Increasing model, data and compute budget scale in the pre-training has been shown to strongly improve model generalization and transfer learning in vast line of work done in language modeling and natural image recognition. However, most studies on the positive effect of larger scale were done in scope of in-domain setting, with source and target data being in close proximity. To study effect of larger scale for both in-domain and out-of-domain setting when performing full and few-shot transfer, we combine here for the first time large, openly available medical X-Ray chest imaging datasets to reach a scale for medical imaging domain comparable to ImageNet-1k, routinely used for pre-training in natural image domain. We then conduct supervised pre-training, while varying network size and source data scale and domain, being either large natural (ImageNet-1k/21k) or large medical chest X-Ray datasets, and transfer pre-trained models to different natural or medical targets. We observe strong improvement due to larger pre-training scale for intra-domain natural-natural and medical-medical transfer. For inter-domain natural-medical transfer, we find improvements due to larger pre-training scale on larger X-Ray targets in full shot regime, while for smaller targets and for few-shot regime the improvement is not visible. Remarkably, large networks pre-trained on very large natural ImageNet-21k are as good or better than networks pre-trained on largest available medical X-Ray data when performing transfer to large X-Ray targets. We conclude that substantially increasing model and generic, medical domain-agnostic natural image source data scale in the pre-training can enable high quality out-of-domain transfer to medical domain specific targets, removing dependency on large medical domain-specific source data often not available in the practice.
Transformers have recently gained increasing attention in computer vision. However, existing studies mostly use Transformers for feature representation learning, e.g. for image classification and dense predictions, and the generalizability of Transformers is unknown. In this work, we further investigate the possibility of applying Transformers for image matching and metric learning given pairs of images. We find that the Vision Transformer (ViT) and the vanilla Transformer with decoders are not adequate for image matching due to their lack of image-to-image attention. Thus, we further design two naive solutions, i.e. query-gallery concatenation in ViT, and query-gallery cross-attention in the vanilla Transformer. The latter improves the performance, but it is still limited. This implies that the attention mechanism in Transformers is primarily designed for global feature aggregation, which is not naturally suitable for image matching. Accordingly, we propose a new simplified decoder, which drops the full attention implementation with the softmax weighting, keeping only the query-key similarity computation. Additionally, global max pooling and a multilayer perceptron (MLP) head are applied to decode the matching result. This way, the simplified decoder is computationally more efficient, while at the same time more effective for image matching. The proposed method, called TransMatcher, achieves state-of-the-art performance in generalizable person re-identification, with up to 6.1% and 5.7% performance gains in Rank-1 and mAP, respectively, on several popular datasets. Code is available at https://github.com/ShengcaiLiao/QAConv.
Deep learning techniques have been widely applied to hyperspectral image (HSI) classification and have achieved great success. However, the deep neural network model has a large parameter space and requires a large number of labeled data. Deep learning methods for HSI classification usually follow a patchwise learning framework. Recently, a fast patch-free global learning (FPGA) architecture was proposed for HSI classification according to global spatial context information. However, FPGA has difficulty extracting the most discriminative features when the sample data is imbalanced. In this paper, a spectral-spatial dependent global learning (SSDGL) framework based on global convolutional long short-term memory (GCL) and global joint attention mechanism (GJAM) is proposed for insufficient and imbalanced HSI classification. In SSDGL, the hierarchically balanced (H-B) sampling strategy and the weighted softmax loss are proposed to address the imbalanced sample problem. To effectively distinguish similar spectral characteristics of land cover types, the GCL module is introduced to extract the long short-term dependency of spectral features. To learn the most discriminative feature representations, the GJAM module is proposed to extract attention areas. The experimental results obtained with three public HSI datasets show that the SSDGL has powerful performance in insufficient and imbalanced sample problems and is superior to other state-of-the-art methods. Code can be obtained at: https://github.com/dengweihuan/SSDGL.
Transformers have become one of the dominant architectures in deep learning, particularly as a powerful alternative to convolutional neural networks (CNNs) in computer vision. However, Transformer training and inference in previous works can be prohibitively expensive due to the quadratic complexity of self-attention over a long sequence of representations, especially for high-resolution dense prediction tasks. To this end, we present a novel Less attention vIsion Transformer (LIT), building upon the fact that the early self-attention layers in Transformers still focus on local patterns and bring minor benefits in recent hierarchical vision Transformers. Specifically, we propose a hierarchical Transformer where we use pure multi-layer perceptrons (MLPs) to encode rich local patterns in the early stages while applying self-attention modules to capture longer dependencies in deeper layers. Moreover, we further propose a learned deformable token merging module to adaptively fuse informative patches in a non-uniform manner. The proposed LIT achieves promising performance on image recognition tasks, including image classification, object detection and instance segmentation, serving as a strong backbone for many vision tasks. Code is available at: https://github.com/zhuang-group/LIT
Transformers have offered a new methodology of designing neural networks for visual recognition. Compared to convolutional networks, Transformers enjoy the ability of referring to global features at each stage, yet the attention module brings higher computational overhead that obstructs the application of Transformers to process high-resolution visual data. This paper aims to alleviate the conflict between efficiency and flexibility, for which we propose a specialized token for each region that serves as a messenger (MSG). Hence, by manipulating these MSG tokens, one can flexibly exchange visual information across regions and the computational complexity is reduced. We then integrate the MSG token into a multi-scale architecture named MSG-Transformer. In standard image classification and object detection, MSG-Transformer achieves competitive performance and the inference on both GPU and CPU is accelerated. Code is available at https://github.com/hustvl/MSG-Transformer.
In this article, we introduce a novel variant of the Tsetlin machine (TM) that randomly drops clauses, the key learning elements of a TM. In effect, TM with drop clause ignores a random selection of the clauses in each epoch, selected according to a predefined probability. In this way, additional stochasticity is introduced in the learning phase of TM. To explore the effects drop clause has on accuracy, training time, interpretability and robustness, we conduct extensive experiments on nine benchmark datasets in natural language processing~(NLP) (IMDb, R8, R52, MR and TREC) and image classification (MNIST, Fashion MNIST, CIFAR-10 and CIFAR-100). Our proposed model outperforms baseline machine learning algorithms by a wide margin and achieves competitive performance in comparison with recent deep learning model such as BERT and AlexNET-DFA. In brief, we observe up to +10% increase in accuracy and 2x to 4x faster learning compared with standard TM. We further employ the Convolutional TM to document interpretable results on the CIFAR datasets, visualizing how the heatmaps produced by the TM become more interpretable with drop clause. We also evaluate how drop clause affects learning robustness by introducing corruptions and alterations in the image/language test data. Our results show that drop clause makes TM more robust towards such changes.
Vision Transformers (ViT) have achieved remarkable success in large-scale image recognition. They split every 2D image into a fixed number of patches, each of which is treated as a token. Generally, representing an image with more tokens would lead to higher prediction accuracy, while it also results in drastically increased computational cost. To achieve a decent trade-off between accuracy and speed, the number of tokens is empirically set to 16x16 or 14x14. In this paper, we argue that every image has its own characteristics, and ideally the token number should be conditioned on each individual input. In fact, we have observed that there exist a considerable number of "easy" images which can be accurately predicted with a mere number of 4x4 tokens, while only a small fraction of "hard" ones need a finer representation. Inspired by this phenomenon, we propose a Dynamic Transformer to automatically configure a proper number of tokens for each input image. This is achieved by cascading multiple Transformers with increasing numbers of tokens, which are sequentially activated in an adaptive fashion at test time, i.e., the inference is terminated once a sufficiently confident prediction is produced. We further design efficient feature reuse and relationship reuse mechanisms across different components of the Dynamic Transformer to reduce redundant computations. Extensive empirical results on ImageNet, CIFAR-10, and CIFAR-100 demonstrate that our method significantly outperforms the competitive baselines in terms of both theoretical computational efficiency and practical inference speed. Code and pre-trained models (based on PyTorch and MindSpore) are available at https://github.com/blackfeather-wang/Dynamic-Vision-Transformer and https://github.com/blackfeather-wang/Dynamic-Vision-Transformer-MindSpore.
Explainable artificial intelligence is the attempt to elucidate the workings of systems too complex to be directly accessible to human cognition through suitable side-information referred to as "explanations". We present a trainable explanation module for convolutional image classifiers we call bounded logit attention (BLA). The BLA module learns to select a subset of the convolutional feature map for each input instance, which then serves as an explanation for the classifier's prediction. BLA overcomes several limitations of the instancewise feature selection method "learning to explain" (L2X) introduced by Chen et al. (2018): 1) BLA scales to real-world sized image classification problems, and 2) BLA offers a canonical way to learn explanations of variable size. Due to its modularity BLA lends itself to transfer learning setups and can also be employed as a post-hoc add-on to trained classifiers. Beyond explainability, BLA may serve as a general purpose method for differentiable approximation of subset selection. In a user study we find that BLA explanations are preferred over explanations generated by the popular (Grad-)CAM method.
Recently, it has been demonstrated that the performance of a deep convolutional neural network can be effectively improved by embedding an attention module into it. In this work, a novel lightweight and effective attention method named Pyramid Squeeze Attention (PSA) module is proposed. By replacing the 3x3 convolution with the PSA module in the bottleneck blocks of the ResNet, a novel representational block named Efficient Pyramid Squeeze Attention (EPSA) is obtained. The EPSA block can be easily added as a plug-and-play component into a well-established backbone network, and significant improvements on model performance can be achieved. Hence, a simple and efficient backbone architecture named EPSANet is developed in this work by stacking these ResNet-style EPSA blocks. Correspondingly, a stronger multi-scale representation ability can be offered by the proposed EPSANet for various computer vision tasks including but not limited to, image classification, object detection, instance segmentation, etc. Without bells and whistles, the performance of the proposed EPSANet outperforms most of the state-of-the-art channel attention methods. As compared to the SENet-50, the Top-1 accuracy is improved by 1.93% on ImageNet dataset, a larger margin of +2.7 box AP for object detection and an improvement of +1.7 mask AP for instance segmentation by using the Mask-RCNN on MS-COCO dataset are obtained. Our source code is available at:https://github.com/murufeng/EPSANet.
For classification tasks, deep neural networks are prone to overfitting in the presence of label noise. Although existing methods are able to alleviate this problem at low noise levels, they encounter significant performance reduction at high noise levels, or even at medium noise levels when the label noise is asymmetric. To train classifiers that are universally robust to all noise levels, and that are not sensitive to any variation in the noise model, we propose a distillation-based framework that incorporates a new subcategory of Positive-Unlabeled learning. In particular, we shall assume that a small subset of any given noisy dataset is known to have correct labels, which we treat as "positive", while the remaining noisy subset is treated as "unlabeled". Our framework consists of the following two components: (1) We shall generate, via iterative updates, an augmented clean subset with additional reliable "positive" samples filtered from "unlabeled" samples; (2) We shall train a teacher model on this larger augmented clean set. With the guidance of the teacher model, we then train a student model on the whole dataset. Experiments were conducted on the CIFAR-10 dataset with synthetic label noise at multiple noise levels for both symmetric and asymmetric noise. The results show that our framework generally outperforms at medium to high noise levels. We also evaluated our framework on Clothing1M, a real-world noisy dataset, and we achieved 2.94% improvement in accuracy over existing state-of-the-art methods.
This paper presents an efficient multi-scale vision Transformer, called ResT, that capably served as a general-purpose backbone for image recognition. Unlike existing Transformer methods, which employ standard Transformer blocks to tackle raw images with a fixed resolution, our ResT have several advantages: (1) A memory-efficient multi-head self-attention is built, which compresses the memory by a simple depth-wise convolution, and projects the interaction across the attention-heads dimension while keeping the diversity ability of multi-heads; (2) Position encoding is constructed as spatial attention, which is more flexible and can tackle with input images of arbitrary size without interpolation or fine-tune; (3) Instead of the straightforward tokenization at the beginning of each stage, we design the patch embedding as a stack of overlapping convolution operation with stride on the 2D-reshaped token map. We comprehensively validate ResT on image classification and downstream tasks. Experimental results show that the proposed ResT can outperform the recently state-of-the-art backbones by a large margin, demonstrating the potential of ResT as strong backbones. The code and models will be made publicly available at https://github.com/wofmanaf/ResT.
Deep Neural Networks (DNNs) training can be difficult due to vanishing and exploding gradients during weight optimization through backpropagation. To address this problem, we propose a general class of Hamiltonian DNNs (H-DNNs) that stem from the discretization of continuous-time Hamiltonian systems and include several existing DNN architectures based on ordinary differential equations. Our main result is that a broad set of H-DNNs ensures non-vanishing gradients by design for an arbitrary network depth. This is obtained by proving that, using a semi-implicit Euler discretization scheme, the backward sensitivity matrices involved in gradient computations are symplectic. We also provide an upper-bound to the magnitude of sensitivity matrices and show that exploding gradients can be controlled through regularization. Finally, we enable distributed implementations of backward and forward propagation algorithms in H-DNNs by characterizing appropriate sparsity constraints on the weight matrices. The good performance of H-DNNs is demonstrated on benchmark classification problems, including image classification with the MNIST dataset.
Pretraining has sparked groundswell of interest in deep learning workflows to learn from limited data and improve generalization. While this is common for 2D image classification tasks, its application to 3D medical imaging tasks like chest CT interpretation is limited. We explore the idea of whether pretraining a model on realistic videos could improve performance rather than training the model from scratch, intended for tuberculosis type classification from chest CT scans. To incorporate both spatial and temporal features, we develop a hybrid convolutional neural network (CNN) and recurrent neural network (RNN) model, where the features are extracted from each axial slice of the CT scan by a CNN, these sequence of image features are input to a RNN for classification of the CT scan. Our model termed as ViPTT-Net, was trained on over 1300 video clips with labels of human activities, and then fine-tuned on chest CT scans with labels of tuberculosis type. We find that pretraining the model on videos lead to better representations and significantly improved model validation performance from a kappa score of 0.17 to 0.35, especially for under-represented class samples. Our best method achieved 2nd place in the ImageCLEF 2021 Tuberculosis - TBT classification task with a kappa score of 0.20 on the final test set with only image information (without using clinical meta-data). All codes and models are made available.
Hierarchical structures are popular in recent vision transformers, however, they require sophisticated designs and massive datasets to work well. In this paper, we explore the idea of nesting basic local transformers on non-overlapping image blocks and aggregating them in a hierarchical way. We find that the block aggregation function plays a critical role in enabling cross-block non-local information communication. This observation leads us to design a simplified architecture that requires minor code changes upon the original vision transformer. The benefits of the proposed judiciously-selected design are threefold: (1) NesT converges faster and requires much less training data to achieve good generalization on both ImageNet and small datasets like CIFAR; (2) when extending our key ideas to image generation, NesT leads to a strong decoder that is 8$\times$ faster than previous transformer-based generators; and (3) we show that decoupling the feature learning and abstraction processes via this nested hierarchy in our design enables constructing a novel method (named GradCAT) for visually interpreting the learned model. Source code is available https://github.com/google-research/nested-transformer.
We propose Predict then Interpolate (PI), a simple algorithm for learning correlations that are stable across environments. The algorithm follows from the intuition that when using a classifier trained on one environment to make predictions on examples from another environment, its mistakes are informative as to which correlations are unstable. In this work, we prove that by interpolating the distributions of the correct predictions and the wrong predictions, we can uncover an oracle distribution where the unstable correlation vanishes. Since the oracle interpolation coefficients are not accessible, we use group distributionally robust optimization to minimize the worst-case risk across all such interpolations. We evaluate our method on both text classification and image classification. Empirical results demonstrate that our algorithm is able to learn robust classifiers (outperforms IRM by 23.85% on synthetic environments and 12.41% on natural environments). Our code and data are available at https://github.com/YujiaBao/Predict-then-Interpolate.
The proliferation of optical, electron, and scanning probe microscopies gives rise to large volumes of imaging data of objects as diversified as cells, bacteria, pollen, to nanoparticles and atoms and molecules. In most cases, the experimental data streams contain images having arbitrary rotations and translations within the image. At the same time, for many cases, small amounts of labeled data are available in the form of prior published results, image collections, and catalogs, or even theoretical models. Here we develop an approach that allows generalizing from a small subset of labeled data with a weak orientational disorder to a large unlabeled dataset with a much stronger orientational (and positional) disorder, i.e., it performs a classification of image data given a small number of examples even in the presence of a distribution shift between the labeled and unlabeled parts. This approach is based on the semi-supervised rotationally invariant variational autoencoder (ss-rVAE) model consisting of the encoder-decoder "block" that learns a rotationally (and translationally) invariant continuous latent representation of data and a classifier that encodes data into a finite number of discrete classes. The classifier part of the trained ss-rVAE inherits the rotational (and translational) invariances and can be deployed independently of the other parts of the model. The performance of the ss-rVAE is illustrated using the synthetic data sets with known factors of variation. We further demonstrate its application for experimental data sets of nanoparticles, creating nanoparticle libraries and disentangling the representations defining the physical factors of variation in the data. The code reproducing the results is available at https://github.com/ziatdinovmax/Semi-Supervised-VAE-nanoparticles.
Large pre-trained language models achieve state-of-the-art results when fine-tuned on downstream NLP tasks. However, they almost exclusively focus on text-only representation, while neglecting cell-level layout information that is important for form image understanding. In this paper, we propose a new pre-training approach, StructuralLM, to jointly leverage cell and layout information from scanned documents. Specifically, we pre-train StructuralLM with two new designs to make the most of the interactions of cell and layout information: 1) each cell as a semantic unit; 2) classification of cell positions. The pre-trained StructuralLM achieves new state-of-the-art results in different types of downstream tasks, including form understanding (from 78.95 to 85.14), document visual question answering (from 72.59 to 83.94) and document image classification (from 94.43 to 96.08).
We present FedScale, a federated learning (FL) benchmarking suite with realistic datasets and a scalable runtime to enable reproducible FL research. FedScale datasets encompass a wide range of critical FL tasks, ranging from image classification and object detection to language modeling and speech recognition. Each dataset comes with a unified evaluation protocol using real-world data splits and evaluation metrics. To reproduce realistic FL behavior, FedScale contains a scalable and extensible runtime. It provides high-level APIs to implement FL algorithms, deploy them at scale across diverse hardware and software backends, and evaluate them at scale, all with minimal developer efforts. We combine the two to perform systematic benchmarking experiments and highlight potential opportunities for heterogeneity-aware co-optimizations in FL. FedScale is open-source and actively maintained by contributors from different institutions at http://fedscale.ai. We welcome feedback and contributions from the community.
The learning rate (LR) schedule is one of the most important hyper-parameters needing careful tuning in training DNNs. However, it is also one of the least automated parts of machine learning systems and usually costs significant manual effort and computing. Though there are pre-defined LR schedules and optimizers with adaptive LR, they introduce new hyperparameters that need to be tuned separately for different tasks/datasets. In this paper, we consider the question: Can we automatically tune the LR over the course of training without human involvement? We propose an efficient method, AutoLRS, which automatically optimizes the LR for each training stage by modeling training dynamics. AutoLRS aims to find an LR applied to every $\tau$ steps that minimizes the resulted validation loss. We solve this black-box optimization on the fly by Bayesian optimization (BO). However, collecting training instances for BO requires a system to evaluate each LR queried by BO's acquisition function for $\tau$ steps, which is prohibitively expensive in practice. Instead, we apply each candidate LR for only $\tau'\ll\tau$ steps and train an exponential model to predict the validation loss after $\tau$ steps. This mutual-training process between BO and the loss-prediction model allows us to limit the training steps invested in the BO search. We demonstrate the advantages and the generality of AutoLRS through extensive experiments of training DNNs for tasks from diverse domains using different optimizers. The LR schedules auto-generated by AutoLRS lead to a speedup of $1.22\times$, $1.43\times$, and $1.5\times$ when training ResNet-50, Transformer, and BERT, respectively, compared to the LR schedules in their original papers, and an average speedup of $1.31\times$ over state-of-the-art heavily-tuned LR schedules.
We propose an efficient and straightforward method for compressing deep convolutional neural networks (CNNs) that uses basis filters to represent the convolutional layers, and optimizes the performance of the compressed network directly in the basis space. Specifically, any spatial convolution layer of the CNN can be replaced by two successive convolution layers: the first is a set of three-dimensional orthonormal basis filters, followed by a layer of one-dimensional filters that represents the original spatial filters in the basis space. We jointly fine-tune both the basis and the filter representation to directly mitigate any performance loss due to the truncation. Generality of the proposed approach is demonstrated by applying it to several well known deep CNN architectures and data sets for image classification and object detection. We also present the execution time and power usage at different compression levels on the Xavier Jetson AGX processor.