In neural networks, continual learning results in gradient interference among sequential tasks, leading to catastrophic forgetting of old tasks while learning new ones. This issue is addressed in recent methods by storing the important gradient spaces for old tasks and updating the model orthogonally during new tasks. However, such restrictive orthogonal gradient updates hamper the learning capability of the new tasks resulting in sub-optimal performance. To improve new learning while minimizing forgetting, in this paper we propose a Scaled Gradient Projection (SGP) method, where we combine the orthogonal gradient projections with scaled gradient steps along the important gradient spaces for the past tasks. The degree of gradient scaling along these spaces depends on the importance of the bases spanning them. We propose an efficient method for computing and accumulating importance of these bases using the singular value decomposition of the input representations for each task. We conduct extensive experiments ranging from continual image classification to reinforcement learning tasks and report better performance with less training overhead than the state-of-the-art approaches.
Hyperspectral image (HSI) classification is used in multiple domains like precision agriculture, mineral exploration, remote sensing, and others. Conventionally, couvolutional neural networks (CNNs) were used in HSI classification, however they have limitations in exploiting spectral-spatial relationships, which is a key factor in understanding HSI. Even though deeper CNN architectures and use of 3-D-CNNs mitigate the above problem to a certain extent, they have increased computational complexity, which inhibits their use in resource-limited devices like IoT and edge computing devices. In this paper, we propose a novel method based on the concept of matrix capsules with Expectation-Maximization (EM) routing algorithm which is specifically designed to accommodate the nuances in the HSI data to efficiently tackle the aforementioned problems. The capsule units enable effective identification of spectral siguatures and part-whole relationships in the data while EM routing ensures viewpoint-invariance. Three representative HSI data sets are used to verify the effectiveness of the proposed method. The empirical results demonstrate that the proposed method is better than the current state-of-the-art methods in terms of accuracy while having 25 times fewer model parameters and requiring over 65 times less storage space. The source code can be found at https://github.com/DeepMatrixCapsules/DeepMatrixCapsules.
Despite the tremendous success of convolutional neural networks (CNNs) in computer vision, the mechanism of CNNs still lacks clear interpretation. Currently, class activation mapping (CAM), a famous visualization technique to interpret CNN's decision, has drawn increasing attention. Gradient-based CAMs are efficient while the performance is heavily affected by gradient vanishing and exploding. In contrast, gradient-free CAMs can avoid computing gradients to produce more understandable results. However, existing gradient-free CAMs are quite time-consuming because hundreds of forward interference per image are required. In this paper, we proposed Cluster-CAM, an effective and efficient gradient-free CNN interpretation algorithm. Cluster-CAM can significantly reduce the times of forward propagation by splitting the feature maps into clusters in an unsupervised manner. Furthermore, we propose an artful strategy to forge a cognition-base map and cognition-scissors from clustered feature maps. The final salience heatmap will be computed by merging the above cognition maps. Qualitative results conspicuously show that Cluster-CAM can produce heatmaps where the highlighted regions match the human's cognition more precisely than existing CAMs. The quantitative evaluation further demonstrates the superiority of Cluster-CAM in both effectiveness and efficiency.
There has been significant recent progress in training differentially private (DP) models which achieve accuracy that approaches the best non-private models. These DP models are typically pretrained on large public datasets and then fine-tuned on private downstream datasets that are relatively large and similar in distribution to the pretraining data. However, in many applications including personalization and federated learning, it is crucial to perform well (i) in the few-shot setting, as obtaining large amounts of labeled data may be problematic; and (ii) on datasets from a wide variety of domains for use in various specialist settings. To understand under which conditions few-shot DP can be effective, we perform an exhaustive set of experiments that reveals how the accuracy and vulnerability to attack of few-shot DP image classification models are affected as the number of shots per class, privacy level, model architecture, downstream dataset, and subset of learnable parameters in the model vary. We show that to achieve DP accuracy on par with non-private models, the shots per class must be increased as the privacy level increases. We also show that learning parameter-efficient FiLM adapters under DP is competitive with learning just the final classifier layer or learning all of the network parameters. Finally, we evaluate DP federated learning systems and establish state-of-the-art performance on the challenging FLAIR benchmark.
A number of recent adaptive optimizers improve the generalisation performance of Adam by essentially reducing the variance of adaptive stepsizes to get closer to SGD with momentum. Following the above motivation, we suppress the range of the adaptive stepsizes of Adam by exploiting the layerwise gradient statistics. In particular, at each iteration, we propose to perform three consecutive operations on the second momentum v_t before using it to update a DNN model: (1): down-scaling, (2): epsilon-embedding, and (3): down-translating. The resulting algorithm is referred to as SET-Adam, where SET is a brief notation of the three operations. The down-scaling operation on v_t is performed layerwise by making use of the angles between the layerwise subvectors of v_t and the corresponding all-one subvectors. Extensive experimental results show that SET-Adam outperforms eight adaptive optimizers when training transformers and LSTMs for NLP, and VGG and ResNet for image classification over CIAF10 and CIFAR100 while matching the best performance of the eight adaptive methods when training WGAN-GP models for image generation tasks. Furthermore, SET-Adam produces higher validation accuracies than Adam and AdaBelief for training ResNet18 over ImageNet.
By investigating iterative methods for a constrained linear model, we propose a new class of fully connected V-cycle MgNet for long-term time series forecasting, which is one of the most difficult tasks in forecasting. MgNet is a CNN model that was proposed for image classification based on the multigrid (MG) methods for solving discretized partial differential equations (PDEs). We replace the convolutional operations with fully connected operations in the existing MgNet and then apply them to forecasting problems. Motivated by the V-cycle structure in MG, we further propose the FV-MgNet, a V-cycle version of the fully connected MgNet, to extract features hierarchically. By evaluating the performance of FV-MgNet on popular data sets and comparing it with state-of-the-art models, we show that the FV-MgNet achieves better results with less memory usage and faster inference speed. In addition, we develop ablation experiments to demonstrate that the structure of FV-MgNet is the best choice among the many variants.
Grammar-Guided Genetic Programming (GGGP) employs a variety of insights from evolutionary theory to autonomously design solutions for a given task. Recent insights from evolutionary biology can lead to further improvements in GGGP algorithms. In this paper, we apply principles from the theory of Facilitated Variation and knowledge about heterogeneous mutation rates and mutation effects to improve the variation operators. We term this new method of variation Facilitated Mutation (FM). We test FM performance on the evolution of neural network optimizers for image classification, a relevant task in evolutionary computation, with important implications for the field of machine learning. We compare FM and FM combined with crossover (FMX) against a typical mutation regime to assess the benefits of the approach. We find that FMX in particular provides statistical improvements in key metrics, creating a superior optimizer overall (+0.48\% average test accuracy), improving the average quality of solutions (+50\% average population fitness), and discovering more diverse high-quality behaviors (+400 high-quality solutions discovered per run on average). Additionally, FM and FMX can reduce the number of fitness evaluations in an evolutionary run, reducing computational costs in some scenarios.
In this paper, we introduce weight prediction into the AdamW optimizer to boost its convergence when training the deep neural network (DNN) models. In particular, ahead of each mini-batch training, we predict the future weights according to the update rule of AdamW and then apply the predicted future weights to do both forward pass and backward propagation. In this way, the AdamW optimizer always utilizes the gradients w.r.t. the future weights instead of current weights to update the DNN parameters, making the AdamW optimizer achieve better convergence. Our proposal is simple and straightforward to implement but effective in boosting the convergence of DNN training. We performed extensive experimental evaluations on image classification and language modeling tasks to verify the effectiveness of our proposal. The experimental results validate that our proposal can boost the convergence of AdamW and achieve better accuracy than AdamW when training the DNN models.
This paper introduces a novel approach to leverage features learned from both supervised and self-supervised paradigms, to improve image classification tasks, specifically for vehicle classification. Two state-of-the-art self-supervised learning methods, DINO and data2vec, were evaluated and compared for their representation learning of vehicle images. The former contrasts local and global views while the latter uses masked prediction on multi-layered representations. In the latter case, supervised learning is employed to finetune a pretrained YOLOR object detector for detecting vehicle wheels, from which definitive wheel positional features are retrieved. The representations learned from these self-supervised learning methods were combined with the wheel positional features for the vehicle classification task. Particularly, a random wheel masking strategy was utilized to finetune the previously learned representations in harmony with the wheel positional features during the training of the classifier. Our experiments show that the data2vec-distilled representations, which are consistent with our wheel masking strategy, outperformed the DINO counterpart, resulting in a celebrated Top-1 classification accuracy of 97.2% for classifying the 13 vehicle classes defined by the Federal Highway Administration.
We propose a simple approach which combines the strengths of probabilistic graphical models and deep learning architectures for solving the multi-label classification task, focusing specifically on image and video data. First, we show that the performance of previous approaches that combine Markov Random Fields with neural networks can be modestly improved by leveraging more powerful methods such as iterative join graph propagation, integer linear programming, and $\ell_1$ regularization-based structure learning. Then we propose a new modeling framework called deep dependency networks, which augments a dependency network, a model that is easy to train and learns more accurate dependencies but is limited to Gibbs sampling for inference, to the output layer of a neural network. We show that despite its simplicity, jointly learning this new architecture yields significant improvements in performance over the baseline neural network. In particular, our experimental evaluation on three video activity classification datasets: Charades, Textually Annotated Cooking Scenes (TACoS), and Wetlab, and three multi-label image classification datasets: MS-COCO, PASCAL VOC, and NUS-WIDE show that deep dependency networks are almost always superior to pure neural architectures that do not use dependency networks.
Recent years have witnessed a big convergence of language, vision, and multi-modal pretraining. In this work, we present mPLUG-2, a new unified paradigm with modularized design for multi-modal pretraining, which can benefit from modality collaboration while addressing the problem of modality entanglement. In contrast to predominant paradigms of solely relying on sequence-to-sequence generation or encoder-based instance discrimination, mPLUG-2 introduces a multi-module composition network by sharing common universal modules for modality collaboration and disentangling different modality modules to deal with modality entanglement. It is flexible to select different modules for different understanding and generation tasks across all modalities including text, image, and video. Empirical study shows that mPLUG-2 achieves state-of-the-art or competitive results on a broad range of over 30 downstream tasks, spanning multi-modal tasks of image-text and video-text understanding and generation, and uni-modal tasks of text-only, image-only, and video-only understanding. Notably, mPLUG-2 shows new state-of-the-art results of 48.0 top-1 accuracy and 80.3 CIDEr on the challenging MSRVTT video QA and video caption tasks with a far smaller model size and data scale. It also demonstrates strong zero-shot transferability on vision-language and video-language tasks. Code and models will be released in https://github.com/alibaba/AliceMind.
Deep active learning (DAL) methods have shown significant improvements in sample efficiency compared to simple random sampling. While these studies are valuable, they nearly always assume that optimal DAL hyperparameter (HP) settings are known in advance, or optimize the HPs through repeating DAL several times with different HP settings. Here, we argue that in real-world settings, or in the wild, there is significant uncertainty regarding good HPs, and their optimization contradicts the premise of using DAL (i.e., we require labeling efficiency). In this study, we evaluate the performance of eleven modern DAL methods on eight benchmark problems as we vary a key HP shared by all methods: the pool ratio. Despite adjusting only one HP, our results indicate that eight of the eleven DAL methods sometimes underperform relative to simple random sampling and some frequently perform worse. Only three methods always outperform random sampling (albeit narrowly), and we find that these methods all utilize diversity to select samples - a relatively simple criterion. Our findings reveal the limitations of existing DAL methods when deployed in the wild, and present this as an important new open problem in the field.
Siamese networks are one of the most trending methods to achieve self-supervised visual representation learning (SSL). Since hand labeling is costly, SSL can play a crucial part by allowing deep learning to train on large unlabeled datasets. Meanwhile, Neural Architecture Search (NAS) is becoming increasingly important as a technique to discover novel deep learning architectures. However, early NAS methods based on reinforcement learning or evolutionary algorithms suffered from ludicrous computational and memory costs. In contrast, differentiable NAS, a gradient-based approach, has the advantage of being much more efficient and has thus retained most of the attention in the past few years. In this article, we present NASiam, a novel approach that uses for the first time differentiable NAS to improve the multilayer perceptron projector and predictor (encoder/predictor pair) architectures inside siamese-networks-based contrastive learning frameworks (e.g., SimCLR, SimSiam, and MoCo) while preserving the simplicity of previous baselines. We crafted a search space designed explicitly for multilayer perceptrons, inside which we explored several alternatives to the standard ReLU activation function. We show that these new architectures allow ResNet backbone convolutional models to learn strong representations efficiently. NASiam reaches competitive performance in both small-scale (i.e., CIFAR-10/CIFAR-100) and large-scale (i.e., ImageNet) image classification datasets while costing only a few GPU hours. We discuss the composition of the NAS-discovered architectures and emit hypotheses on why they manage to prevent collapsing behavior. Our code is available at https://github.com/aheuillet/NASiam.
When training deep neural networks, keeping all tensors in high precision (e.g., 32-bit or even 16-bit floats) is often wasteful. However, keeping all tensors in low precision (e.g., 8-bit floats) can lead to unacceptable accuracy loss. Hence, it is important to use a precision assignment -- a mapping from all tensors (arising in training) to precision levels (high or low) -- that keeps most of the tensors in low precision and leads to sufficiently accurate models. We provide a technique that explores this memory-accuracy tradeoff by generating precision assignments for convolutional neural networks that (i) use less memory and (ii) lead to more accurate convolutional networks at the same time, compared to the precision assignments considered by prior work in low-precision floating-point training. We evaluate our technique on image classification tasks by training convolutional networks on CIFAR-10, CIFAR-100, and ImageNet. Our method typically provides > 2x memory reduction over a baseline precision assignment while preserving training accuracy, and gives further reductions by trading off accuracy. Compared to other baselines which sometimes cause training to diverge, our method provides similar or better memory reduction while avoiding divergence.
Real-world data contains a vast amount of multimodal information, among which vision and language are the two most representative modalities. Moreover, increasingly heavier models, \textit{e}.\textit{g}., Transformers, have attracted the attention of researchers to model compression. However, how to compress multimodal models, especially vison-language Transformers, is still under-explored. This paper proposes the \textbf{U}nified and \textbf{P}r\textbf{o}gressive \textbf{P}runing (\textbf{\emph{UPop}}) as a universal vison-language Transformer compression framework, which incorporates 1) unifiedly searching multimodal subnets in a continuous optimization space from the original model, which enables automatic assignment of pruning ratios among compressible modalities and structures; 2) progressively searching and retraining the subnet, which maintains convergence between the search and retrain to attain higher compression ratios. Experiments on various tasks, datasets, and model architectures demonstrate the effectiveness and versatility of the proposed UPop framework. The code is available at https://github.com/sdc17/UPop.
Semi-supervised learning (SSL) has been widely explored in recent years, and it is an effective way of leveraging unlabeled data to reduce the reliance on labeled data. In this work, we adjust neural processes (NPs) to the semi-supervised image classification task, resulting in a new method named NP-Match. NP-Match is suited to this task for two reasons. Firstly, NP-Match implicitly compares data points when making predictions, and as a result, the prediction of each unlabeled data point is affected by the labeled data points that are similar to it, which improves the quality of pseudo-labels. Secondly, NP-Match is able to estimate uncertainty that can be used as a tool for selecting unlabeled samples with reliable pseudo-labels. Compared with uncertainty-based SSL methods implemented with Monte-Carlo (MC) dropout, NP-Match estimates uncertainty with much less computational overhead, which can save time at both the training and the testing phases. We conducted extensive experiments on five public datasets under three semi-supervised image classification settings, namely, the standard semi-supervised image classification, the imbalanced semi-supervised image classification, and the multi-label semi-supervised image classification, and NP-Match outperforms state-of-the-art (SOTA) approaches or achieves competitive results on them, which shows the effectiveness of NP-Match and its potential for SSL. The codes are at https://github.com/Jianf-Wang/NP-Match
The primary goal of training in early convolutional neural networks (CNN) is the higher generalization performance of the model. However, as the expected calibration error (ECE), which quantifies the explanatory power of model inference, was recently introduced, research on training models that can be explained is in progress. We hypothesized that a gap in supervision criteria during training and inference leads to overconfidence, and investigated that performing label distribution learning (LDL) would enhance the model calibration in CNN training. To verify this assumption, we used a simple LDL setting with recent data augmentation techniques. Based on a series of experiments, the following results are obtained: 1) State-of-the-art KD methods significantly impede model calibration. 2) Training using LDL with recent data augmentation can have excellent effects on model calibration and even in generalization performance. 3) Online LDL brings additional improvements in model calibration and accuracy with long training, especially in large-size models. Using the proposed approach, we simultaneously achieved a lower ECE and higher generalization performance for the image classification datasets CIFAR10, 100, STL10, and ImageNet. We performed several visualizations and analyses and witnessed several interesting behaviors in CNN training with the LDL.
Vision Transformers (ViTs) are becoming a very popular paradigm for vision tasks as they achieve state-of-the-art performance on image classification. However, although early works implied that this network structure had increased robustness against adversarial attacks, some works argue ViTs are still vulnerable. This paper presents our first attempt toward detecting adversarial attacks during inference time using the network's input and outputs as well as latent features. We design four quantifications (or derivatives) of input, output, and latent vectors of ViT-based models that provide a signature of the inference, which could be beneficial for the attack detection, and empirically study their behavior over clean samples and adversarial samples. The results demonstrate that the quantifications from input (images) and output (posterior probabilities) are promising for distinguishing clean and adversarial samples, while latent vectors offer less discriminative power, though they give some insights on how adversarial perturbations work.
A good feature representation is the key to image classification. In practice, image classifiers may be applied in scenarios different from what they have been trained on. This so-called domain shift leads to a significant performance drop in image classification. Unsupervised domain adaptation (UDA) reduces the domain shift by transferring the knowledge learned from a labeled source domain to an unlabeled target domain. We perform feature disentanglement for UDA by distilling category-relevant features and excluding category-irrelevant features from the global feature maps. This disentanglement prevents the network from overfitting to category-irrelevant information and makes it focus on information useful for classification. This reduces the difficulty of domain alignment and improves the classification accuracy on the target domain. We propose a coarse-to-fine domain adaptation method called Domain Adaptation via Feature Disentanglement~(DAFD), which has two components: (1)the Category-Relevant Feature Selection (CRFS) module, which disentangles the category-relevant features from the category-irrelevant features, and (2)the Dynamic Local Maximum Mean Discrepancy (DLMMD) module, which achieves fine-grained alignment by reducing the discrepancy within the category-relevant features from different domains. Combined with the CRFS, the DLMMD module can align the category-relevant features properly. We conduct comprehensive experiment on four standard datasets. Our results clearly demonstrate the robustness and effectiveness of our approach in domain adaptive image classification tasks and its competitiveness to the state of the art.
Early diagnosis of Alzheimer's disease (AD) is essential in preventing the disease's progression. Therefore, detecting AD from neuroimaging data such as structural magnetic resonance imaging (sMRI) has been a topic of intense investigation in recent years. Deep learning has gained considerable attention in Alzheimer's detection. However, training a convolutional neural network from scratch is challenging since it demands more computational time and a significant amount of annotated data. By transferring knowledge learned from other image recognition tasks to medical image classification, transfer learning can provide a promising and effective solution. Irregularities in the dataset distribution present another difficulty. Class decomposition can tackle this issue by simplifying learning a dataset's class boundaries. Motivated by these approaches, this paper proposes a transfer learning method using class decomposition to detect Alzheimer's disease from sMRI images. We use two ImageNet-trained architectures: VGG19 and ResNet50, and an entropy-based technique to determine the most informative images. The proposed model achieved state-of-the-art performance in the Alzheimer's disease (AD) vs mild cognitive impairment (MCI) vs cognitively normal (CN) classification task with a 3\% increase in accuracy from what is reported in the literature.