Continual learning in neural networks suffers from a phenomenon called catastrophic forgetting, in which a network quickly forgets what was learned in a previous task. The human brain, however, is able to continually learn new tasks and accumulate knowledge throughout life. Neuroscience findings suggest that continual learning success in the human brain is potentially associated with its modular structure and memory consolidation mechanisms. In this paper we propose a novel topological regularization that penalizes cycle structure in a neural network during training using principled theory from persistent homology and optimal transport. The penalty encourages the network to learn modular structure during training. The penalization is based on the closed-form expressions of the Wasserstein distance and barycenter for the topological features of a 1-skeleton representation for the network. Our topological continual learning method combines the proposed regularization with a tiny episodic memory to mitigate forgetting. We demonstrate that our method is effective in both shallow and deep network architectures for multiple image classification datasets.
Convolutional neural networks (CNNs) have achieved superhuman performance in multiple vision tasks, especially image classification. However, unlike humans, CNNs leverage spurious features, such as background information to make decisions. This tendency creates different problems in terms of robustness or weak generalization performance. Through our work, we introduce a contrastive learning-based approach (CLAD) to mitigate the background bias in CNNs. CLAD encourages semantic focus on object foregrounds and penalizes learning features from irrelavant backgrounds. Our method also introduces an efficient way of sampling negative samples. We achieve state-of-the-art results on the Background Challenge dataset, outperforming the previous benchmark with a margin of 4.1\%. Our paper shows how CLAD serves as a proof of concept for debiasing of spurious features, such as background and texture (in supplementary material).
Random masks define surprisingly effective sparse neural network models, as has been shown empirically. The resulting sparse networks can often compete with dense architectures and state-of-the-art lottery ticket pruning algorithms, even though they do not rely on computationally expensive prune-train iterations and can be drawn initially without significant computational overhead. We offer a theoretical explanation of how random masks can approximate arbitrary target networks if they are wider by a logarithmic factor in the inverse sparsity $1 / \log(1/\text{sparsity})$. This overparameterization factor is necessary at least for 3-layer random networks, which elucidates the observed degrading performance of random networks at higher sparsity. At moderate to high sparsity levels, however, our results imply that sparser networks are contained within random source networks so that any dense-to-sparse training scheme can be turned into a computationally more efficient sparse-to-sparse one by constraining the search to a fixed random mask. We demonstrate the feasibility of this approach in experiments for different pruning methods and propose particularly effective choices of initial layer-wise sparsity ratios of the random source network. As a special case, we show theoretically and experimentally that random source networks also contain strong lottery tickets.
This work proposes a new method to sequentially train deep neural networks on multiple tasks without suffering catastrophic forgetting, while endowing it with the capability to quickly adapt to unseen tasks. Starting from existing work on network masking (Wortsman et al., 2020), we show that simply learning a linear combination of a small number of task-specific supermasks (impressions) on a randomly initialized backbone network is sufficient to both retain accuracy on previously learned tasks, as well as achieve high accuracy on unseen tasks. In contrast to previous methods, we do not require to generate dedicated masks or contexts for each new task, instead leveraging transfer learning to keep per-task parameter overhead small. Our work illustrates the power of linearly combining individual impressions, each of which fares poorly in isolation, to achieve performance comparable to a dedicated mask. Moreover, even repeated impressions from the same task (homogeneous masks), when combined, can approach the performance of heterogeneous combinations if sufficiently many impressions are used. Our approach scales more efficiently than existing methods, often requiring orders of magnitude fewer parameters and can function without modification even when task identity is missing. In addition, in the setting where task labels are not given at inference, our algorithm gives an often favorable alternative to the one-shot procedure used by Wortsman et al., 2020. We evaluate our method on a number of well-known image classification datasets and network architectures.
This paper presents MOAT, a family of neural networks that build on top of MObile convolution (i.e., inverted residual blocks) and ATtention. Unlike the current works that stack separate mobile convolution and transformer blocks, we effectively merge them into a MOAT block. Starting with a standard Transformer block, we replace its multi-layer perceptron with a mobile convolution block, and further reorder it before the self-attention operation. The mobile convolution block not only enhances the network representation capacity, but also produces better downsampled features. Our conceptually simple MOAT networks are surprisingly effective, achieving 89.1% / 81.5% top-1 accuracy on ImageNet-1K / ImageNet-1K-V2 with ImageNet22K pretraining. Additionally, MOAT can be seamlessly applied to downstream tasks that require large resolution inputs by simply converting the global attention to window attention. Thanks to the mobile convolution that effectively exchanges local information between pixels (and thus cross-windows), MOAT does not need the extra window-shifting mechanism. As a result, on COCO object detection, MOAT achieves 59.2% box AP with 227M model parameters (single-scale inference, and hard NMS), and on ADE20K semantic segmentation, MOAT attains 57.6% mIoU with 496M model parameters (single-scale inference). Finally, the tiny-MOAT family, obtained by simply reducing the channel sizes, also surprisingly outperforms several mobile-specific transformer-based models on ImageNet. The tiny-MOAT family is also benchmarked on downstream tasks, serving as a baseline for the community. We hope our simple yet effective MOAT will inspire more seamless integration of convolution and self-attention. Code is publicly available.
Classifying pill categories from real-world images is crucial for various smart healthcare applications. Although existing approaches in image classification might achieve a good performance on fixed pill categories, they fail to handle novel instances of pill categories that are frequently presented to the learning algorithm. To this end, a trivial solution is to train the model with novel classes. However, this may result in a phenomenon known as catastrophic forgetting, in which the system forgets what it learned in previous classes. In this paper, we address this challenge by introducing the class incremental learning (CIL) ability to traditional pill image classification systems. Specifically, we propose a novel incremental multi-stream intermediate fusion framework enabling incorporation of an additional guidance information stream that best matches the domain of the problem into various state-of-the-art CIL methods. From this framework, we consider color-specific information of pill images as a guidance stream and devise an approach, namely "Color Guidance with Multi-stream intermediate fusion"(CG-IMIF) for solving CIL pill image classification task. We conduct comprehensive experiments on real-world incremental pill image classification dataset, namely VAIPE-PCIL, and find that the CG-IMIF consistently outperforms several state-of-the-art methods by a large margin in different task settings. Our code, data, and trained model are available at https://github.com/vinuni-vishc/CG-IMIF.
To deploy deep learning algorithms on resource-limited scenarios, an emerging device-resistive random access memory (ReRAM) has been regarded as promising via analog computing. However, the practicability of ReRAM is primarily limited due to the weight drifting of ReRAM neural networks due to multi-factor reasons, including manufacturing, thermal noises, and etc. In this paper, we propose a novel Bayesian optimization method for fault tolerant neural network architecture (BayesFT). For neural architecture search space design, instead of conducting neural architecture search on the whole feasible neural architecture search space, we first systematically explore the weight drifting tolerance of different neural network components, such as dropout, normalization, number of layers, and activation functions in which dropout is found to be able to improve the neural network robustness to weight drifting. Based on our analysis, we propose an efficient search space by only searching for dropout rates for each layer. Then, we use Bayesian optimization to search for the optimal neural architecture robust to weight drifting. Empirical experiments demonstrate that our algorithmic framework has outperformed the state-of-the-art methods by up to 10 times on various tasks, such as image classification and object detection.
Given the detrimental effect of spectral variations in a hyperspectral image (HSI), this article investigates to recover its discriminative representation to improve the classification performance. We propose a new method, namely local low-rank approximation with superpixel-guided locality preserving graph (LLRA-SLPG), which can reduce the spectral variations and preserve the local manifold structure of an HSI. Specifically, the LLRA-SLPG method first clusters pixels of an HSI into several groups (i.e., superpixels). By taking advantage of the local manifold structure, a Laplacian graph is constructed from the superpixels to ensure that a typical pixel should be similar to its neighbors within the same superpixel. The LLRA-SLPG model can increase the compactness of pixels belonging to the same class by reducing spectral variations while promoting local consistency via the Laplacian graph. The objective function of the LLRA-SLPG model can be solved efficiently in an iterative manner. Experimental results on four benchmark datasets validate the superiority of the LLRA-SLPG model over state-of-the-art methods, particularly in cases where only extremely few training pixels are available.
Real-world deployment of machine learning models is challenging because data evolves over time. While no model can work when data evolves in an arbitrary fashion, if there is some pattern to these changes, we might be able to design methods to address it. This paper addresses situations when data evolves gradually. We introduce a time-varying propensity score that can detect gradual shifts in the distribution of data which allows us to selectively sample past data to update the model -- not just similar data from the past like that of a standard propensity score but also data that evolved in a similar fashion in the past. The time-varying propensity score is quite general: we demonstrate different ways of implementing it and evaluate it on a variety of problems ranging from supervised learning (e.g., image classification problems) where data undergoes a sequence of gradual shifts, to reinforcement learning tasks (e.g., robotic manipulation and continuous control) where data shifts as the policy or the task changes.
Training deep learning models on medical datasets that perform well for all classes is a challenging task. It is often the case that a suboptimal performance is obtained on some classes due to the natural class imbalance issue that comes with medical data. An effective way to tackle this problem is by using targeted active learning, where we iteratively add data points to the training data that belong to the rare classes. However, existing active learning methods are ineffective in targeting rare classes in medical datasets. In this work, we propose Clinical (targeted aCtive Learning for ImbalaNced medICal imAge cLassification) a framework that uses submodular mutual information functions as acquisition functions to mine critical data points from rare classes. We apply our framework to a wide-array of medical imaging datasets on a variety of real-world class imbalance scenarios - namely, binary imbalance and long-tail imbalance. We show that Clinical outperforms the state-of-the-art active learning methods by acquiring a diverse set of data points that belong to the rare classes.
Most classifiers rely on discriminative boundaries that separate instances of each class from everything else. We argue that discriminative boundaries are counter-intuitive as they define semantics by what-they-are-not; and should be replaced by generative classifiers which define semantics by what-they-are. Unfortunately, generative classifiers are significantly less accurate. This may be caused by the tendency of generative models to focus on easy to model semantic generative factors and ignore non-semantic factors that are important but difficult to model. We propose a new generative model in which semantic factors are accommodated by shell theory's hierarchical generative process and non-semantic factors by an instance specific noise term. We use the model to develop a classification scheme which suppresses the impact of noise while preserving semantic cues. The result is a surprisingly accurate generative classifier, that takes the form of a modified nearest-neighbor algorithm; we term it distance classification. Unlike discriminative classifiers, a distance classifier: defines semantics by what-they-are; is amenable to incremental updates; and scales well with the number of classes.
Federated Learning (FL) framework enables training over distributed datasets while keeping the data local. However, it is difficult to customize a model fitting for all unknown local data. A pre-determined model is most likely to lead to slow convergence or low accuracy, especially when the distributed data is non-i.i.d.. To resolve the issue, we propose a model searching method in the federated learning scenario, and the method automatically searches a model structure fitting for the unseen local data. We novelly design a reinforcement learning-based framework that samples and distributes sub-models to the participants and updates its model selection policy by maximizing the reward. In practice, the model search algorithm takes a long time to converge, and hence we adaptively assign sub-models to participants according to the transmission condition. We further propose delay-compensated synchronization to mitigate loss over late updates to facilitate convergence. Extensive experiments show that our federated model search algorithm produces highly accurate models efficiently, particularly on non-i.i.d. data.
While deep neural networks (DNNs) have led to a paradigm shift, its exorbitant computational requirement has always been a roadblock in its deployment to the edge, such as wearable devices and smartphones. Hence a hybrid edge-cloud computational framework is proposed to transfer part of the computation to the cloud, by naively partitioning the DNN operations under the constant network condition assumption. However, real-world network state varies greatly depending on the context, and DNN partitioning only has limited strategy space. In this paper, we explore the structural flexibility of DNN to fit the edge model to varying network contexts and different deployment platforms. Specifically, we designed a reinforcement learning-based decision engine to search for model transformation strategies in response to a combined objective of model accuracy and computation latency. The engine generates a context-aware model tree so that the DNN can decide the model branch to switch to at runtime. By the emulation and field experimental results, our approach enjoys a 30% − 50% latency reduction while retaining the model accuracy.
We present a dynamic model in which the weights are conditioned on an input sample x and are learned to match those that would be obtained by finetuning a base model on x and its label y. This mapping between an input sample and network weights is approximated by a denoising diffusion model. The diffusion model we employ focuses on modifying a single layer of the base model and is conditioned on the input, activations, and output of this layer. Since the diffusion model is stochastic in nature, multiple initializations generate different networks, forming an ensemble, which leads to further improvements. Our experiments demonstrate the wide applicability of the method for image classification, 3D reconstruction, tabular data, speech separation, and natural language processing. Our code is available at https://github.com/ShaharLutatiPersonal/OCD
Text-to-image generation models have recently attracted unprecedented attention as they unlatch imaginative applications in all areas of life. However, developing such models requires huge amounts of data that might contain privacy-sensitive information, e.g., face identity. While privacy risks have been extensively demonstrated in the image classification and GAN generation domains, privacy risks in the text-to-image generation domain are largely unexplored. In this paper, we perform the first privacy analysis of text-to-image generation models through the lens of membership inference. Specifically, we propose three key intuitions about membership information and design four attack methodologies accordingly. We conduct comprehensive evaluations on two mainstream text-to-image generation models including sequence-to-sequence modeling and diffusion-based modeling. The empirical results show that all of the proposed attacks can achieve significant performance, in some cases even close to an accuracy of 1, and thus the corresponding risk is much more severe than that shown by existing membership inference attacks. We further conduct an extensive ablation study to analyze the factors that may affect the attack performance, which can guide developers and researchers to be alert to vulnerabilities in text-to-image generation models. All these findings indicate that our proposed attacks pose a realistic privacy threat to the text-to-image generation models.
In this paper, we provide an in-depth study of Stochastic Backpropagation (SBP) when training deep neural networks for standard image classification and object detection tasks. During backward propagation, SBP calculates the gradients by only using a subset of feature maps to save the GPU memory and computational cost. We interpret SBP as an efficient way to implement stochastic gradient decent by performing backpropagation dropout, which leads to considerable memory saving and training process speedup, with a minimal impact on the overall model accuracy. We offer some good practices to apply SBP in training image recognition models, which can be adopted in learning a wide range of deep neural networks. Experiments on image classification and object detection show that SBP can save up to 40% of GPU memory with less than 1% accuracy degradation.
Convolution blocks serve as local feature extractors and are the key to success of the neural networks. To make local semantic feature embedding rather explicit, we reformulate convolution blocks as feature selection according to the best matching kernel. In this manner, we show that typical ResNet blocks indeed perform local feature embedding via template matching once batch normalization (BN) followed by a rectified linear unit (ReLU) is interpreted as arg-max optimizer. Following this perspective, we tailor a residual block that explicitly forces semantically meaningful local feature embedding through using label information. Specifically, we assign a feature vector to each local region according to the classes that the corresponding region matches. We evaluate our method on three popular benchmark datasets with several architectures for image classification and consistently show that our approach substantially improves the performance of the baseline architectures.
Rapid economic growth gives rise to the urgent demand for a more efficient waste recycling system. This work thereby developed an innovative recycling bin that automatically separates urban waste to increase the recycling rate. We collected 1800 recycling waste images and combined them with an existing public dataset to train classification models for two embedded systems, Jetson Nano and K210, targeting different markets. The model reached an accuracy of 95.98% on Jetson Nano and 96.64% on K210. A bin program was designed to collect feedback from users. On Jetson Nano, the overall power consumption of the application was reduced by 30% from the previous work to 4.7 W, while the second system, K210, only needed 0.89 W of power to operate. In summary, our work demonstrated a fully functional prototype of an energy-saving, high-accuracy smart recycling bin, which can be commercialized in the future to improve urban waste recycling.
Few-shot learning is an important, but challenging problem of machine learning aimed at learning from only fewer labeled training examples. It has become an active area of research due to deep learning requiring huge amounts of labeled dataset, which is not feasible in the real world. Learning from a few examples is also an important attempt towards learning like humans. Few-shot learning has proven a very good promise in different areas of machine learning applications, particularly in image classification. As it is a recent technique, most researchers focus on understanding and solving the issues related to its concept by focusing only on common image datasets like Mini-ImageNet and Omniglot. Few-shot learning also opens an opportunity to address low resource languages like Amharic. In this study, offline handwritten Amharic character recognition using few-shot learning is addressed. Particularly, prototypical networks, the popular and simpler type of few-shot learning, is implemented as a baseline. Using the opportunities explored in the nature of Amharic alphabet having row-wise and column-wise similarities, a novel way of augmenting the training episodes is proposed. The experimental results show that the proposed method outperformed the baseline method. This study has implemented few-shot learning for Amharic characters for the first time. More importantly, the findings of the study open new ways of examining the influence of training episodes in few-shot learning, which is one of the important issues that needs exploration. The datasets used for this study are collected from native Amharic language writers using an Android App developed as a part of this study.
Vision transformers have recently achieved competitive results across various vision tasks but still suffer from heavy computation costs when processing a large number of tokens. Many advanced approaches have been developed to reduce the total number of tokens in large-scale vision transformers, especially for image classification tasks. Typically, they select a small group of essential tokens according to their relevance with the class token, then fine-tune the weights of the vision transformer. Such fine-tuning is less practical for dense prediction due to the much heavier computation and GPU memory cost than image classification. In this paper, we focus on a more challenging problem, i.e., accelerating large-scale vision transformers for dense prediction without any additional re-training or fine-tuning. In response to the fact that high-resolution representations are necessary for dense prediction, we present two non-parametric operators, a token clustering layer to decrease the number of tokens and a token reconstruction layer to increase the number of tokens. The following steps are performed to achieve this: (i) we use the token clustering layer to cluster the neighboring tokens together, resulting in low-resolution representations that maintain the spatial structures; (ii) we apply the following transformer layers only to these low-resolution representations or clustered tokens; and (iii) we use the token reconstruction layer to re-create the high-resolution representations from the refined low-resolution representations. The results obtained by our method are promising on five dense prediction tasks, including object detection, semantic segmentation, panoptic segmentation, instance segmentation, and depth estimation.