Low-rank adaptation~(LoRA) has recently gained much interest in fine-tuning foundation models. It effectively reduces the number of trainable parameters by incorporating low-rank matrices $A$ and $B$ to represent the weight change, i.e., $\Delta W=BA$. Despite LoRA's progress, it faces storage challenges when handling extensive customization adaptations or larger base models. In this work, we aim to further compress trainable parameters by enjoying the powerful expressiveness of the Fourier transform. Specifically, we introduce FourierFT, which treats $\Delta W$ as a matrix in the spatial domain and learns only a small fraction of its spectral coefficients. With the trained spectral coefficients, we implement the inverse discrete Fourier transform to recover $\Delta W$. Empirically, our FourierFT method shows comparable or better performance with fewer parameters than LoRA on various tasks, including natural language understanding, natural language generation, instruction tuning, and image classification. For example, when performing instruction tuning on the LLaMA2-7B model, FourierFT surpasses LoRA with only 0.064M trainable parameters, compared to LoRA's 33.5M. Our code is released at \url{https://github.com/Chaos96/fourierft}.
Adversarial training (AT) trains models using adversarial examples (AEs), which are natural images modified with specific perturbations to mislead the model. These perturbations are constrained by a predefined perturbation budget $\epsilon$ and are equally applied to each pixel within an image. However, in this paper, we discover that not all pixels contribute equally to the accuracy on AEs (i.e., robustness) and accuracy on natural images (i.e., accuracy). Motivated by this finding, we propose Pixel-reweighted AdveRsarial Training (PART), a new framework that partially reduces $\epsilon$ for less influential pixels, guiding the model to focus more on key regions that affect its outputs. Specifically, we first use class activation mapping (CAM) methods to identify important pixel regions, then we keep the perturbation budget for these regions while lowering it for the remaining regions when generating AEs. In the end, we use these pixel-reweighted AEs to train a model. PART achieves a notable improvement in accuracy without compromising robustness on CIFAR-10, SVHN and TinyImagenet-200, justifying the necessity to allocate distinct weights to different pixel regions in robust classification.
The Segment Anything Model (SAM) was originally designed for label-agnostic mask generation. Does this model also possess inherent semantic understanding, of value to broader visual tasks? In this work we follow a multi-staged approach towards exploring this question. We firstly quantify SAM's semantic capabilities by comparing base image encoder efficacy under classification tasks, in comparison with established models (CLIP and DINOv2). Our findings reveal a significant lack of semantic discriminability in SAM feature representations, limiting potential for tasks that require class differentiation. This initial result motivates our exploratory study that attempts to enable semantic information via in-context learning with lightweight fine-tuning where we observe that generalisability to unseen classes remains limited. Our observations culminate in the proposal of a training-free approach that leverages DINOv2 features, towards better endowing SAM with semantic understanding and achieving instance-level class differentiation through feature-based similarity. Our study suggests that incorporation of external semantic sources provides a promising direction for the enhancement of SAM's utility with respect to complex visual tasks that require semantic understanding.
Label Smoothing (LS) is widely adopted to curb overconfidence in neural network predictions and enhance generalization. However, previous research shows that LS can force feature representations into excessively tight clusters, eroding intra-class distinctions. More recent findings suggest that LS also induces overconfidence in misclassifications, yet the precise mechanism remained unclear. In this work, we decompose the loss term introduced by LS, revealing two key components: (i) a regularization term that functions only when the prediction is correct, and (ii) an error-enhancement term that emerges under misclassifications. This latter term compels the model to reinforce incorrect predictions with exaggerated certainty, further collapsing the feature space. To address these issues, we propose Max Suppression (MaxSup), which uniformly applies the intended regularization to both correct and incorrect predictions by penalizing the top-1 logit instead of the ground-truth logit. Through feature analyses, we show that MaxSup restores intra-class variation and sharpens inter-class boundaries. Extensive experiments on image classification and downstream tasks confirm that MaxSup is a more robust alternative to LS. Code is available at: https://github.com/ZhouYuxuanYX/Maximum-Suppression-Regularization.
One of the challenges of computer vision is that it needs to adapt to color deviations in changeable environments. Therefore, minimizing the adverse effects of color deviation on the prediction is one of the main goals of vision task. Current solutions focus on using generative models to augment training data to enhance the invariance of input variation. However, such methods often introduce new noise, which limits the gain from generated data. To this end, this paper proposes a strategy eliminate deviation with deviation, which is named Random Color Dropout (RCD). Our hypothesis is that if there are color deviation between the query image and the gallery image, the retrieval results of some examples will be better after ignoring the color information. Specifically, this strategy balances the weights between color features and color-independent features in the neural network by dropouting partial color information in the training data, so as to overcome the effect of color devitaion. The proposed RCD can be combined with various existing ReID models without changing the learning strategy, and can be applied to other computer vision fields, such as object detection. Experiments on several ReID baselines and three common large-scale datasets such as Market1501, DukeMTMC, and MSMT17 have verified the effectiveness of this method. Experiments on Cross-domain tests have shown that this strategy is significant eliminating the domain gap. Furthermore, in order to understand the working mechanism of RCD, we analyzed the effectiveness of this strategy from the perspective of classification, which reveals that it may be better to utilize many instead of all of color information in visual tasks with strong domain variations.
The self-attention mechanism has attracted wide publicity for its most important advantage of modeling long dependency, and its variations in computer vision tasks, the non-local block tries to model the global dependency of the input feature maps. Gathering global contextual information will inevitably need a tremendous amount of memory and computing resources, which has been extensively studied in the past several years. However, there is a further problem with the self-attention scheme: is all information gathered from the global scope helpful for the contextual modelling? To our knowledge, few studies have focused on the problem. Aimed at both questions this paper proposes the salient positions-based attention scheme SPANet, which is inspired by some interesting observations on the attention maps and affinity matrices generated in self-attention scheme. We believe these observations are beneficial for better understanding of the self-attention. SPANet uses the salient positions selection algorithm to select only a limited amount of salient points to attend in the attention map computing. This approach will not only spare a lot of memory and computing resources, but also try to distill the positive information from the transformation of the input feature maps. In the implementation, considering the feature maps with channel high dimensions, which are completely different from the general visual image, we take the squared power of the feature maps along the channel dimension as the saliency metric of the positions. In general, different from the non-local block method, SPANet models the contextual information using only the selected positions instead of all, along the channel dimension instead of space dimension. Our source code is available at https://github.com/likyoo/SPANet.
In self-supervised learning, multi-granular features are heavily desired though rarely investigated, as different downstream tasks (e.g., general and fine-grained classification) often require different or multi-granular features, e.g.~fine- or coarse-grained one or their mixture. In this work, for the first time, we propose an effective MUlti-Granular Self-supervised learning (Mugs) framework to explicitly learn multi-granular visual features. Mugs has three complementary granular supervisions: 1) an instance discrimination supervision (IDS), 2) a novel local-group discrimination supervision (LGDS), and 3) a group discrimination supervision (GDS). IDS distinguishes different instances to learn instance-level fine-grained features. LGDS aggregates features of an image and its neighbors into a local-group feature, and pulls local-group features from different crops of the same image together and push them away for others. It provides complementary instance supervision to IDS via an extra alignment on local neighbors, and scatters different local-groups separately to increase discriminability. Accordingly, it helps learn high-level fine-grained features at a local-group level. Finally, to prevent similar local-groups from being scattered randomly or far away, GDS brings similar samples close and thus pulls similar local-groups together, capturing coarse-grained features at a (semantic) group level. Consequently, Mugs can capture three granular features that often enjoy higher generality on diverse downstream tasks over single-granular features, e.g.~instance-level fine-grained features in contrastive learning. By only pretraining on ImageNet-1K, Mugs sets new SoTA linear probing accuracy 82.1$\%$ on ImageNet-1K and improves previous SoTA by $1.1\%$. It also surpasses SoTAs on other tasks, e.g. transfer learning, detection and segmentation.
In this paper, the fourth version the Sloan Digital Sky Survey (SDSS-4), Data Release 16 dataset was used to classify the SDSS dataset into galaxies, stars, and quasars using machine learning and deep learning architectures. We efficiently utilize both image and metadata in tabular format to build a novel multi-modal architecture and achieve state-of-the-art results. In addition, our experiments on transfer learning using Imagenet weights on five different architectures (Resnet-50, DenseNet-121 VGG-16, Xception, and EfficientNet) reveal that freezing all layers and adding a final trainable layer may not be an optimal solution for transfer learning. It is hypothesized that higher the number of trainable layers, higher will be the training time and accuracy of predictions. It is also hypothesized that any subsequent increase in the number of training layers towards the base layers will not increase in accuracy as the pre trained lower layers only help in low level feature extraction which would be quite similar in all the datasets. Hence the ideal level of trainable layers needs to be identified for each model in respect to the number of parameters. For the tabular data, we compared classical machine learning algorithms (Logistic Regression, Random Forest, Decision Trees, Adaboost, LightGBM etc.,) with artificial neural networks. Our works shed new light on transfer learning and multi-modal deep learning architectures. The multi-modal architecture not only resulted in higher metrics (accuracy, precision, recall, F1 score) than models using only image data or tabular data. Furthermore, multi-modal architecture achieved the best metrics in lesser training epochs and improved the metrics on all classes.
Deep learning in computer vision has achieved great success with the price of large-scale labeled training data. However, exhaustive data annotation is impracticable for each task of all domains of interest, due to high labor costs and unguaranteed labeling accuracy. Besides, the uncontrollable data collection process produces non-IID training and test data, where undesired duplication may exist. All these nuisances may hinder the verification of typical theories and exposure to new findings. To circumvent them, an alternative is to generate synthetic data via 3D rendering with domain randomization. We in this work push forward along this line by doing profound and extensive research on bare supervised learning and downstream domain adaptation. Specifically, under the well-controlled, IID data setting enabled by 3D rendering, we systematically verify the typical, important learning insights, e.g., shortcut learning, and discover the new laws of various data regimes and network architectures in generalization. We further investigate the effect of image formation factors on generalization, e.g., object scale, material texture, illumination, camera viewpoint, and background in a 3D scene. Moreover, we use the simulation-to-reality adaptation as a downstream task for comparing the transferability between synthetic and real data when used for pre-training, which demonstrates that synthetic data pre-training is also promising to improve real test results. Lastly, to promote future research, we develop a new large-scale synthetic-to-real benchmark for image classification, termed S2RDA, which provides more significant challenges for transfer from simulation to reality. The code and datasets are available at https://github.com/huitangtang/On_the_Utility_of_Synthetic_Data.
In multimedia understanding tasks, corrupted samples pose a critical challenge, because when fed to machine learning models they lead to performance degradation. In the past, three groups of approaches have been proposed to handle noisy data: i) enhancer and denoiser modules to improve the quality of the noisy data, ii) data augmentation approaches, and iii) domain adaptation strategies. All the aforementioned approaches come with drawbacks that limit their applicability; the first has high computational costs and requires pairs of clean-corrupted data for training, while the others only allow deployment of the same task/network they were trained on (\ie, when upstream and downstream task/network are the same). In this paper, we propose SyMPIE to solve these shortcomings. To this end, we design a small, modular, and efficient (just 2GFLOPs to process a Full HD image) system to enhance input data for robust downstream multimedia understanding with minimal computational cost. Our SyMPIE is pre-trained on an upstream task/network that should not match the downstream ones and does not need paired clean-corrupted samples. Our key insight is that most input corruptions found in real-world tasks can be modeled through global operations on color channels of images or spatial filters with small kernels. We validate our approach on multiple datasets and tasks, such as image classification (on ImageNetC, ImageNetC-Bar, VizWiz, and a newly proposed mixed corruption benchmark named ImageNetC-mixed) and semantic segmentation (on Cityscapes, ACDC, and DarkZurich) with consistent improvements of about 5\% relative accuracy gain across the board. The code of our approach and the new ImageNetC-mixed benchmark will be made available upon publication.
Individual neurons in convolutional neural networks supervised for image-level classification tasks have been shown to implicitly learn semantically meaningful concepts ranging from simple textures and shapes to whole or partial objects - forming a "dictionary" of concepts acquired through the learning process. In this work we introduce a simple, efficient zero-shot learning approach based on this observation. Our approach, which we call Neuron Importance-AwareWeight Transfer (NIWT), learns to map domain knowledge about novel "unseen" classes onto this dictionary of learned concepts and then optimizes for network parameters that can effectively combine these concepts - essentially learning classifiers by discovering and composing learned semantic concepts in deep networks. Our approach shows improvements over previous approaches on the CUBirds and AWA2 generalized zero-shot learning benchmarks. We demonstrate our approach on a diverse set of semantic inputs as external domain knowledge including attributes and natural language captions. Moreover by learning inverse mappings, NIWT can provide visual and textual explanations for the predictions made by the newly learned classifiers and provide neuron names. Our code is available at https://github.com/ramprs/neuron-importance-zsl.
In this paper, we propose a novel edge-labeling graph neural network (EGNN), which adapts a deep neural network on the edge-labeling graph, for few-shot learning. The previous graph neural network (GNN) approaches in few-shot learning have been based on the node-labeling framework, which implicitly models the intra-cluster similarity and the inter-cluster dissimilarity. In contrast, the proposed EGNN learns to predict the edge-labels rather than the node-labels on the graph that enables the evolution of an explicit clustering by iteratively updating the edge-labels with direct exploitation of both intra-cluster similarity and the inter-cluster dissimilarity. It is also well suited for performing on various numbers of classes without retraining, and can be easily extended to perform a transductive inference. The parameters of the EGNN are learned by episodic training with an edge-labeling loss to obtain a well-generalizable model for unseen low-data problem. On both of the supervised and semi-supervised few-shot image classification tasks with two benchmark datasets, the proposed EGNN significantly improves the performances over the existing GNNs.
The performance of existing point cloud-based 3D object detection methods heavily relies on large-scale high-quality 3D annotations. However, such annotations are often tedious and expensive to collect. Semi-supervised learning is a good alternative to mitigate the data annotation issue, but has remained largely unexplored in 3D object detection. Inspired by the recent success of self-ensembling technique in semi-supervised image classification task, we propose SESS, a self-ensembling semi-supervised 3D object detection framework. Specifically, we design a thorough perturbation scheme to enhance generalization of the network on unlabeled and new unseen data. Furthermore, we propose three consistency losses to enforce the consistency between two sets of predicted 3D object proposals, to facilitate the learning of structure and semantic invariances of objects. Extensive experiments conducted on SUN RGB-D and ScanNet datasets demonstrate the effectiveness of SESS in both inductive and transductive semi-supervised 3D object detection. Our SESS achieves competitive performance compared to the state-of-the-art fully-supervised method by using only 50% labeled data. Our code is available at https://github.com/Na-Z/sess.
We present DarkneTZ, a framework that uses an edge device's Trusted Execution Environment (TEE) in conjunction with model partitioning to limit the attack surface against Deep Neural Networks (DNNs). Increasingly, edge devices (smartphones and consumer IoT devices) are equipped with pre-trained DNNs for a variety of applications. This trend comes with privacy risks as models can leak information about their training data through effective membership inference attacks (MIAs). We evaluate the performance of DarkneTZ, including CPU execution time, memory usage, and accurate power consumption, using two small and six large image classification models. Due to the limited memory of the edge device's TEE, we partition model layers into more sensitive layers (to be executed inside the device TEE), and a set of layers to be executed in the untrusted part of the operating system. Our results show that even if a single layer is hidden, we can provide reliable model privacy and defend against state of the art MIAs, with only 3% performance overhead. When fully utilizing the TEE, DarkneTZ provides model protections with up to 10% overhead.
We generalize a graph-based multiclass semi-supervised classification technique based on diffuse interface methods to multilayer graphs. Besides the treatment of various applications with an inherent multilayer structure, we present a very flexible approach that interprets high-dimensional data in a low-dimensional multilayer graph representation. Highly efficient numerical methods involving the spectral decomposition of the corresponding differential graph operators as well as fast matrix-vector products based on the nonequispaced fast Fourier transform (NFFT) enable the rapid treatment of large and high-dimensional data sets. We perform various numerical tests putting a special focus on image segmentation. In particular, we test the performance of our method on data sets with up to 10 million nodes per layer as well as up to 104 dimensions resulting in graphs with up to 52 layers. While all presented numerical experiments can be run on an average laptop computer, the linear dependence per iteration step of the runtime on the network size in all stages of our algorithm makes it scalable to even larger and higher-dimensional problems.
Sequential manipulation tasks require a robot to perceive the state of an environment and plan a sequence of actions leading to a desired goal state. In such tasks, the ability to reason about spatial relations among object entities from raw sensor inputs is crucial in order to determine when a task has been completed and which actions can be executed. In this work, we propose SORNet (Spatial Object-Centric Representation Network), a framework for learning object-centric representations from RGB images conditioned on a set of object queries, represented as image patches called canonical object views. With only a single canonical view per object and no annotation, SORNet generalizes zero-shot to object entities whose shape and texture are both unseen during training. We evaluate SORNet on various spatial reasoning tasks such as spatial relation classification and relative direction regression in complex tabletop manipulation scenarios and show that SORNet significantly outperforms baselines including state-of-the-art representation learning techniques. We also demonstrate the application of the representation learned by SORNet on visual-servoing and task planning for sequential manipulation on a real robot.
Although Deep Neural Networks (DNNs) achieve excellent performance on many real-world tasks, they are highly vulnerable to adversarial attacks. A leading defense against such attacks is adversarial training, a technique in which a DNN is trained to be robust to adversarial attacks by introducing adversarial noise to its input. This procedure is effective but must be done during the training phase. In this work, we propose Augmented Random Forest (ARF), a simple and easy-to-use strategy for robustifying an existing pretrained DNN without modifying its weights. For every image, we generate randomized test time augmentations by applying diverse color, blur, noise, and geometric transforms. Then we use the DNN's logits output to train a simple random forest to predict the real class label. Our method achieves state-of-the-art adversarial robustness on a diversity of white and black box attacks with minimal compromise on the natural images' classification. We test ARF also against numerous adaptive white-box attacks and it shows excellent results when combined with adversarial training. Code is available at https://github.com/giladcohen/ARF.
Knowledge Graph (KG) and its variant of ontology have been widely used for knowledge representation, and have shown to be quite effective in augmenting Zero-shot Learning (ZSL). However, existing ZSL methods that utilize KGs all neglect the intrinsic complexity of inter-class relationships represented in KGs. One typical feature is that a class is often related to other classes in different semantic aspects. In this paper, we focus on ontologies for augmenting ZSL, and propose to learn disentangled ontology embeddings guided by ontology properties to capture and utilize more fine-grained class relationships in different aspects. We also contribute a new ZSL framework named DOZSL, which contains two new ZSL solutions based on generative models and graph propagation models, respectively, for effectively utilizing the disentangled ontology embeddings. Extensive evaluations have been conducted on five benchmarks across zero-shot image classification (ZS-IMGC) and zero-shot KG completion (ZS-KGC). DOZSL often achieves better performance than the state-of-the-art, and its components have been verified by ablation studies and case studies. Our codes and datasets are available at https://github.com/zjukg/DOZSL.
In recent years, predictive machine learning methods have gained prominence in various scientific domains. However, due to their black-box nature, it is essential to establish trust in these models before accepting them as accurate. One promising strategy for assigning trust involves employing explanation techniques that elucidate the rationale behind a black-box model's predictions in a manner that humans can understand. However, assessing the degree of human interpretability of the rationale generated by such methods is a nontrivial challenge. In this work, we introduce interpretation entropy as a universal solution for assessing the degree of human interpretability associated with any linear model. Using this concept and drawing inspiration from classical thermodynamics, we present Thermodynamics-inspired Explainable Representations of AI and other black-box Paradigms (TERP), a method for generating accurate, and human-interpretable explanations for black-box predictions in a model-agnostic manner. To demonstrate the wide-ranging applicability of TERP, we successfully employ it to explain various black-box model architectures, including deep learning Autoencoders, Recurrent Neural Networks, and Convolutional Neural Networks, across diverse domains such as molecular simulations, text, and image classification.
Tree perception is an essential building block toward autonomous forestry operations. Current developments generally consider input data from lidar sensors to solve forest navigation, tree detection and diameter estimation problems. Whereas cameras paired with deep learning algorithms usually address species classification or forest anomaly detection. In either of these cases, data unavailability and forest diversity restrain deep learning developments for autonomous systems. So, we propose two densely annotated image datasets - 43k synthetic, 100 real - for bounding box, segmentation mask and keypoint detections to assess the potential of vision-based methods. Deep neural network models trained on our datasets achieve a precision of 90.4% for tree detection, 87.2% for tree segmentation, and centimeter accurate keypoint estimations. We measure our models' generalizability when testing it on other forest datasets, and their scalability with different dataset sizes and architectural improvements. Overall, the experimental results offer promising avenues toward autonomous tree felling operations and other applied forestry problems. The datasets and pre-trained models in this article are publicly available on \href{https://github.com/norlab-ulaval/PercepTreeV1}{GitHub} (https://github.com/norlab-ulaval/PercepTreeV1).