In the last few years, the success of Transformers in computer vision has stimulated the discovery of many alternative models that compete with Transformers, such as the MLP-Mixer. Despite their weak inductive bias, these models have achieved performance comparable to well-studied convolutional neural networks. Recent studies on modern Hopfield networks suggest the correspondence between certain energy-based associative memory models and Transformers or MLP-Mixer, and shed some light on the theoretical background of the Transformer-type architectures design. In this paper, we generalize the correspondence to the recently introduced hierarchical Hopfield network, and find iMixer, a novel generalization of MLP-Mixer model. Unlike ordinary feedforward neural networks, iMixer involves MLP layers that propagate forward from the output side to the input side. We characterize the module as an example of invertible, implicit, and iterative mixing module. We evaluate the model performance with various datasets on image classification tasks, and find that iMixer, despite its unique architecture, exhibits stable learning capabilities and achieves performance comparable to or better than the baseline vanilla MLP-Mixer. The results imply that the correspondence between the Hopfield networks and the Mixer models serves as a principle for understanding a broader class of Transformer-like architecture designs.
Large-scale pre-trained transformers have demonstrated remarkable success in various computer vision tasks. However, it is still highly challenging to fully fine-tune these models for downstream tasks due to their high computational and storage costs. Recently, Parameter-Efficient Tuning (PETuning) techniques, e.g., Visual Prompt Tuning (VPT) and Low-Rank Adaptation (LoRA), have significantly reduced the computation and storage cost by inserting lightweight prompt modules into the pre-trained models and tuning these prompt modules with a small number of trainable parameters, while keeping the transformer backbone frozen. Although only a few parameters need to be adjusted, most PETuning methods still require a significant amount of downstream task training data to achieve good results. The performance is inadequate on low-data regimes, especially when there are only one or two examples per class. To this end, we first empirically identify the poor performance is mainly due to the inappropriate way of initializing prompt modules, which has also been verified in the pre-trained language models. Next, we propose a Pre-trained Visual Parameter-efficient (PVP) Tuning framework, which pre-trains the parameter-efficient tuning modules first and then leverages the pre-trained modules along with the pre-trained transformer backbone to perform parameter-efficient tuning on downstream tasks. Experiment results on five Fine-Grained Visual Classification (FGVC) and VTAB-1k datasets demonstrate that our proposed method significantly outperforms state-of-the-art PETuning methods.
Modern neural networks have revolutionized the fields of computer vision (CV) and Natural Language Processing (NLP). They are widely used for solving complex CV tasks and NLP tasks such as image classification, image generation, and machine translation. Most state-of-the-art neural networks are over-parameterized and require a high computational cost. One straightforward solution is to replace the layers of the networks with their low-rank tensor approximations using different tensor decomposition methods. This paper reviews six tensor decomposition methods and illustrates their ability to compress model parameters of convolutional neural networks (CNNs), recurrent neural networks (RNNs) and Transformers. The accuracy of some compressed models can be higher than the original versions. Evaluations indicate that tensor decompositions can achieve significant reductions in model size, run-time and energy consumption, and are well suited for implementing neural networks on edge devices.
Self-supervised learning (SSL) techniques have recently been integrated into the few-shot learning (FSL) framework and have shown promising results in improving the few-shot image classification performance. However, existing SSL approaches used in FSL typically seek the supervision signals from the global embedding of every single image. Therefore, during the episodic training of FSL, these methods cannot capture and fully utilize the local visual information in image samples and the data structure information of the whole episode, which are beneficial to FSL. To this end, we propose to augment the few-shot learning objective with a novel self-supervised Episodic Spatial Pretext Task (ESPT). Specifically, for each few-shot episode, we generate its corresponding transformed episode by applying a random geometric transformation to all the images in it. Based on these, our ESPT objective is defined as maximizing the local spatial relationship consistency between the original episode and the transformed one. With this definition, the ESPT-augmented FSL objective promotes learning more transferable feature representations that capture the local spatial features of different images and their inter-relational structural information in each input episode, thus enabling the model to generalize better to new categories with only a few samples. Extensive experiments indicate that our ESPT method achieves new state-of-the-art performance for few-shot image classification on three mainstay benchmark datasets. The source code will be available at: https://github.com/Whut-YiRong/ESPT.
With the development of Vision-Language Pre-training Models (VLPMs) represented by CLIP and ALIGN, significant breakthroughs have been achieved for association-based visual tasks such as image classification and image-text retrieval by the zero-shot capability of CLIP without fine-tuning. However, CLIP is hard to apply to generation-based tasks. This is due to the lack of decoder architecture and pre-training tasks for generation. Although previous works have created generation capacity for CLIP through additional language models, a modality gap between the CLIP representations of different modalities and the inability of CLIP to model the offset of this gap, which fails the concept to transfer across modalities. To solve the problem, we try to map images/videos to the language modality and generate captions from the language modality. In this paper, we propose the K-nearest-neighbor Cross-modality Mapping (Knight), a zero-shot method from association to generation. With text-only unsupervised training, Knight achieves State-of-the-Art performance in zero-shot methods for image captioning and video captioning. Our code is available at https://github.com/junyangwang0410/Knight.
Incorporating deep learning (DL) classification models into unmanned aerial vehicles (UAVs) can significantly augment search-and-rescue operations and disaster management efforts. In such critical situations, the UAV's ability to promptly comprehend the crisis and optimally utilize its limited power and processing resources to narrow down search areas is crucial. Therefore, developing an efficient and lightweight method for scene classification is of utmost importance. However, current approaches tend to prioritize accuracy on benchmark datasets at the expense of computational efficiency. To address this shortcoming, we introduce the Wider ATTENTION EfficientNet (WATT-EffNet), a novel method that achieves higher accuracy with a more lightweight architecture compared to the baseline EfficientNet. The WATT-EffNet leverages width-wise incremental feature modules and attention mechanisms over width-wise features to ensure the network structure remains lightweight. We evaluate our method on a UAV-based aerial disaster image classification dataset and demonstrate that it outperforms the baseline by up to 15 times in terms of classification accuracy and 38.3% in terms of computing efficiency as measured by Floating Point Operations per second (FLOPs). Additionally, we conduct an ablation study to investigate the effect of varying the width of WATT-EffNet on accuracy and computational efficiency. Our code is available at \url{https://github.com/TanmDL/WATT-EffNet}.
Meta-learning, also known as ``learning to learn'', enables models to acquire great generalization abilities by learning from various tasks. Recent advancements have made these models applicable across various fields without data constraints, offering new opportunities for general artificial intelligence. However, applying these models can be challenging due to their often task-specific, standalone nature and the technical barriers involved. To address this challenge, we develop AwesomeMeta+, a prototyping and learning system designed to standardize the key components of meta-learning within the context of systems engineering. It standardizes different components of meta-learning and uses a building block metaphor to assist in model construction. By employing a modular, building-block approach, AwesomeMeta+ facilitates the construction of meta-learning models that can be adapted and optimized for specific application needs in real-world systems. The system is developed to support the full lifecycle of meta-learning system engineering, from design to deployment, by enabling users to assemble compatible algorithmic modules. We evaluate AwesomeMeta+ through feedback from 50 researchers and a series of machine-based tests and user studies. The results demonstrate that AwesomeMeta+ enhances users' understanding of meta-learning principles, accelerates system engineering processes, and provides valuable decision-making support for efficient deployment of meta-learning systems in complex application scenarios.
Bayesian optimization (BO) has contributed greatly to improving model performance by suggesting promising hyperparameter configurations iteratively based on observations from multiple training trials. However, only partial knowledge (i.e., the measured performances of trained models and their hyperparameter configurations) from previous trials is transferred. On the other hand, Self-Distillation (SD) only transfers partial knowledge learned by the task model itself. To fully leverage the various knowledge gained from all training trials, we propose the BOSS framework, which combines BO and SD. BOSS suggests promising hyperparameter configurations through BO and carefully selects pre-trained models from previous trials for SD, which are otherwise abandoned in the conventional BO process. BOSS achieves significantly better performance than both BO and SD in a wide range of tasks including general image classification, learning with noisy labels, semi-supervised learning, and medical image analysis tasks.
Due to a huge volume of information in many domains, the need for classification methods is imperious. In spite of many advances, most of the approaches require a large amount of labeled data, which is often not available, due to costs and difficulties of manual labeling processes. In this scenario, unsupervised and semi-supervised approaches have been gaining increasing attention. The GCNs (Graph Convolutional Neural Networks) represent a promising solution since they encode the neighborhood information and have achieved state-of-the-art results on scenarios with limited labeled data. However, since GCNs require graph-structured data, their use for semi-supervised image classification is still scarce in the literature. In this work, we propose a novel approach, the Manifold-GCN, based on GCNs for semi-supervised image classification. The main hypothesis of this paper is that the use of manifold learning to model the graph structure can further improve the GCN classification. To the best of our knowledge, this is the first framework that allows the combination of GCNs with different types of manifold learning approaches for image classification. All manifold learning algorithms employed are completely unsupervised, which is especially useful for scenarios where the availability of labeled data is a concern. A broad experimental evaluation was conducted considering 5 GCN models, 3 manifold learning approaches, 3 image datasets, and 5 deep features. The results reveal that our approach presents better accuracy than traditional and recent state-of-the-art methods with very efficient run times for both training and testing.
Adversarial attacks and defenses have gained increasing interest on computer vision systems in recent years, but as of today, most investigations are limited to images. However, many artificial intelligence models actually handle documentary data, which is very different from real world images. Hence, in this work, we try to apply the adversarial attack philosophy on documentary and natural data and to protect models against such attacks. We focus our work on untargeted gradient-based, transfer-based and score-based attacks and evaluate the impact of adversarial training, JPEG input compression and grey-scale input transformation on the robustness of ResNet50 and EfficientNetB0 model architectures. To the best of our knowledge, no such work has been conducted by the community in order to study the impact of these attacks on the document image classification task.
Medical image analysis is a hot research topic because of its usefulness in different clinical applications, such as early disease diagnosis and treatment. Convolutional neural networks (CNNs) have become the de-facto standard in medical image analysis tasks because of their ability to learn complex features from the available datasets, which makes them surpass humans in many image-understanding tasks. In addition to CNNs, transformer architectures also have gained popularity for medical image analysis tasks. However, despite progress in the field, there are still potential areas for improvement. This study uses different CNNs and transformer-based methods with a wide range of data augmentation techniques. We evaluated their performance on three medical image datasets from different modalities. We evaluated and compared the performance of the vision transformer model with other state-of-the-art (SOTA) pre-trained CNN networks. For Chest X-ray, our vision transformer model achieved the highest F1 score of 0.9532, recall of 0.9533, Matthews correlation coefficient (MCC) of 0.9259, and ROC-AUC score of 0.97. Similarly, for the Kvasir dataset, we achieved an F1 score of 0.9436, recall of 0.9437, MCC of 0.9360, and ROC-AUC score of 0.97. For the Kvasir-Capsule (a large-scale VCE dataset), our ViT model achieved a weighted F1-score of 0.7156, recall of 0.7182, MCC of 0.3705, and ROC-AUC score of 0.57. We found that our transformer-based models were better or more effective than various CNN models for classifying different anatomical structures, findings, and abnormalities. Our model showed improvement over the CNN-based approaches and suggests that it could be used as a new benchmarking algorithm for algorithm development.
The recently proposed data augmentation TransMix employs attention labels to help visual transformers (ViT) achieve better robustness and performance. However, TransMix is deficient in two aspects: 1) The image cropping method of TransMix may not be suitable for ViTs. 2) At the early stage of training, the model produces unreliable attention maps. TransMix uses unreliable attention maps to compute mixed attention labels that can affect the model. To address the aforementioned issues, we propose MaskMix and Progressive Attention Labeling (PAL) in image and label space, respectively. In detail, from the perspective of image space, we design MaskMix, which mixes two images based on a patch-like grid mask. In particular, the size of each mask patch is adjustable and is a multiple of the image patch size, which ensures each image patch comes from only one image and contains more global contents. From the perspective of label space, we design PAL, which utilizes a progressive factor to dynamically re-weight the attention weights of the mixed attention label. Finally, we combine MaskMix and Progressive Attention Labeling as our new data augmentation method, named MixPro. The experimental results show that our method can improve various ViT-based models at scales on ImageNet classification (73.8\% top-1 accuracy based on DeiT-T for 300 epochs). After being pre-trained with MixPro on ImageNet, the ViT-based models also demonstrate better transferability to semantic segmentation, object detection, and instance segmentation. Furthermore, compared to TransMix, MixPro also shows stronger robustness on several benchmarks. The code is available at https://github.com/fistyee/MixPro.
Feature distillation makes the student mimic the intermediate features of the teacher. Nearly all existing feature-distillation methods use L2 distance or its slight variants as the distance metric between teacher and student features. However, while L2 distance is isotropic w.r.t. all dimensions, the neural network's operation on different dimensions is usually anisotropic, i.e., perturbations with the same 2-norm but in different dimensions of intermediate features lead to changes in the final output with largely different magnitude. Considering this, we argue that the similarity between teacher and student features should not be measured merely based on their appearance (i.e., L2 distance), but should, more importantly, be measured by their difference in function, namely how later layers of the network will read, decode, and process them. Therefore, we propose Function-Consistent Feature Distillation (FCFD), which explicitly optimizes the functional similarity between teacher and student features. The core idea of FCFD is to make teacher and student features not only numerically similar, but more importantly produce similar outputs when fed to the later part of the same network. With FCFD, the student mimics the teacher more faithfully and learns more from the teacher. Extensive experiments on image classification and object detection demonstrate the superiority of FCFD to existing methods. Furthermore, we can combine FCFD with many existing methods to obtain even higher accuracy. Our codes are available at https://github.com/LiuDongyang6/FCFD.
Occlusions of objects is one of the indispensable problems in Computer vision. While Convolutional Neural Net-works (CNNs) provide various state of the art approaches for regular image classification, they however, prove to be not as effective for the classification of images with partial occlusions. Partial occlusion is scenario where an object is occluded partially by some other object/space. This problem when solved,holds tremendous potential to facilitate various scenarios. We in particular are interested in autonomous driving scenario and its implications in the same. Autonomous vehicle research is one of the hot topics of this decade, there are ample situations of partial occlusions of a driving sign or a person or other objects at different angles. Considering its prime importance in situations which can be further extended to video analytics of traffic data to handle crimes, anticipate income levels of various groups etc.,this holds the potential to be exploited in many ways. In this paper, we introduce our own synthetically created dataset by utilising Stanford Car Dataset and adding occlusions of various sizes and nature to it. On this created dataset, we conducted a comprehensive analysis using various state of the art CNN models such as VGG-19, ResNet 50/101, GoogleNet, DenseNet 121. We further in depth study the effect of varying occlusion proportions and nature on the performance of these models by fine tuning and training these from scratch on dataset and how is it likely to perform when trained in different scenarios, i.e., performance when training with occluded images and unoccluded images, which model is more robust to partial occlusions and soon.
Resource-constrained Edge Devices (EDs), e.g., IoT sensors and microcontroller units, are expected to make intelligent decisions using Deep Learning (DL) inference at the edge of the network. Toward this end, there is a significant research effort in developing tinyML models - Deep Learning (DL) models with reduced computation and memory storage requirements - that can be embedded on these devices. However, tinyML models have lower inference accuracy. On a different front, DNN partitioning and inference offloading techniques were studied for distributed DL inference between EDs and Edge Servers (ESs). In this paper, we explore Hierarchical Inference (HI), a novel approach proposed by Vishnu et al. 2023, arXiv:2304.00891v1 , for performing distributed DL inference at the edge. Under HI, for each data sample, an ED first uses a local algorithm (e.g., a tinyML model) for inference. Depending on the application, if the inference provided by the local algorithm is incorrect or further assistance is required from large DL models on edge or cloud, only then the ED offloads the data sample. At the outset, HI seems infeasible as the ED, in general, cannot know if the local inference is sufficient or not. Nevertheless, we present the feasibility of implementing HI for machine fault detection and image classification applications. We demonstrate its benefits using quantitative analysis and argue that using HI will result in low latency, bandwidth savings, and energy savings in edge AI systems.
The paper discusses the use of the Absolute activation function in classification neural networks. An examples are shown of using this activation function in simple and more complex problems. Using as a baseline LeNet-5 network for solving the MNIST problem, the efficiency of Absolute activation function is shown in comparison with the use of Tanh, ReLU and SeLU activations. It is shown that in deep networks Absolute activation does not cause vanishing and exploding gradients, and therefore Absolute activation can be used in both simple and deep neural networks. Due to high volatility of training networks with Absolute activation, a special modification of ADAM training algorithm is used, that estimates lower bound of accuracy at any test dataset using validation dataset analysis at each training epoch, and uses this value to stop/decrease learning rate, and re-initializes ADAM algorithm between these steps. It is shown that solving the MNIST problem with the LeNet-like architectures based on Absolute activation allows to significantly reduce the number of trained parameters in the neural network with improving the prediction accuracy.
This work presents a novel label-efficient selfsupervised representation learning-based approach for classifying diabetic retinopathy (DR) images in cross-domain settings. Most of the existing DR image classification methods are based on supervised learning which requires a lot of time-consuming and expensive medical domain experts-annotated data for training. The proposed approach uses the prior learning from the source DR image dataset to classify images drawn from the target datasets. The image representations learned from the unlabeled source domain dataset through contrastive learning are used to classify DR images from the target domain dataset. Moreover, the proposed approach requires a few labeled images to perform successfully on DR image classification tasks in cross-domain settings. The proposed work experiments with four publicly available datasets: EyePACS, APTOS 2019, MESSIDOR-I, and Fundus Images for self-supervised representation learning-based DR image classification in cross-domain settings. The proposed method achieves state-of-the-art results on binary and multiclassification of DR images, even in cross-domain settings. The proposed method outperforms the existing DR image binary and multi-class classification methods proposed in the literature. The proposed method is also validated qualitatively using class activation maps, revealing that the method can learn explainable image representations. The source code and trained models are published on GitHub.
Interpreting remote sensing imagery enables numerous downstream applications ranging from land-use planning to deforestation monitoring. Robustly classifying this data is challenging due to the Earth's geographic diversity. While many distinct satellite and aerial image classification datasets exist, there is yet to be a benchmark curated that suitably covers this diversity. In this work, we introduce SATellite ImageNet (SATIN), a metadataset curated from 27 existing remotely sensed datasets, and comprehensively evaluate the zero-shot transfer classification capabilities of a broad range of vision-language (VL) models on SATIN. We find SATIN to be a challenging benchmark-the strongest method we evaluate achieves a classification accuracy of 52.0%. We provide a $\href{https://satinbenchmark.github.io}{\text{public leaderboard}}$ to guide and track the progress of VL models in this important domain.
Visual representation based on covariance matrix has demonstrates its efficacy for image classification by characterising the pairwise correlation of different channels in convolutional feature maps. However, pairwise correlation will become misleading once there is another channel correlating with both channels of interest, resulting in the ``confounding'' effect. For this case, ``partial correlation'' which removes the confounding effect shall be estimated instead. Nevertheless, reliably estimating partial correlation requires to solve a symmetric positive definite matrix optimisation, known as sparse inverse covariance estimation (SICE). How to incorporate this process into CNN remains an open issue. In this work, we formulate SICE as a novel structured layer of CNN. To ensure end-to-end trainability, we develop an iterative method to solve the above matrix optimisation during forward and backward propagation steps. Our work obtains a partial correlation based deep visual representation and mitigates the small sample problem often encountered by covariance matrix estimation in CNN. Computationally, our model can be effectively trained with GPU and works well with a large number of channels of advanced CNNs. Experiments show the efficacy and superior classification performance of our deep visual representation compared to covariance matrix based counterparts.
Recent years have witnessed the outstanding success of deep learning in various fields such as vision and natural language processing. This success is largely indebted to the massive size of deep learning models that is expected to increase unceasingly. This growth of the deep learning models is accompanied by issues related to their considerable energy consumption, both during the training and inference phases, as well as their scalability. Although a number of work based on unconventional physical systems have been proposed which addresses the issue of energy efficiency in the inference phase, efficient training of deep learning models has remained unaddressed. So far, training of digital deep learning models mainly relies on backpropagation, which is not suitable for physical implementation as it requires perfect knowledge of the computation performed in the so-called forward pass of the neural network. Here, we tackle this issue by proposing a simple deep neural network architecture augmented by a biologically plausible learning algorithm, referred to as "model-free forward-forward training". The proposed architecture enables training deep physical neural networks consisting of layers of physical nonlinear systems, without requiring detailed knowledge of the nonlinear physical layers' properties. We show that our method outperforms state-of-the-art hardware-aware training methods by improving training speed, decreasing digital computations, and reducing power consumption in physical systems. We demonstrate the adaptability of the proposed method, even in systems exposed to dynamic or unpredictable external perturbations. To showcase the universality of our approach, we train diverse wave-based physical neural networks that vary in the underlying wave phenomenon and the type of non-linearity they use, to perform vowel and image classification tasks experimentally.