Large datasets have been crucial to the success of deep learning models in the recent years, which keep performing better as they are trained with more labelled data. While there have been sustained efforts to make these models more data-efficient, the potential benefit of understanding the data itself, is largely untapped. Specifically, focusing on object recognition tasks, we wonder if for common benchmark datasets we can do better than random subsets of the data and find a subset that can generalize on par with the full dataset when trained on. To our knowledge, this is the first result that can find notable redundancies in CIFAR-10 and ImageNet datasets (at least 10%). Interestingly, we observe semantic correlations between required and redundant images. We hope that our findings can motivate further research into identifying additional redundancies and exploiting them for more efficient training or data-collection.
How can prior knowledge on the transformation invariances of a domain be incorporated into the architecture of a neural network? We propose Equivariant Transformers (ETs), a family of differentiable image-to-image mappings that improve the robustness of models towards pre-defined continuous transformation groups. Through the use of specially-derived canonical coordinate systems, ETs incorporate functions that are equivariant by construction with respect to these transformations. We show empirically that ETs can be flexibly composed to improve model robustness towards more complicated transformation groups in several parameters. On a real-world image classification task, ETs improve the sample efficiency of ResNet classifiers, achieving relative improvements in error rate of up to 15% in the limited data regime while increasing model parameter count by less than 1%.
Sketching is a randomized dimensionality-reduction method that aims to preserve relevant information in large-scale datasets. Count sketch is a simple popular sketch which uses a randomized hash function to achieve compression. In this paper, we propose a novel extension known as Higher-order Count Sketch (HCS). While count sketch uses a single hash function, HCS uses multiple (smaller) hash functions for sketching. HCS reshapes the input (vector) data into a higher-order tensor and employs a tensor product of the random hash functions to compute the sketch. This results in an exponential saving (with respect to the order of the tensor) in the memory requirements of the hash functions, under certain conditions on the input data. Furthermore, when the input data itself has an underlying structure in the form of various tensor representations such as the Tucker decomposition, we obtain significant advantages. We derive efficient (approximate) computation of various tensor operations such as tensor products and tensor contractions directly on the sketched data. Thus, HCS is the first sketch to fully exploit the multi-dimensional nature of higher-order tensors. We apply HCS to tensorized neural networks where we replace fully connected layers with sketched tensor operations. We achieve nearly state of the art accuracy with significant compression on the image classification benchmark.
Classification of histologic patterns in lung adenocarcinoma is critical for determining tumor grade and treatment for patients. However, this task is often challenging due to the heterogeneous nature of lung adenocarcinoma and the subjective criteria for evaluation. In this study, we propose a deep learning model that automatically classifies the histologic patterns of lung adenocarcinoma on surgical resection slides. Our model uses a convolutional neural network to identify regions of neoplastic cells, then aggregates those classifications to infer predominant and minor histologic patterns for any given whole-slide image. We evaluated our model on an independent set of 143 whole-slide images. It achieved a kappa score of 0.525 and an agreement of 66.6% with three pathologists for classifying the predominant patterns, slightly higher than the inter-pathologist kappa score of 0.485 and agreement of 62.7% on this test set. All evaluation metrics for our model and the three pathologists were within 95% confidence intervals of agreement. If confirmed in clinical practice, our model can assist pathologists in improving classification of lung adenocarcinoma patterns by automatically pre-screening and highlighting cancerous regions prior to review. Our approach can be generalized to any whole-slide image classification task, and code is made publicly available at https://github.com/BMIRDS/deepslide.
Computations for the softmax function are significantly expensive when the number of output classes is large. In this paper, we present a novel softmax inference speedup method, Doubly Sparse Softmax (DS-Softmax), that leverages sparse mixture of sparse experts to efficiently retrieve top-k classes. Different from most existing methods that require and approximate a fixed softmax, our method is learning-based and can adapt softmax weights for a better inference speedup. In particular, our method learns a two-level hierarchy which divides entire output class space into several partially overlapping experts. Each expert is sparse and only contains a subset of output classes. To find top-k classes, a sparse mixture enables us to find the most probable expert quickly, and the sparse expert enables us to search within a small-scale softmax. We empirically conduct evaluation on several real-world tasks, including neural machine translation, language modeling and image classification, and demonstrate that significant computation reductions can be achieved at no performance loss.
Deep neural networks have seen tremendous success for different modalities of data including images, videos, and speech. This success has led to their deployment in mobile and embedded systems for real-time applications. However, making repeated inferences using deep networks on embedded systems poses significant challenges due to constrained resources (e.g., energy and computing power). To address these challenges, we develop a principled co-design approach. Building on prior work, we develop a formalism referred to as Coarse-to-Fine Networks (C2F Nets) that allow us to employ classifiers of varying complexity to make predictions. We propose a principled optimization algorithm to automatically configure C2F Nets for a specified trade-off between accuracy and energy consumption for inference. The key idea is to select a classifier on-the-fly whose complexity is proportional to the hardness of the input example: simple classifiers for easy inputs and complex classifiers for hard inputs. We perform comprehensive experimental evaluation using four different C2F Net architectures on multiple real-world image classification tasks. Our results show that optimized C2F Net can reduce the Energy Delay Product (EDP) by 27 to 60 percent with no loss in accuracy when compared to the baseline solution, where all predictions are made using the most complex classifier in C2F Net.
Deep neural network models have recently draw lots of attention, as it consistently produce impressive results in many computer vision tasks such as image classification, object detection, etc. However, interpreting such model and show the reason why it performs quite well becomes a challenging question. In this paper, we propose a novel method to interpret the neural network models with attention mechanism. Inspired by the heatmap visualization, we analyze the relation between classification accuracy with the attention based heatmap. An improved attention based method is also included and illustrate that a better classifier can be interpreted by the attention based heatmap.
Compressed domain image classification performs classification directly on compressive measurements acquired from the single-pixel camera, bypassing the image reconstruction step. It is of great importance for extending high-speed object detection and classification beyond the visible spectrum in a cost-effective manner especially for resource-limited platforms. Previous neural network methods require training a dedicated neural network for each different measurement rate (MR), which is costly in computation and storage. In this work, we develop an efficient training scheme that provides a neural network with dynamic-rate property, where a single neural network is capable of classifying over any MR within the range of interest with a given sensing matrix. This training scheme uses only a few selected MRs for training and the trained neural network is valid over the full range of MRs of interest. We demonstrate the performance of the dynamic-rate neural network on datasets of MNIST, CIFAR-10, Fashion-MNIST, COIL-100, and show that it generates approximately equal performance at each MR as that of a single-rate neural network valid only for one MR. Robustness to noise of the dynamic-rate model is also demonstrated. The dynamic-rate training scheme can be regarded as a general approach compatible with different types of sensing matrices, various neural network architectures, and is a valuable step towards wider adoption of compressive inference techniques and other compressive sensing related tasks via neural networks.
We propose a new layer in Convolutional Neural Networks (CNNs) to increase their robustness to several types of noise perturbations of the input images. We call this a push-pull layer and compute its response as the combination of two half-wave rectified convolutions, with kernels of opposite polarity. It is based on a biologically-motivated non-linear model of certain neurons in the visual system that exhibit a response suppression phenomenon, known as push-pull inhibition. We validate our method by substituting the first convolutional layer of the LeNet-5 and WideResNet architectures with our push-pull layer. We train the networks on nonperturbed training images from the MNIST, CIFAR-10 and CIFAR-100 data sets, and test on images perturbed by noise that is unseen by the training process. We demonstrate that our push-pull layers contribute to a considerable improvement in robustness of classification of images perturbed by noise, while maintaining state-of-the-art performance on the original image classification task.
Image classification is a challenging problem which aims to identify the category of object in the image. In recent years, deep Convolutional Neural Networks (CNNs) have been applied to handle this task, and impressive improvement has been achieved. However, some research showed the output of CNNs can be easily altered by adding relatively small perturbations to the input image, such as modifying few pixels. Recently, Capsule Networks (CapsNets) are proposed, which can help eliminating this limitation. Experiments on MNIST dataset revealed that capsules can better characterize the features of object than CNNs. But it's hard to find a suitable quantitative method to compare the generalization ability of CNNs and CapsNets. In this paper, we propose a new image classification task called Top-2 classification to evaluate the generalization ability of CNNs and CapsNets. The models are trained on single label image samples same as the traditional image classification task. But in the test stage, we randomly concatenate two test image samples which contain different labels, and then use the trained models to predict the top-2 labels on the unseen newly-created two label image samples. This task can provide us precise quantitative results to compare the generalization ability of CNNs and CapsNets. Back to the CapsNet, because it uses Full Connectivity (FC) mechanism among all capsules, it requires many parameters. To reduce the number of parameters, we introduce the Parameter-Sharing (PS) mechanism between capsules. Experiments on five widely used benchmark image datasets demonstrate the method significantly reduces the number of parameters, without losing the effectiveness of extracting features. Further, on the Top-2 classification task, the proposed PS CapsNets obtain impressive higher accuracy compared to the traditional CNNs and FC CapsNets by a large margin.
We develop a unified model, known as MgNet, that simultaneously recovers some convolutional neural networks (CNN) for image classification and multigrid (MG) methods for solving discretized partial differential equations (PDEs). This model is based on close connections that we have observed and uncovered between the CNN and MG methodologies. For example, pooling operation and feature extraction in CNN correspond directly to restriction operation and iterative smoothers in MG, respectively. As the solution space is often the dual of the data space in PDEs, the analogous concept of feature space and data space (which are dual to each other) is introduced in CNN. With such connections and new concept in the unified model, the function of various convolution operations and pooling used in CNN can be better understood. As a result, modified CNN models (with fewer weights and hyper parameters) are developed that exhibit competitive and sometimes better performance in comparison with existing CNN models when applied to both CIFAR-10 and CIFAR-100 data sets.
It is intuitive that NLP tasks for logographic languages like Chinese should benefit from the use of the glyph information in those languages. However, due to the lack of rich pictographic evidence in glyphs and the weak generalization ability of standard computer vision models on character data, an effective way to utilize the glyph information remains to be found. In this paper, we address this gap by presenting Glyce, the glyph-vectors for Chinese character representations. We make three major innovations: (1) We use historical Chinese scripts (e.g., bronzeware script, seal script, traditional Chinese, etc) to enrich the pictographic evidence in characters; (2) We design CNN structures (called tianzege-CNN) tailored to Chinese character image processing; and (3) We use image-classification as an auxiliary task in a multi-task learning setup to increase the model's ability to generalize. We show that glyph-based models are able to consistently outperform word/char ID-based models in a wide range of Chinese NLP tasks. We are able to set new state-of-the-art results for a variety of Chinese NLP tasks, including tagging (NER, CWS, POS), sentence pair classification, single sentence classification tasks, dependency parsing, and semantic role labeling. For example, the proposed model achieves an F1 score of 80.6 on the OntoNotes dataset of NER, +1.5 over BERT; it achieves an almost perfect accuracy of 99.8\% on the Fudan corpus for text classification. Code found at https://github.com/ShannonAI/glyce.
Many real-world problems exhibit the coexistence of multiple types of heterogeneity, such as view heterogeneity (i.e., multi-view property) and task heterogeneity (i.e., multi-task property). For example, in an image classification problem containing multiple poses of the same object, each pose can be considered as one view, and the detection of each type of object can be treated as one task. Furthermore, in some problems, the data type of multiple views might be different. In a web classification problem, for instance, we might be provided an image and text mixed data set, where the web pages are characterized by both images and texts. A common strategy to solve this kind of problem is to leverage the consistency of views and the relatedness of tasks to build the prediction model. In the context of deep neural network, multi-task relatedness is usually realized by grouping tasks at each layer, while multi-view consistency is usually enforced by finding the maximal correlation coefficient between views. However, there is no existing deep learning algorithm that jointly models task and view dual heterogeneity, particularly for a data set with multiple modalities (text and image mixed data set or text and video mixed data set, etc.). In this paper, we bridge this gap by proposing a deep multi-task multi-view learning framework that learns a deep representation for such dual-heterogeneity problems. Empirical studies on multiple real-world data sets demonstrate the effectiveness of our proposed Deep-MTMV algorithm.
We address two questions for training a convolutional neural network (CNN) for hyperspectral image classification: i) is it possible to build a pre-trained network? and ii) is the pre-training effective in furthering the performance? To answer the first question, we have devised an approach that pre-trains a network on multiple source datasets that differ in their hyperspectral characteristics and fine-tunes on a target dataset. This approach effectively resolves the architectural issue that arises when transferring meaningful information between the source and the target networks. To answer the second question, we carried out several ablation experiments. Based on the experimental results, a network trained from scratch performs as good as a network fine-tuned from a pre-trained network. However, we observed that pre-training the network has its own advantage in achieving better performances when deeper networks are required.
High-dimensional representations often have a lower dimensional underlying structure. This is particularly the case in many decision making settings. For example, when the representation of actions is generated from a deep neural network, it is reasonable to expect a low-rank structure whereas conventional structures like sparsity are not valid anymore. Subspace recovery methods, such as Principle Component Analysis (PCA) can find the underlying low-rank structures in the feature space and reduce the complexity of the learning tasks. In this work, we propose Projected Stochastic Linear Bandit (PSLB), an algorithm for high dimensional stochastic linear bandits (SLB) when the representation of actions has an underlying low-dimensional subspace structure. PSLB deploys PCA based projection to iteratively find the low rank structure in SLBs. We show that deploying projection methods assures dimensionality reduction and results in a tighter regret upper bound that is in terms of the dimensionality of the subspace and its properties, rather than the dimensionality of the ambient space. We modify the image classification task into the SLB setting and empirically show that, when a pre-trained DNN provides the high dimensional feature representations, deploying PSLB results in significant reduction of regret and faster convergence to an accurate model compared to state-of-art algorithm.
We present ADMM-Softmax, an alternating direction method of multipliers (ADMM) for solving multinomial logistic regression (MLR) problems. Our method is geared toward supervised classification tasks with many examples and features. It decouples the nonlinear optimization problem in MLR into three steps that can be solved efficiently. In particular, each iteration of ADMM-Softmax consists of a linear least-squares problem, a set of independent small-scale smooth, convex problems, and a trivial dual variable update. Solution of the least-squares problem can be be accelerated by pre-computing a factorization or preconditioner, and the separability in the smooth, convex problem can be easily parallelized across examples. For two image classification problems, we demonstrate that ADMM-Softmax leads to improved generalization compared to a Newton-Krylov, a quasi Newton, and a stochastic gradient descent method.
Normalization layers are a staple in state-of-the-art deep neural network architectures. They are widely believed to stabilize training, enable higher learning rate, accelerate convergence and improve generalization, though the reason for their effectiveness is still an active research topic. In this work, we challenge the commonly-held beliefs by showing that none of the perceived benefits is unique to normalization. Specifically, we propose fixed-update initialization (Fixup), an initialization motivated by solving the exploding and vanishing gradient problem at the beginning of training via properly rescaling a standard initialization. We find training residual networks with Fixup to be as stable as training with normalization -- even for networks with 10,000 layers. Furthermore, with proper regularization, Fixup enables residual networks without normalization to achieve state-of-the-art performance in image classification and machine translation.
Capsule Networks preserve the hierarchical spatial relationships between objects, and thereby bears a potential to surpass the performance of traditional Convolutional Neural Networks (CNNs) in performing tasks like image classification. A large body of work has explored adversarial examples for CNNs, but their effectiveness on Capsule Networks has not yet been well studied. In our work, we perform an analysis to study the vulnerabilities in Capsule Networks to adversarial attacks. These perturbations, added to the test inputs, are small and imperceptible to humans, but can fool the network to mispredict. We propose a greedy algorithm to automatically generate targeted imperceptible adversarial examples in a black-box attack scenario. We show that this kind of attacks, when applied to the German Traffic Sign Recognition Benchmark (GTSRB), mislead Capsule Networks. Moreover, we apply the same kind of adversarial attacks to a 5-layer CNN and a 9-layer CNN, and analyze the outcome, compared to the Capsule Networks to study differences in their behavior.
A convolutional layer in a Convolutional Neural Network (CNN) consists of many filters which apply convolution operation to the input, capture some special patterns and pass the result to the next layer. If the same patterns also occur at the deeper layers of the network, why wouldn't the same convolutional filters be used also in those layers? In this paper, we propose a CNN architecture, Layer Reuse Network (LruNet), where the convolutional layers are used repeatedly without the need of introducing new layers to get a better performance. This approach introduces several advantages: (i) Considerable amount of parameters are saved since we are reusing the layers instead of introducing new layers, (ii) the Memory Access Cost (MAC) can be reduced since reused layer parameters can be fetched only once, (iii) the number of nonlinearities increases with layer reuse, and (iv) reused layers get gradient updates from multiple parts of the network. The proposed approach is evaluated on CIFAR-10, CIFAR-100 and Fashion-MNIST datasets for image classification task, and layer reuse improves the performance by 5.14%, 5.85% and 2.29%, respectively. The source code and pretrained models are publicly available.
A commonly cited inefficiency of neural network training by back-propagation is the update locking problem: each layer must wait for the signal to propagate through the full network before updating. Several alternatives that can alleviate this issue have been proposed. In this context, we consider a simpler, but more effective, substitute that uses minimal feedback, which we call Decoupled Greedy Learning (DGL). It is based on a greedy relaxation of the joint training objective, recently shown to be effective in the context of Convolutional Neural Networks (CNNs) on large-scale image classification. We consider an optimization of this objective that permits us to decouple the layer training, allowing for layers or modules in networks to be trained with a potentially linear parallelization in layers. With the use of a replay buffer we show this approach can be extended to asynchronous settings, where modules can operate with possibly large communication delays. We show theoretically and empirically that this approach converges. Then, we empirically find that it can lead to better generalization than sequential greedy optimization. We demonstrate the effectiveness of DGL against alternative approaches on the CIFAR-10 dataset and on the large-scale ImageNet dataset.