auto-encoders (MAE) or its variations are now being successfully applied to satellite imagery. Most of the ongoing technical validation of foundation models have been applied to optical images like RGB or multi-spectral images. Due to difficulty in semantic labeling to create datasets and higher noise content with respect to optical images, Synthetic Aperture Radar (SAR) data has not been explored a lot in the field for foundation models. Therefore, in this work as a pre-training approach, we explored masked auto-encoder, specifically MixMAE on Sentinel-1 SAR images and its impact on SAR image classification tasks. Moreover, we proposed to use the physical characteristic of SAR data for applying weighting parameter on the auto-encoder training loss (MSE) to reduce the effect of speckle noise and very high values on the SAR images. Proposed SAR intensity-based weighting of the reconstruction loss demonstrates promising results both on SAR pre-training and downstream tasks specifically on flood detection compared with the baseline model.
Vision Transformers (ViTs) have significantly advanced image classification by applying self-attention on patch embeddings. However, the standard MLP blocks in each Transformer layer may not capture complex nonlinear dependencies optimally. In this paper, we propose ViKANformer, a Vision Transformer where we replace the MLP sub-layers with Kolmogorov-Arnold Network (KAN) expansions, including Vanilla KAN, Efficient-KAN, Fast-KAN, SineKAN, and FourierKAN, while also examining a Flash Attention variant. By leveraging the Kolmogorov-Arnold theorem, which guarantees that multivariate continuous functions can be expressed via sums of univariate continuous functions, we aim to boost representational power. Experimental results on MNIST demonstrate that SineKAN, Fast-KAN, and a well-tuned Vanilla KAN can achieve over 97% accuracy, albeit with increased training overhead. This trade-off highlights that KAN expansions may be beneficial if computational cost is acceptable. We detail the expansions, present training/test accuracy and F1/ROC metrics, and provide pseudocode and hyperparameters for reproducibility. Finally, we compare ViKANformer to a simple MLP and a small CNN baseline on MNIST, illustrating the efficiency of Transformer-based methods even on a small-scale dataset.
Recent advances in medical vision-language models (VLMs) demonstrate impressive performance in image classification tasks, driven by their strong zero-shot generalization capabilities. However, given the high variability and complexity inherent in medical imaging data, the ability of these models to detect out-of-distribution (OOD) data in this domain remains underexplored. In this work, we conduct the first systematic investigation into the OOD detection potential of medical VLMs. We evaluate state-of-the-art VLM-based OOD detection methods across a diverse set of medical VLMs, including both general and domain-specific purposes. To accurately reflect real-world challenges, we introduce a cross-modality evaluation pipeline for benchmarking full-spectrum OOD detection, rigorously assessing model robustness against both semantic shifts and covariate shifts. Furthermore, we propose a novel hierarchical prompt-based method that significantly enhances OOD detection performance. Extensive experiments are conducted to validate the effectiveness of our approach. The codes are available at https://github.com/PyJulie/Medical-VLMs-OOD-Detection.