Multi-label image classification, which can be categorized into label-dependency and region-based methods, is a challenging problem due to the complex underlying object layouts. Although region-based methods are less likely to encounter issues with model generalizability than label-dependency methods, they often generate hundreds of meaningless or noisy proposals with non-discriminative information, and the contextual dependency among the localized regions is often ignored or over-simplified. This paper builds a unified framework to perform effective noisy-proposal suppression and to interact between global and local features for robust feature learning. Specifically, we propose category-aware weak supervision to concentrate on non-existent categories so as to provide deterministic information for local feature learning, restricting the local branch to focus on more high-quality regions of interest. Moreover, we develop a cross-granularity attention module to explore the complementary information between global and local features, which can build the high-order feature correlation containing not only global-to-local, but also local-to-local relations. Both advantages guarantee a boost in the performance of the whole network. Extensive experiments on two large-scale datasets (MS-COCO and VOC 2007) demonstrate that our framework achieves superior performance over state-of-the-art methods.
In this paper, we propose a simple and general approach to augment regular convolution operator by injecting extra group-wise transformation during training and recover it during inference. Extra transformation is carefully selected to ensure it can be merged with regular convolution in each group and will not change the topological structure of regular convolution during inference. Compared with regular convolution operator, our approach (AugConv) can introduce larger learning capacity to improve model performance during training but will not increase extra computational overhead for model deployment. Based on ResNet, we utilize AugConv to build convolutional neural networks named AugResNet. Result on image classification dataset Cifar-10 shows that AugResNet outperforms its baseline in terms of model performance.
Label noise and class imbalance commonly coexist in real-world data. Previous works for robust learning, however, usually address either one type of the data biases and underperform when facing them both. To mitigate this gap, this work presents a novel meta-learning based dynamic loss that automatically adjusts the objective functions with the training process to robustly learn a classifier from long-tailed noisy data. Concretely, our dynamic loss comprises a label corrector and a margin generator, which respectively correct noisy labels and generate additive per-class classification margins by perceiving the underlying data distribution as well as the learning state of the classifier. Equipped with a new hierarchical sampling strategy that enriches a small amount of unbiased metadata with diverse and hard samples, the two components in the dynamic loss are optimized jointly through meta-learning and cultivate the classifier to well adapt to clean and balanced test data. Extensive experiments show our method achieves state-of-the-art accuracy on multiple real-world and synthetic datasets with various types of data biases, including CIFAR-10/100, Animal-10N, ImageNet-LT, and Webvision. Code will soon be publicly available.
This work presents two astonishing findings on neural networks learned for large-scale image classification. 1) Given a well-trained model, the logits predicted for some category can be directly obtained by linearly combining the predictions of a few other categories, which we call \textbf{neural dependency}. 2) Neural dependencies exist not only within a single model, but even between two independently learned models, regardless of their architectures. Towards a theoretical analysis of such phenomena, we demonstrate that identifying neural dependencies is equivalent to solving the Covariance Lasso (CovLasso) regression problem proposed in this paper. Through investigating the properties of the problem solution, we confirm that neural dependency is guaranteed by a redundant logit covariance matrix, which condition is easily met given massive categories, and that neural dependency is highly sparse, implying that one category correlates to only a few others. We further empirically show the potential of neural dependencies in understanding internal data correlations, generalizing models to unseen categories, and improving model robustness with a dependency-derived regularizer. Code for this work will be made publicly available.
Implicit Neural Representations (INRs) are powerful to parameterize continuous signals in computer vision. However, almost all INRs methods are limited to low-level tasks, e.g., image/video compression, super-resolution, and image generation. The questions on how to explore INRs to high-level tasks and deep networks are still under-explored. Existing INRs methods suffer from two problems: 1) narrow theoretical definitions of INRs are inapplicable to high-level tasks; 2) lack of representation capabilities to deep networks. Motivated by the above facts, we reformulate the definitions of INRs from a novel perspective and propose an innovative Implicit Neural Representation Network (INRN), which is the first study of INRs to tackle both low-level and high-level tasks. Specifically, we present three key designs for basic blocks in INRN along with two different stacking ways and corresponding loss functions. Extensive experiments with analysis on both low-level tasks (image fitting) and high-level vision tasks (image classification, object detection, instance segmentation) demonstrate the effectiveness of the proposed method.
Multi-spectral imagery is invaluable for remote sensing due to different spectral signatures exhibited by materials that often appear identical in greyscale and RGB imagery. Paired with modern deep learning methods, this modality has great potential utility in a variety of remote sensing applications, such as humanitarian assistance and disaster recovery efforts. State-of-the-art deep learning methods have greatly benefited from large-scale annotations like in ImageNet, but existing MSI image datasets lack annotations at a similar scale. As an alternative to transfer learning on such data with few annotations, we apply complex-valued co-domain symmetric models to classify real-valued MSI images. Our experiments on 8-band xView data show that our ultra-lean model trained on xView from scratch without data augmentations can outperform ResNet with data augmentation and modified transfer learning on xView. Our work is the first to demonstrate the value of complex-valued deep learning on real-valued MSI data.
Hierarchical classification (HC) assigns each object with multiple labels organized into a hierarchical structure. The existing deep learning based HC methods usually predict an instance starting from the root node until a leaf node is reached. However, in the real world, images interfered by noise, occlusion, blur, or low resolution may not provide sufficient information for the classification at subordinate levels. To address this issue, we propose a novel semantic guided level-category hybrid prediction network (SGLCHPN) that can jointly perform the level and category prediction in an end-to-end manner. SGLCHPN comprises two modules: a visual transformer that extracts feature vectors from the input images, and a semantic guided cross-attention module that uses categories word embeddings as queries to guide learning category-specific representations. In order to evaluate the proposed method, we construct two new datasets in which images are at a broad range of quality and thus are labeled to different levels (depths) in the hierarchy according to their individual quality. Experimental results demonstrate the effectiveness of our proposed HC method.
Explaining black-box Artificial Intelligence (AI) models is a cornerstone for trustworthy AI and a prerequisite for its use in safety critical applications such that AI models can reliably assist humans in critical decisions. However, instead of trying to explain our models post-hoc, we need models which are interpretable-by-design built on a reasoning process similar to humans that exploits meaningful high-level concepts such as shapes, texture or object parts. Learning such concepts is often hindered by its need for explicit specification and annotation up front. Instead, prototype-based learning approaches such as ProtoPNet claim to discover visually meaningful prototypes in an unsupervised way. In this work, we propose a set of properties that those prototypes have to fulfill to enable human analysis, e.g. as part of a reliable model assessment case, and analyse such existing methods in the light of these properties. Given a 'Guess who?' game, we find that these prototypes still have a long way ahead towards definite explanations. We quantitatively validate our findings by conducting a user study indicating that many of the learnt prototypes are not considered useful towards human understanding. We discuss about the missing links in the existing methods and present a potential real-world application motivating the need to progress towards truly human-interpretable prototypes.
We propose a routing algorithm that takes a sequence of vectors and computes a new sequence with specified length and vector size. Each output vector maximizes "bang per bit," the difference between a net benefit to use and net cost to ignore data, by better predicting the input vectors. We describe output vectors as geometric objects, as latent variables that assign credit, as query states in a model of associative memory, and as agents in a model of a Society of Mind. We implement the algorithm with optimizations that reduce parameter count, computation, and memory use by orders of magnitude, enabling us to route sequences of greater length than previously possible. We evaluate our implementation on natural language and visual classification tasks, obtaining competitive or state-of-the-art accuracy and end-to-end credit assignments that are interpretable.
We study the generalization behavior of transfer learning of deep neural networks (DNNs). We adopt the overparameterization perspective -- featuring interpolation of the training data (i.e., approximately zero train error) and the double descent phenomenon -- to explain the delicate effect of the transfer learning setting on generalization performance. We study how the generalization behavior of transfer learning is affected by the dataset size in the source and target tasks, the number of transferred layers that are kept frozen in the target DNN training, and the similarity between the source and target tasks. We show that the test error evolution during the target DNN training has a more significant double descent effect when the target training dataset is sufficiently large. In addition, a larger source training dataset can yield a slower target DNN training. Moreover, we demonstrate that the number of frozen layers can determine whether the transfer learning is effectively underparameterized or overparameterized and, in turn, this may induce a freezing-wise double descent phenomenon that determines the relative success or failure of learning. Also, we show that the double descent phenomenon may make a transfer from a less related source task better than a transfer from a more related source task. We establish our results using image classification experiments with the ResNet, DenseNet and the vision transformer (ViT) architectures.
Implementing Decentralized Gradient Descent (DGD) in wireless systems is challenging due to noise, fading, and limited bandwidth, necessitating topology awareness, transmission scheduling, and the acquisition of channel state information (CSI) to mitigate interference and maintain reliable communications. These operations may result in substantial signaling overhead and scalability challenges in large networks lacking central coordination. This paper introduces a scalable DGD algorithm that eliminates the need for scheduling, topology information, or CSI (both average and instantaneous). At its core is a Non-Coherent Over-The-Air (NCOTA) consensus scheme that exploits a noisy energy superposition property of wireless channels. Nodes encode their local optimization signals into energy levels within an OFDM frame and transmit simultaneously, without coordination. The key insight is that the received energy equals, on average, the sum of the energies of the transmitted signals, scaled by their respective average channel gains, akin to a consensus step. This property enables unbiased consensus estimation, utilizing average channel gains as mixing weights, thereby removing the need for their explicit design or for CSI. Introducing a consensus stepsize mitigates consensus estimation errors due to energy fluctuations around their expected values. For strongly-convex problems, it is shown that the expected squared distance between the local and globally optimum models vanishes at a rate of O(1/sqrt{k}) after k iterations, with suitable decreasing learning and consensus stepsizes. Extensions accommodate a broad class of fading models and frequency-selective channels. Numerical experiments on image classification demonstrate faster convergence in terms of running time compared to state-of-the-art schemes, especially in dense network scenarios.
Deep learning methods have achieved great success in solving computer vision tasks, and they have been widely utilized in artificially intelligent systems for image processing, analysis, and understanding. However, deep neural networks have been shown to be vulnerable to adversarial perturbations in input data. The security issues of deep neural networks have thus come to the fore. It is imperative to study the adversarial robustness of deep vision algorithms comprehensively. This talk focuses on the adversarial robustness of image classification models and image denoisers. We will discuss the robustness of deep vision algorithms from three perspectives: 1) robustness evaluation (we propose the ObsAtk to evaluate the robustness of denoisers), 2) robustness improvement (HAT, TisODE, and CIFS are developed to robustify vision models), and 3) the connection between adversarial robustness and generalization capability to new domains (we find that adversarially robust denoisers can deal with unseen types of real-world noise).
Concept Bottleneck Models (CBM) are inherently interpretable models that factor model decisions into human-readable concepts. They allow people to easily understand why a model is failing, a critical feature for high-stakes applications. CBMs require manually specified concepts and often under-perform their black box counterparts, preventing their broad adoption. We address these shortcomings and are first to show how to construct high-performance CBMs without manual specification of similar accuracy to black box models. Our approach, Language Guided Bottlenecks (LaBo), leverages a language model, GPT-3, to define a large space of possible bottlenecks. Given a problem domain, LaBo uses GPT-3 to produce factual sentences about categories to form candidate concepts. LaBo efficiently searches possible bottlenecks through a novel submodular utility that promotes the selection of discriminative and diverse information. Ultimately, GPT-3's sentential concepts can be aligned to images using CLIP, to form a bottleneck layer. Experiments demonstrate that LaBo is a highly effective prior for concepts important to visual recognition. In the evaluation with 11 diverse datasets, LaBo bottlenecks excel at few-shot classification: they are 11.7% more accurate than black box linear probes at 1 shot and comparable with more data. Overall, LaBo demonstrates that inherently interpretable models can be widely applied at similar, or better, performance than black box approaches.
Recently, vision architectures based exclusively on multi-layer perceptrons (MLPs) have gained much attention in the computer vision community. MLP-like models achieve competitive performance on a single 2D image classification with less inductive bias without hand-crafted convolution layers. In this work, we explore the effectiveness of MLP-based architecture for the view-based 3D object recognition task. We present an MLP-based architecture termed as Round-Roll MLP (R$^2$-MLP). It extends the spatial-shift MLP backbone by considering the communications between patches from different views. R$^2$-MLP rolls part of the channels along the view dimension and promotes information exchange between neighboring views. We benchmark MLP results on ModelNet10 and ModelNet40 datasets with ablations in various aspects. The experimental results show that, with a conceptually simple structure, our R$^2$-MLP achieves competitive performance compared with existing state-of-the-art methods.
Small on-device models have been successfully trained with user-level differential privacy (DP) for next word prediction and image classification tasks in the past. However, existing methods can fail when directly applied to learn embedding models using supervised training data with a large class space. To achieve user-level DP for large image-to-embedding feature extractors, we propose DP-FedEmb, a variant of federated learning algorithms with per-user sensitivity control and noise addition, to train from user-partitioned data centralized in the datacenter. DP-FedEmb combines virtual clients, partial aggregation, private local fine-tuning, and public pretraining to achieve strong privacy utility trade-offs. We apply DP-FedEmb to train image embedding models for faces, landmarks and natural species, and demonstrate its superior utility under same privacy budget on benchmark datasets DigiFace, EMNIST, GLD and iNaturalist. We further illustrate it is possible to achieve strong user-level DP guarantees of $\epsilon<4$ while controlling the utility drop within 5%, when millions of users can participate in training.
Annotating datasets for object detection is an expensive and time-consuming endeavor. To minimize this burden, active learning (AL) techniques are employed to select the most informative samples for annotation within a constrained "annotation budget". Traditional AL strategies typically rely on model uncertainty or sample diversity for query sampling, while more advanced methods have focused on developing AL-specific object detector architectures to enhance performance. However, these specialized approaches are not readily adaptable to different object detectors due to the significant engineering effort required for integration. To overcome this challenge, we introduce Plug and Play Active Learning (PPAL), a simple and effective AL strategy for object detection. PPAL is a two-stage method comprising uncertainty-based and diversity-based sampling phases. In the first stage, our Difficulty Calibrated Uncertainty Sampling leverage a category-wise difficulty coefficient that combines both classification and localisation difficulties to re-weight instance uncertainties, from which we sample a candidate pool for the subsequent diversity-based sampling. In the second stage, we propose Category Conditioned Matching Similarity to better compute the similarities of multi-instance images as ensembles of their instance similarities, which is used by the k-Means++ algorithm to sample the final AL queries. PPAL makes no change to model architectures or detector training pipelines; hence it can be easily generalized to different object detectors. We benchmark PPAL on the MS-COCO and Pascal VOC datasets using different detector architectures and show that our method outperforms prior work by a large margin. Code is available at https://github.com/ChenhongyiYang/PPAL
Unsupervised domain adaptation (UDA) aims to estimate a transferable model for unlabeled target domains by exploiting labeled source data. Optimal Transport (OT) based methods have recently been proven to be a promising solution for UDA with a solid theoretical foundation and competitive performance. However, most of these methods solely focus on domain-level OT alignment by leveraging the geometry of domains for domain-invariant features based on the global embeddings of images. However, global representations of images may destroy image structure, leading to the loss of local details that offer category-discriminative information. This study proposes an end-to-end Deep Hierarchical Optimal Transport method (DeepHOT), which aims to learn both domain-invariant and category-discriminative representations by mining hierarchical structural relations among domains. The main idea is to incorporate a domain-level OT and image-level OT into a unified OT framework, hierarchical optimal transport, to model the underlying geometry in both domain space and image space. In DeepHOT framework, an image-level OT serves as the ground distance metric for the domain-level OT, leading to the hierarchical structural distance. Compared with the ground distance of the conventional domain-level OT, the image-level OT captures structural associations among local regions of images that are beneficial to classification. In this way, DeepHOT, a unified OT framework, not only aligns domains by domain-level OT, but also enhances the discriminative power through image-level OT. Moreover, to overcome the limitation of high computational complexity, we propose a robust and efficient implementation of DeepHOT by approximating origin OT with sliced Wasserstein distance in image-level OT and accomplishing the mini-batch unbalanced domain-level OT.
Optimizing neural networks with noisy labels is a challenging task, especially if the label set contains real-world noise. Networks tend to generalize to reasonable patterns in the early training stages and overfit to specific details of noisy samples in the latter ones. We introduce Blind Knowledge Distillation - a novel teacher-student approach for learning with noisy labels by masking the ground truth related teacher output to filter out potentially corrupted knowledge and to estimate the tipping point from generalizing to overfitting. Based on this, we enable the estimation of noise in the training data with Otsus algorithm. With this estimation, we train the network with a modified weighted cross-entropy loss function. We show in our experiments that Blind Knowledge Distillation detects overfitting effectively during training and improves the detection of clean and noisy labels on the recently published CIFAR-N dataset. Code is available at GitHub.
Federated learning (FL for simplification) is a distributed machine learning technique that utilizes global servers and collaborative clients to achieve privacy-preserving global model training without direct data sharing. However, heterogeneous data problem, as one of FL's main problems, makes it difficult for the global model to perform effectively on each client's local data. Thus, personalized federated learning (PFL for simplification) aims to improve the performance of the model on local data as much as possible. Bayesian learning, where the parameters of the model are seen as random variables with a prior assumption, is a feasible solution to the heterogeneous data problem due to the tendency that the more local data the model use, the more it focuses on the local data, otherwise focuses on the prior. When Bayesian learning is applied to PFL, the global model provides global knowledge as a prior to the local training process. In this paper, we employ Bayesian learning to model PFL by assuming a prior in the scaled exponential family, and therefore propose pFedBreD, a framework to solve the problem we model using Bregman divergence regularization. Empirically, our experiments show that, under the prior assumption of the spherical Gaussian and the first order strategy of mean selection, our proposal significantly outcompetes other PFL algorithms on multiple public benchmarks.
Deep learning usually relies on training large-scale data samples to achieve better performance. However, over-fitting based on training data always remains a problem. Scholars have proposed various strategies, such as feature dropping and feature mixing, to improve the generalization continuously. For the same purpose, we subversively propose a novel training method, Feature Weaken, which can be regarded as a data augmentation method. Feature Weaken constructs the vicinal data distribution with the same cosine similarity for model training by weakening features of the original samples. In especially, Feature Weaken changes the spatial distribution of samples, adjusts sample boundaries, and reduces the gradient optimization value of back-propagation. This work can not only improve the classification performance and generalization of the model, but also stabilize the model training and accelerate the model convergence. We conduct extensive experiments on classical deep convolution neural models with five common image classification datasets and the Bert model with four common text classification datasets. Compared with the classical models or the generalization improvement methods, such as Dropout, Mixup, Cutout, and CutMix, Feature Weaken shows good compatibility and performance. We also use adversarial samples to perform the robustness experiments, and the results show that Feature Weaken is effective in improving the robustness of the model.