Attention guidance is an approach to addressing dataset bias in deep learning, where the model relies on incorrect features to make decisions. Focusing on image classification tasks, we propose an efficient human-in-the-loop system to interactively direct the attention of classifiers to the regions specified by users, thereby reducing the influence of co-occurrence bias and improving the transferability and interpretability of a DNN. Previous approaches for attention guidance require the preparation of pixel-level annotations and are not designed as interactive systems. We present a new interactive method to allow users to annotate images with simple clicks, and study a novel active learning strategy to significantly reduce the number of annotations. We conducted both a numerical evaluation and a user study to evaluate the proposed system on multiple datasets. Compared to the existing non-active-learning approach which usually relies on huge amounts of polygon-based segmentation masks to fine-tune or train the DNNs, our system can save lots of labor and money and obtain a fine-tuned network that works better even when the dataset is biased. The experiment results indicate that the proposed system is efficient, reasonable, and reliable.
Convolutional Neural Networks (ConvNets) have been thoroughly researched for document image classification and are known for their exceptional performance in unimodal image-based document classification. Recently, however, there has been a sudden shift in the field towards multimodal approaches that simultaneously learn from the visual and textual features of the documents. While this has led to significant advances in the field, it has also led to a waning interest in improving pure ConvNets-based approaches. This is not desirable, as many of the multimodal approaches still use ConvNets as their visual backbone, and thus improving ConvNets is essential to improving these approaches. In this paper, we present DocXClassifier, a ConvNet-based approach that, using state-of-the-art model design patterns together with modern data augmentation and training strategies, not only achieves significant performance improvements in image-based document classification, but also outperforms some of the recently proposed multimodal approaches. Moreover, DocXClassifier is capable of generating transformer-like attention maps, which makes it inherently interpretable, a property not found in previous image-based classification models. Our approach achieves a new peak performance in image-based classification on two popular document datasets, namely RVL-CDIP and Tobacco3482, with a top-1 classification accuracy of 94.17% and 95.57% on the two datasets, respectively. Moreover, it sets a new record for the highest image-based classification accuracy of 90.14% on Tobacco3482 without transfer learning from RVL-CDIP. Finally, our proposed model may serve as a powerful visual backbone for future multimodal approaches, by providing much richer visual features than existing counterparts.
Embedding of large but redundant data, such as images or text, in a hierarchy of lower-dimensional spaces is one of the key features of representation learning approaches, which nowadays provide state-of-the-art solutions to problems once believed hard or impossible to solve. In this work, in a plot twist with a strong meta aftertaste, we show how trained deep models are as redundant as the data they are optimized to process, and how it is therefore possible to use deep learning models to embed deep learning models. In particular, we show that it is possible to use representation learning to learn a fixed-size, low-dimensional embedding space of trained deep models and that such space can be explored by interpolation or optimization to attain ready-to-use models. We find that it is possible to learn an embedding space of multiple instances of the same architecture and of multiple architectures. We address image classification and neural representation of signals, showing how our embedding space can be learnt so as to capture the notions of performance and 3D shape, respectively. In the Multi-Architecture setting we also show how an embedding trained only on a subset of architectures can learn to generate already-trained instances of architectures it never sees instantiated at training time.
We examine how the saccade mechanism from biological vision can be used to make deep neural networks more efficient for classification and object detection problems. Our proposed approach is based on the ideas of attention-driven visual processing and saccades, miniature eye movements influenced by attention. We conduct experiments by analyzing: i) the robustness of different deep neural network (DNN) feature extractors to partially-sensed images for image classification and object detection, and ii) the utility of saccades in masking image patches for image classification and object tracking. Experiments with convolutional nets (ResNet-18) and transformer-based models (ViT, DETR, TransTrack) are conducted on several datasets (CIFAR-10, DAVSOD, MSCOCO, and MOT17). Our experiments show intelligent data reduction via learning to mimic human saccades when used in conjunction with state-of-the-art DNNs for classification, detection, and tracking tasks. We observed minimal drop in performance for the classification and detection tasks while only using about 30\% of the original sensor data. We discuss how the saccade mechanism can inform hardware design via ``in-pixel'' processing.
Deep neural networks are capable of learning powerful representations to tackle complex vision tasks but expose undesirable properties like the over-fitting issue. To this end, regularization techniques like image augmentation are necessary for deep neural networks to generalize well. Nevertheless, most prevalent image augmentation recipes confine themselves to off-the-shelf linear transformations like scale, flip, and colorjitter. Due to their hand-crafted property, these augmentations are insufficient to generate truly hard augmented examples. In this paper, we propose a novel perspective of augmentation to regularize the training process. Inspired by the recent success of applying masked image modeling to self-supervised learning, we adopt the self-supervised masked autoencoder to generate the distorted view of the input images. We show that utilizing such model-based nonlinear transformation as data augmentation can improve high-level recognition tasks. We term the proposed method as \textbf{M}ask-\textbf{R}econstruct \textbf{A}ugmentation (MRA). The extensive experiments on various image classification benchmarks verify the effectiveness of the proposed augmentation. Specifically, MRA consistently enhances the performance on supervised, semi-supervised as well as few-shot classification.
We propose an algorithm to explore the global optimization method, using SAT solvers, for training a neural net. Deep Neural Networks have achieved great feats in tasks like-image recognition, speech recognition, etc. Much of their success can be attributed to the gradient-based optimisation methods, which scale well to huge datasets while still giving solutions, better than any other existing methods. However, there exist learning problems like the parity function and the Fast Fourier Transform, where a neural network using gradient-based optimisation algorithm can not capture the underlying structure of the learning task properly. Thus, exploring global optimisation methods is of utmost interest as the gradient-based methods get stuck in local optima. In the experiments, we demonstrate the effectiveness of our algorithm against the ADAM optimiser in certain tasks like parity learning. However, in the case of image classification on the MNIST Dataset, the performance of our algorithm was less than satisfactory. We further discuss the role of the size of the training dataset and the hyper-parameter settings in keeping things scalable for a SAT solver.
Overparameterized neural networks enjoy great representation power on complex data, and more importantly yield sufficiently smooth output, which is crucial to their generalization and robustness. Most existing function approximation theories suggest that with sufficiently many parameters, neural networks can well approximate certain classes of functions in terms of the function value. The neural network themselves, however, can be highly nonsmooth. To bridge this gap, we take convolutional residual networks (ConvResNets) as an example, and prove that large ConvResNets can not only approximate a target function in terms of function value, but also exhibit sufficient first-order smoothness. Moreover, we extend our theory to approximating functions supported on a low-dimensional manifold. Our theory partially justifies the benefits of using deep and wide networks in practice. Numerical experiments on adversarial robust image classification are provided to support our theory.
The success of meta-learning on existing benchmarks is predicated on the assumption that the distribution of meta-training tasks covers meta-testing tasks. Frequent violation of the assumption in applications with either insufficient tasks or a very narrow meta-training task distribution leads to memorization or learner overfitting. Recent solutions have pursued augmentation of meta-training tasks, while it is still an open question to generate both correct and sufficiently imaginary tasks. In this paper, we seek an approach that up-samples meta-training tasks from the task representation via a task up-sampling network. Besides, the resulting approach named Adversarial Task Up-sampling (ATU) suffices to generate tasks that can maximally contribute to the latest meta-learner by maximizing an adversarial loss. On few-shot sine regression and image classification datasets, we empirically validate the marked improvement of ATU over state-of-the-art task augmentation strategies in the meta-testing performance and also the quality of up-sampled tasks.
We report our investigation on the use of the popular PlantVillage dataset for training deep learning based plant disease detection models. We trained a machine learning model using only 8 pixels from the PlantVillage image backgrounds. The model achieved 49.0% accuracy on the held-out test set, well above the random guessing accuracy of 2.6%. This result indicates that the PlantVillage dataset contains noise correlated with the labels and deep learning models can easily exploit this bias to make predictions. Possible approaches to alleviate this problem are discussed.
Despite advances in image classification methods, detecting the samples not belonging to the training classes is still a challenging problem. There has been a burst of interest in this subject recently, which is called Open-Set Recognition (OSR). In OSR, the goal is to achieve both the classification and detecting out-of-distribution (OOD) samples. Several ideas have been proposed to push the empirical result further through complicated techniques. We believe that such complication is indeed not necessary. To this end, we have shown that Maximum Softmax Probability (MSP), as the simplest baseline for OSR, applied on Vision Transformers (ViTs) as the base classifier that is trained with non-OOD augmentations can surprisingly outperform many recent methods. Non-OOD augmentations are the ones that do not alter the data distribution by much. Our results outperform state-of-the-art in CIFAR-10 datasets, and is also better than most of the current methods in SVHN and MNIST. We show that training augmentation has a significant effect on the performance of ViTs in the OSR tasks, and while they should produce significant diversity in the augmented samples, the generated sample OOD-ness must remain limited.
To build an artificial neural network like the biological intelligence system, recent works have unified numerous tasks into a generalist model, which can process various tasks with shared parameters and do not have any task-specific modules. While generalist models achieve promising results on various benchmarks, they have performance degradation on some tasks compared with task-specialized models. In this work, we find that interference among different tasks and modalities is the main factor to this phenomenon. To mitigate such interference, we introduce the Conditional Mixture-of-Experts (Conditional MoEs) to generalist models. Routing strategies under different levels of conditions are proposed to take both the training/inference cost and generalization ability into account. By incorporating the proposed Conditional MoEs, the recently proposed generalist model Uni-Perceiver can effectively mitigate the interference across tasks and modalities, and achieves state-of-the-art results on a series of downstream tasks via prompt tuning on 1% of downstream data. Moreover, the introduction of Conditional MoEs still holds the generalization ability of generalist models to conduct zero-shot inference on new tasks, e.g., video-text retrieval and video caption. Code and pre-trained generalist models shall be released.
Federated learning has been proposed as a privacy-preserving machine learning framework that enables multiple clients to collaborate without sharing raw data. However, client privacy protection is not guaranteed by design in this framework. Prior work has shown that the gradient sharing strategies in federated learning can be vulnerable to data reconstruction attacks. In practice, though, clients may not transmit raw gradients considering the high communication cost or due to privacy enhancement requirements. Empirical studies have demonstrated that gradient obfuscation, including intentional obfuscation via gradient noise injection and unintentional obfuscation via gradient compression, can provide more privacy protection against reconstruction attacks. In this work, we present a new data reconstruction attack framework targeting the image classification task in federated learning. We show that commonly adopted gradient postprocessing procedures, such as gradient quantization, gradient sparsification, and gradient perturbation, may give a false sense of security in federated learning. Contrary to prior studies, we argue that privacy enhancement should not be treated as a byproduct of gradient compression. Additionally, we design a new method under the proposed framework to reconstruct the image at the semantic level. We quantify the semantic privacy leakage and compare with conventional based on image similarity scores. Our comparisons challenge the image data leakage evaluation schemes in the literature. The results emphasize the importance of revisiting and redesigning the privacy protection mechanisms for client data in existing federated learning algorithms.
The size of vision models has grown exponentially over the last few years, especially after the emergence of Vision Transformer. This has motivated the development of parameter-efficient tuning methods, such as learning adapter layers or visual prompt tokens, which allow a tiny portion of model parameters to be trained whereas the vast majority obtained from pre-training are frozen. However, designing a proper tuning method is non-trivial: one might need to try out a lengthy list of design choices, not to mention that each downstream dataset often requires custom designs. In this paper, we view the existing parameter-efficient tuning methods as "prompt modules" and propose Neural prOmpt seArcH (NOAH), a novel approach that learns, for large vision models, the optimal design of prompt modules through a neural architecture search algorithm, specifically for each downstream dataset. By conducting extensive experiments on over 20 vision datasets, we demonstrate that NOAH (i) is superior to individual prompt modules, (ii) has a good few-shot learning ability, and (iii) is domain-generalizable. The code and models are available at https://github.com/Davidzhangyuanhan/NOAH.
Deep neural networks for image classification are well-known to be vulnerable to adversarial attacks. One such attack that has garnered recent attention is the adversarial backdoor attack, which has demonstrated the capability to perform targeted misclassification of specific examples. In particular, backdoor attacks attempt to force a model to learn spurious relations between backdoor trigger patterns and false labels. In response to this threat, numerous defensive measures have been proposed; however, defenses against backdoor attacks focus on backdoor pattern detection, which may be unreliable against novel or unexpected types of backdoor pattern designs. We introduce a novel re-contextualization of the adversarial setting, where the presence of an adversary implicitly admits the existence of multiple database contributors. Then, under the mild assumption of contributor awareness, it becomes possible to exploit this knowledge to defend against backdoor attacks by destroying the false label associations. We propose a contributor-aware universal defensive framework for learning in the presence of multiple, potentially adversarial data sources that utilizes semi-supervised ensembles and learning from crowds to filter the false labels produced by adversarial triggers. Importantly, this defensive strategy is agnostic to backdoor pattern design, as it functions without needing -- or even attempting -- to perform either adversary identification or backdoor pattern detection during either training or inference. Our empirical studies demonstrate the robustness of the proposed framework against adversarial backdoor attacks from multiple simultaneous adversaries.
Efficient neural network backbones for mobile devices are often optimized for metrics such as FLOPs or parameter count. However, these metrics may not correlate well with latency of the network when deployed on a mobile device. Therefore, we perform extensive analysis of different metrics by deploying several mobile-friendly networks on a mobile device. We identify and analyze architectural and optimization bottlenecks in recent efficient neural networks and provide ways to mitigate these bottlenecks. To this end, we design an efficient backbone MobileOne, with variants achieving an inference time under 1 ms on an iPhone12 with 75.9% top-1 accuracy on ImageNet. We show that MobileOne achieves state-of-the-art performance within the efficient architectures while being many times faster on mobile. Our best model obtains similar performance on ImageNet as MobileFormer while being 38x faster. Our model obtains 2.3% better top-1 accuracy on ImageNet than EfficientNet at similar latency. Furthermore, we show that our model generalizes to multiple tasks - image classification, object detection, and semantic segmentation with significant improvements in latency and accuracy as compared to existing efficient architectures when deployed on a mobile device. Code and models are available at https://github.com/apple/ml-mobileone
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.
Image classifiers often rely overly on peripheral attributes that have a strong correlation with the target class (i.e., dataset bias) when making predictions. Due to the dataset bias, the model correctly classifies data samples including bias attributes (i.e., bias-aligned samples) while failing to correctly predict those without bias attributes (i.e., bias-conflicting samples). Recently, a myriad of studies focus on mitigating such dataset bias, the task of which is referred to as debiasing. However, our comprehensive study indicates several issues need to be improved when conducting evaluation of debiasing in image classification. First, most of the previous studies do not specify how they select their hyper-parameters and model checkpoints (i.e., tuning criterion). Second, the debiasing studies until now evaluated their proposed methods on datasets with excessively high bias-severities, showing degraded performance on datasets with low bias severity. Third, the debiasing studies do not share consistent experimental settings (e.g., datasets and neural networks) which need to be standardized for fair comparisons. Based on such issues, this paper 1) proposes an evaluation metric `Align-Conflict (AC) score' for the tuning criterion, 2) includes experimental settings with low bias severity and shows that they are yet to be explored, and 3) unifies the standardized experimental settings to promote fair comparisons between debiasing methods. We believe that our findings and lessons inspire future researchers in debiasing to further push state-of-the-art performances with fair comparisons.
Brain inspired spiking neural networks (SNNs) have been successfully applied to many pattern recognition domains. The SNNs based deep structure have achieved considerable results in perceptual tasks, such as image classification, target detection. However, the application of deep SNNs in reinforcement learning (RL) tasks is still a problem to be explored. Although there have been previous studies on the combination of SNNs and RL, most of them focus on robotic control problems with shallow networks or using ANN-SNN conversion method to implement spiking deep Q Network (SDQN). In this work, we mathematically analyzed the problem of the disappearance of spiking signal features in SDQN and proposed a potential based layer normalization(pbLN) method to directly train spiking deep Q networks. Experiment shows that compared with state-of-art ANN-SNN conversion method and other SDQN works, the proposed pbLN spiking deep Q networks (PL-SDQN) achieved better performance on Atari game tasks.
The rapid on-site evaluation (ROSE) technique can significantly ac-celerate the diagnostic workflow of pancreatic cancer by immediately analyzing the fast-stained cytopathological images with on-site pathologists. Computer-aided diagnosis (CAD) using the deep learning method has the potential to solve the problem of insufficient pathology staffing. However, the cancerous patterns of ROSE images vary greatly between different samples, making the CAD task extremely challenging. Besides, due to different staining qualities and various types of acquisition devices, the ROSE images also have compli-cated perturbations in terms of color distribution, brightness, and contrast. To address these challenges, we proposed a novel multiple instance learning (MIL) approach using shuffle patches containing the instances, which adopts the patch-based learning strategy of Vision Transformers. With the re-grouped bags of shuffle instances and their bag-level soft labels, the approach utilizes a MIL head to make the model focus on the features from the pancreatic cancer cells, rather than that from various perturbations in ROSE images. Simultaneously, combined with a classification head, the model can effectively identify the gen-eral distributive patterns across different instances. The results demonstrate the significant improvements in the classification accuracy with more accurate at-tention regions, indicating that the diverse patterns of ROSE images are effec-tively extracted, and the complicated perturbations of ROSE images are signifi-cantly eliminated. It also suggests that the MIL with shuffle instances has great potential in the analysis of cytopathological images.
Given the current trend of increasing size and complexity of machine learning architectures, it has become of critical importance to identify new approaches to improve the computational efficiency of model training. In this context, we address the advantages of floating-point over fixed-point representation, and present an in-depth study on the use of 8-bit floating-point number formats for activations, weights, and gradients for both training and inference. We explore the effect of different bit-widths for exponents and significands and different exponent biases. The experimental results demonstrate that a suitable choice of these low-precision formats enables faster training and reduced power consumption without any degradation in accuracy for a range of deep learning models for image classification and language processing.