Adaptive gradient methods such as RMSProp and Adam use exponential moving estimate of the squared gradient to compute adaptive step sizes, achieving better convergence than SGD in face of noisy objectives. However, Adam can have undesirable convergence behaviors due to unstable or extreme adaptive learning rates. Methods such as AMSGrad and AdaBound have been proposed to stabilize the adaptive learning rates of Adam in the later stage of training, but they do not outperform Adam in some practical tasks such as training Transformers \cite{transformer}. In this paper, we propose an adaptive learning rate principle, in which the running mean of squared gradient in Adam is replaced by a weighted mean, with weights chosen to maximize the estimated variance of each coordinate. This results in a faster adaptation to the local gradient variance, which leads to more desirable empirical convergence behaviors than Adam. We prove the proposed algorithm converges under mild assumptions for nonconvex stochastic optimization problems, and demonstrate the improved efficacy of our adaptive averaging approach on machine translation, natural language understanding and large-batch pretraining of BERT. The code is available at https://github.com/zhuchen03/MaxVA.
Bayesian meta-learning enables robust and fast adaptation to new tasks with uncertainty assessment. The key idea behind Bayesian meta-learning is empirical Bayes inference of hierarchical model. In this work, we extend this framework to include a variety of existing methods, before proposing our variant based on gradient-EM algorithm. Our method improves computational efficiency by avoiding back-propagation computation in the meta-update step, which is exhausting for deep neural networks. Furthermore, it provides flexibility to the inner-update optimization procedure by decoupling it from meta-update. Experiments on sinusoidal regression, few-shot image classification, and policy-based reinforcement learning show that our method not only achieves better accuracy with less computation cost, but is also more robust to uncertainty.
In the absence of large labelled datasets, self-supervised learning techniques can boost performance by learning useful representations from unlabelled data, which is often more readily available. However, there is often a domain shift between the unlabelled collection and the downstream target problem data. We show that by learning Bayesian instance weights for the unlabelled data, we can improve the downstream classification accuracy by prioritising the most useful instances. Additionally, we show that the training time can be reduced by discarding unnecessary datapoints. Our method, BetaDataWeighter is evaluated using the popular self-supervised rotation prediction task on STL-10 and Visual Decathlon. We compare to related instance weighting schemes, both hand-designed heuristics and meta-learning, as well as conventional self-supervised learning. BetaDataWeighter achieves both the highest average accuracy and rank across datasets, and on STL-10 it prunes up to 78% of unlabelled images without significant loss in accuracy, corresponding to over 50% reduction in training time.
The recent growth in the number of satellite images fosters the development of effective deep-learning techniques for Remote Sensing (RS). However, their full potential is untapped due to the lack of large annotated datasets. Such a problem is usually countered by fine-tuning a feature extractor that is previously trained on the ImageNet dataset. Unfortunately, the domain of natural images differs from the RS one, which hinders the final performance. In this work, we propose to learn meaningful representations from satellite imagery, leveraging its high-dimensionality spectral bands to reconstruct the visible colors. We conduct experiments on land cover classification (BigEarthNet) and West Nile Virus detection, showing that colorization is a solid pretext task for training a feature extractor. Furthermore, we qualitatively observe that guesses based on natural images and colorization rely on different parts of the input. This paves the way to an ensemble model that eventually outperforms both the above-mentioned techniques.
Texture is a fundamental characteristic of many types of images, especially those with significant rotation, scale illumination, and viewpoint change. Texture image classification is one of the challenging problems that have various applications such as remote sensing, material recognition, and computer-aided medical diagnosis, etc. Various Computer vision techniques have been used. More recently, Deep learning architectures demonstrated impressive results. This paper aims to investigate combining two feature extraction methods: Handcrafted-based and CNN-based in a two-stream neural network architecture. We believe that Statistical features could enhance the performance of the CNN architecture, especially in the case of small datasets. To test our approach we used two challenging datasets, the Describable Textures Dataset (DTD) and Flicker Material Database (FMD). Results showed that our two-stream neural network which has an image as a first stream and a statistical feature vector as a second stream achieve better results than a Convolutional neural network achieved with just the RGB image as input. The Xception network [9] combined with SIFT-FV demonstrated an accuracy superiority for both datasets.
The past years have witnessed great progress on remote sensing (RS) image interpretation and its wide applications. With RS images becoming more accessible than ever before, there is an increasing demand for the automatic interpretation of these images. In this context, the benchmark datasets serve as essential prerequisites for developing and testing intelligent interpretation algorithms. After reviewing existing benchmark datasets in the research community of RS image interpretation, this article discusses the problem of how to efficiently prepare a suitable benchmark dataset for RS image interpretation. Specifically, we first analyze the current challenges of developing intelligent algorithms for RS image interpretation with bibliometric investigations. We then present the general guidances on creating benchmark datasets in efficient manners. Following the presented guidances, we also provide an example on building RS image dataset, i.e., Million-AID, a new large-scale benchmark dataset containing a million instances for RS image scene classification. Several challenges and perspectives in RS image annotation are finally discussed to facilitate the research in benchmark dataset construction. We do hope this paper will provide the RS community an overall perspective on constructing large-scale and practical image datasets for further research, especially data-driven ones.
In active learning, sampling bias could pose a serious inconsistency problem and hinder the algorithm from finding the optimal hypothesis. However, many methods for neural networks are hypothesis space agnostic and do not address this problem. We examine active learning with convolutional neural networks through the principled lens of version space reduction. We identify the connection between two approaches---prior mass reduction and diameter reduction---and propose a new diameter-based querying method---the minimum Gibbs-vote disagreement. By estimating version space diameter and bias, we illustrate how version space of neural networks evolves and examine the realizability assumption. With experiments on MNIST, Fashion-MNIST, SVHN and STL-10 datasets, we demonstrate that diameter reduction methods reduce the version space more effectively and perform better than prior mass reduction and other baselines, and that the Gibbs vote disagreement is on par with the best query method.
Recommendation is a prevalent application of machine learning that affects many users; therefore, it is important for recommender models to be accurate and interpretable. In this work, we propose a method to both interpret and augment the predictions of black-box recommender systems. In particular, we propose to interpret feature interactions from a source recommender model and explicitly encode these interactions in a target recommender model, where both source and target models are black-boxes. By not assuming the structure of the recommender system, our approach can be used in general settings. In our experiments, we focus on a prominent use of machine learning recommendation: ad-click prediction. We found that our interaction interpretations are both informative and predictive, e.g., significantly outperforming existing recommender models. What's more, the same approach to interpret interactions can provide new insights into domains even beyond recommendation, such as text and image classification.
According to the World Health Organization, distracted driving is one of the leading cause of motor accidents and deaths in the world. In our study, we tackle the problem of distracted driving by aiming to build a robust multi-class classifier to detect and identify different forms of driver inattention using the State Farm Distracted Driving Dataset. We utilize combinations of pretrained image classification models, classical data augmentation, OpenCV based image preprocessing and skin segmentation augmentation approaches. Our best performing model combines several augmentation techniques, including skin segmentation, facial blurring, and classical augmentation techniques. This model achieves an approximately 15% increase in F1 score over the baseline, thus showing the promise in these techniques in enhancing the power of neural networks for the task of distracted driver detection.
Deep Convolutional Neural Networks are able to identify complex patterns and perform tasks with super-human capabilities. However, besides the exceptional results, they are not completely understood and it is still impractical to hand-engineer similar solutions. In this work, an image classification Convolutional Neural Network and its building blocks are described from a frequency domain perspective. Some network layers have established counterparts in the frequency domain like the convolutional and pooling layers. We propose the 2SReLU layer, a novel non-linear activation function that preserves high frequency components in deep networks. It is demonstrated that in the frequency domain it is possible to achieve competitive results without using the computationally costly convolution operation. A source code implementation in PyTorch is provided at: https://gitlab.com/thomio/2srelu
Fully Convolutional Networks have been achieving remarkable results in image semantic segmentation, while being efficient. Such efficiency results from the capability of segmenting several voxels in a single forward pass. So, there is a direct spatial correspondence between a unit in a feature map and the voxel in the same location. In a convolutional layer, the kernel spans over all channels and extracts information from them. We observe that linear recombination of feature maps by increasing the number of channels followed by compression may enhance their discriminative power. Moreover, not all feature maps have the same relevance for the classes being predicted. In order to learn the inter-channel relationships and recalibrate the channels to suppress the less relevant ones, Squeeze and Excitation blocks were proposed in the context of image classification with Convolutional Neural Networks. However, this is not well adapted for segmentation with Fully Convolutional Networks since they segment several objects simultaneously, hence a feature map may contain relevant information only in some locations. In this paper, we propose recombination of features and a spatially adaptive recalibration block that is adapted for semantic segmentation with Fully Convolutional Networks - the SegSE block. Feature maps are recalibrated by considering the cross-channel information together with spatial relevance. Experimental results indicate that Recombination and Recalibration improve the results of a competitive baseline, and generalize across three different problems: brain tumor segmentation, stroke penumbra estimation, and ischemic stroke lesion outcome prediction. The obtained results are competitive or outperform the state of the art in the three applications.
Second-order optimizers hold intriguing potential for deep learning, but suffer from increased cost and sensitivity to the non-convexity of the loss surface as compared to gradient-based approaches. We introduce a coordinate descent method to train deep neural networks for classification tasks that exploits global convexity of the cross-entropy loss in the weights of the linear layer. Our hybrid Newton/Gradient Descent (NGD) method is consistent with the interpretation of hidden layers as providing an adaptive basis and the linear layer as providing an optimal fit of the basis to data. By alternating between a second-order method to find globally optimal parameters for the linear layer and gradient descent to train the hidden layers, we ensure an optimal fit of the adaptive basis to data throughout training. The size of the Hessian in the second-order step scales only with the number weights in the linear layer and not the depth and width of the hidden layers; furthermore, the approach is applicable to arbitrary hidden layer architecture. Previous work applying this adaptive basis perspective to regression problems demonstrated significant improvements in accuracy at reduced training cost, and this work can be viewed as an extension of this approach to classification problems. We first prove that the resulting Hessian matrix is symmetric semi-definite, and that the Newton step realizes a global minimizer. By studying classification of manufactured two-dimensional point cloud data, we demonstrate both an improvement in validation error and a striking qualitative difference in the basis functions encoded in the hidden layer when trained using NGD. Application to image classification benchmarks for both dense and convolutional architectures reveals improved training accuracy, suggesting possible gains of second-order methods over gradient descent.
We propose a method for efficiently incorporating constraints into a stochastic gradient Langevin framework for the training of deep neural networks. Constraints allow direct control of the parameter space of the model. Appropriately designed, they reduce the vanishing/exploding gradient problem, control weight magnitudes and stabilize deep neural networks and thus improve the robustness of training algorithms and the generalization capabilities of the trained neural network. We present examples of constrained training methods motivated by orthogonality preservation for weight matrices and explicit weight normalizations. We describe the methods in the overdamped formulation of Langevin dynamics and the underdamped form, in which momenta help to improve sampling efficiency. The methods are explored in test examples in image classification and natural language processing.
Solving long-tail large vocabulary object detection with deep learning based models is a challenging and demanding task, which is however under-explored.In this work, we provide the first systematic analysis on the underperformance of state-of-the-art models in front of long-tail distribution. We find existing detection methods are unable to model few-shot classes when the dataset is extremely skewed, which can result in classifier imbalance in terms of parameter magnitude. Directly adapting long-tail classification models to detection frameworks can not solve this problem due to the intrinsic difference between detection and classification.In this work, we propose a novel balanced group softmax (BAGS) module for balancing the classifiers within the detection frameworks through group-wise training. It implicitly modulates the training process for the head and tail classes and ensures they are both sufficiently trained, without requiring any extra sampling for the instances from the tail classes.Extensive experiments on the very recent long-tail large vocabulary object recognition benchmark LVIS show that our proposed BAGS significantly improves the performance of detectors with various backbones and frameworks on both object detection and instance segmentation. It beats all state-of-the-art methods transferred from long-tail image classification and establishes new state-of-the-art.Code is available at https://github.com/FishYuLi/BalancedGroupSoftmax.
We propose a novel pool-based Active Learning framework constructed on a sequential Graph Convolution Network (GCN). Each image's feature from a pool of data represents a node in the graph and the edges encode their similarities. With a small number of randomly sampled images as seed labelled examples, we learn the parameters of the graph to distinguish labelled vs unlabelled nodes by minimising the binary cross-entropy loss. GCN performs message-passing operations between the nodes, and hence, induces similar representations of the strongly associated nodes. We exploit these characteristics of GCN to select the unlabelled examples which are sufficiently different from labelled ones. To this end, we utilise the graph node embeddings and their confidence scores and adapt sampling techniques such as CoreSet and uncertainty-based methods to query the nodes. We flip the label of newly queried nodes from unlabelled to labelled, re-train the learner to optimise the downstream task and the graph to minimise its modified objective. We continue this process within a fixed budget. We evaluate our method on 6 different benchmarks:4 real image classification, 1 depth-based hand pose estimation and 1 synthetic RGB image classification datasets. Our method outperforms several competitive baselines such as VAAL, Learning Loss, CoreSet and attains the new state-of-the-art performance on multiple applications The implementations can be found here: https://github.com/razvancaramalau/Sequential-GCN-for-Active-Learning
A model must adapt itself to generalize to new and different data during testing. In this setting of fully test-time adaptation the model has only the test data and its own parameters. We propose to adapt by test entropy minimization (tent): we optimize the model for confidence as measured by the entropy of its predictions. Our method estimates normalization statistics and optimizes channel-wise affine transformations to update online on each batch. Tent reduces generalization error for image classification on corrupted ImageNet and CIFAR-10/100 and reaches a new state-of-the-art error on ImageNet-C. Tent handles source-free domain adaptation on digit recognition from SVHN to MNIST/MNIST-M/USPS, on semantic segmentation from GTA to Cityscapes, and on the VisDA-C benchmark. These results are achieved in one epoch of test-time optimization without altering training.
Simi-Supervised Recognition Challenge-FGVC7 is a challenging fine-grained recognition competition. One of the difficulties of this competition is how to use unlabeled data. We adopted pseudo-tag data mining to increase the amount of training data. The other one is how to identify similar birds with a very small difference, especially those have a relatively tiny main-body in examples. We combined generic image recognition and fine-grained image recognition method to solve the problem. All generic image recognition models were training using PaddleClas . Using the combination of two different ways of deep recognition models, we finally won the third place in the competition.
Automatic supervised classification with complex modelling such as deep neural networks requires the availability of representative training data sets. While there exists a plethora of data sets that can be used for this purpose, they are usually very heterogeneous and not interoperable. In this context, the present work has a twofold objective: i) to describe procedures of open-source training data management, integration, and data retrieval, and ii) to demonstrate the practical use of varying source training data for remote sensing image classification. For the former, we propose SatImNet, a collection of open training data, structured and harmonized according to specific rules. For the latter, two modelling approaches based on convolutional neural networks have been designed and configured to deal with satellite image classification and segmentation.
Training neural networks requires increasing amounts of memory. Parameter sharing can reduce memory and communication costs, but existing methods assume networks have many identical layers and utilize hand-crafted sharing strategies that fail to generalize. We introduce Neural Parameter Allocation Search (NPAS), a novel task where the goal is to train a neural network given an arbitrary, fixed parameter budget. NPAS covers both low-budget regimes, which produce compact networks, as well as a novel high-budget regime, where additional capacity can be added to boost performance without increasing inference FLOPs. To address NPAS, we introduce Shapeshifter Networks (SSNs), which automatically learn where and how to share parameters in a network to support any parameter budget without requiring any changes to the architecture or loss function. NPAS and SSNs provide a complete framework for addressing generalized parameter sharing, and can also be combined with prior work for additional performance gains. We demonstrate the effectiveness of our approach using nine network architectures across four diverse tasks, including ImageNet classification and transformers.
Spiking Neural Networks (SNNs), the third generation NNs, have come under the spotlight for machine learning based applications due to their biological plausibility and reduced complexity compared to traditional artificial Deep Neural Networks (DNNs). These SNNs can be implemented with extreme energy efficiency on neuromorphic processors like the Intel Loihi research chip, and fed by event-based sensors, such as DVS cameras. However, DNNs with many layers can achieve relatively high accuracy on image classification and recognition tasks, as the research on learning rules for SNNs for real-world applications is still not mature. The accuracy results for SNNs are typically obtained either by converting the trained DNNs into SNNs, or by directly designing and training SNNs in the spiking domain. Towards the conversion from a DNN to an SNN, we perform a comprehensive analysis of such process, specifically designed for Intel Loihi, showing our methodology for the design of an SNN that achieves nearly the same accuracy results as its corresponding DNN. Towards the usage of the event-based sensors, we design a pre-processing method, evaluated for the DvsGesture dataset, which makes it possible to be used in the DNN domain. Hence, based on the outcome of the first analysis, we train a DNN for the pre-processed DvsGesture dataset, and convert it into the spike domain for its deployment on Intel Loihi, which enables real-time gesture recognition. The results show that our SNN achieves 89.64% classification accuracy and occupies only 37 Loihi cores. The source code for generating our experiments is available online at https://github.com/albertomarchisio/EfficientSNN.