Data Augmentation Algorithm Based on Local Dynamic Transformation

PurposeThe purpose of this study is to explore the potential of trainable activation functions to enhance the performance of deep neural networks, specifically ResNet architectures, in the task of image classification. By introducing activation functions that adapt during training, the authors aim to determine whether such flexibility can lead to improved learning outcomes and generalization capabilities compared to static activation functions like ReLU. This research seeks to provide insights into how dynamic nonlinearities might influence deep learning models' efficiency and accuracy in handling complex image data sets.Design/methodology/approachThis research integrates three novel trainable activation functions – CosLU, DELU and ReLUN – into various ResNet-n architectures, where “n” denotes the number of convolutional layers. Using CIFAR-10 and CIFAR-100 data sets, the authors conducted a comparative study to assess the impact of these functions on image classification accuracy. The approach included modifying the traditional ResNet models by replacing their static activation functions with the trainable variants, allowing for dynamic adaptation during training. The performance was evaluated based on accuracy metrics and loss profiles across different network depths.FindingsThe findings indicate that trainable activation functions, particularly CosLU, can significantly enhance the performance of deep learning models, outperforming the traditional ReLU in deeper network configurations on the CIFAR-10 data set. CosLU showed the highest improvement in accuracy, whereas DELU and ReLUN offered varying levels of performance enhancements. These functions also demonstrated potential in reducing overfitting and improving model generalization across more complex data sets like CIFAR-100, suggesting that the adaptability of activation functions plays a crucial role in the training dynamics of deep neural networks.Originality/valueThis study contributes to the field of deep learning by introducing and evaluating the impact of three novel trainable activation functions within widely used ResNet architectures. Unlike previous works that primarily focused on static activation functions, this research demonstrates that incorporating trainable nonlinearities can lead to significant improvements in model performance and adaptability. The introduction of CosLU, DELU and ReLUN provides a new pathway for enhancing the flexibility and efficiency of neural networks, potentially setting a new standard for future deep learning applications in image classification and beyond.

Capsule networks exhibit the potential to enhance computer vision tasks through their utilization of equivariance for capturing spatial relationships. However, the broader adoption of these networks has been impeded by the computational complexity of their routing mechanism and shallow backbone model. To address these challenges, this paper introduces an innovative hybrid architecture that seamlessly integrates a pretrained backbone model with a task-specific capsule head (CapsHead). Our methodology is extensively evaluated across a range of classification and segmentation tasks, encompassing diverse datasets. The empirical findings robustly underscore the efficacy and practical feasibility of our proposed approach in real-world vision applications. Notably, our approach yields substantial 3.45% and 6.24% enhancement in linear evaluation on the CIFAR10 dataset and segmentation on the VOC2012 dataset, respectively, compared to baselines that do not incorporate the capsule head. This research offers a noteworthy contribution by not only advancing the application of capsule networks, but also mitigating their computational complexities. The results substantiate the feasibility of our hybrid architecture, thereby paving the way for a wider integration of capsule networks into various computer vision tasks.

Deep neural networks are a branch in machine learning that has seen a meteoric rise in popularity due to its powerful abilities to represent and model high-level abstractions in highly complex data. One area in deep neural networks that are ripe for exploration is neural connectivity formation. A pivotal study on the brain tissue of rats found that synaptic formation for specific functional connectivity in neocortical neural microcircuits can be surprisingly well modeled and predicted as a random formation. Motivated by this intriguing finding, we introduce the concept of StochasticNet where deep neural networks are formed via stochastic connectivity between neurons. As a result, any type of deep neural networks can be formed as a StochasticNet by allowing the neuron connectivity to be stochastic. Stochastic synaptic formations in a deep neural network architecture can allow for efficient utilization of neurons for performing specific tasks. To evaluate the feasibility of such a deep neural network architecture, we train a StochasticNet using four different image datasets (CIFAR-10, MNIST, SVHN, and STL-10). Experimental results show that a StochasticNet using less than half the number of neural connections as a conventional deep neural network achieves comparable accuracy and reduces overfitting on the CIFAR-10, MNIST, and SVHN data sets. Interestingly, StochasticNet with less than half the number of neural connections, achieved a higher accuracy (relative improvement in test error rate of $\sim 6$ % compared to ConvNet) on the STL-10 data set than a conventional deep neural network. Finally, the StochasticNets have faster operational speeds while achieving better or similar accuracy performances.

Dropout is a well-known regularization method by sampling a sub-network from a larger deep neural network and training different sub-networks on different subsets of the data. Inspired by the dropout concept, we propose EDropout as an energy-based framework for pruning neural networks in classification tasks. In this approach, a set of binary pruning state vectors (population) represents a set of corresponding sub-networks from an arbitrary original neural network. An energy loss function assigns a scalar energy loss value to each pruning state. The energy-based model (EBM) stochastically evolves the population to find states with lower energy loss. The best pruning state is then selected and applied to the original network. Similar to dropout, the kept weights are updated using backpropagation in a probabilistic model. The EBM again searches for better pruning states and the cycle continuous. This procedure is a switching between the energy model, which manages the pruning states, and the probabilistic model, which updates the kept weights, in each iteration. The population can dynamically converge to a pruning state. This can be interpreted as dropout leading to pruning the network. From an implementation perspective, unlike most of the pruning methods, EDropout can prune neural networks without manually modifying the network architecture code. We have evaluated the proposed method on different flavors of ResNets, AlexNet, l1 pruning, ThinNet, ChannelNet, and SqueezeNet on the Kuzushiji, Fashion, CIFAR-10, CIFAR-100, Flowers, and ImageNet data sets, and compared the pruning rate and classification performance of the models. The networks trained with EDropout on average achieved a pruning rate of more than 50% of the trainable parameters with approximately < 5% and < 1% drop of Top-1 and Top-5 classification accuracy, respectively.

Large-scale image retrieval using transductive support vector machines

The capsule network is a novel architecture to encode feature attributes and spatial relationships of an image. By using the dynamic routing (DR) algorithm, a capsule network (CapsNet) model can be trained. However, the original three-layer CapsNet with the DR algorithm performs poorly on complex data sets, such as FashionMNIST, CIFAR-10, and CIFAR-100. This deficiency limits the wider application of capsule networks. In this article, we propose a deep capsule network model combined with a U-Net preprocessing module (DCN-UN). Local connection and weight-sharing strategies are adopted from convolutional neural networks to design a convolutional capsule layer in the DCN-UN model. This allows considerably reducing the number of parameters in the network model. Moreover, a greedy strategy is incorporated into the design of a mask DR (MDR) algorithm to improve the performance of network models. DCN-UN requires up to five times fewer parameters compared with the original CapsNet and other CapsNet-based models. The performance improvement achieved by the DCN-UN model with the MDR algorithm over the original CapsNet model with the DR algorithm is approximately 12% and 17% on the CIFAR-10 and CIFAR-100 data sets, respectively. The experimental results confirm that the proposed DCN-UN model allows preserving advantages of image reconstruction and equivariance mechanism in capsule networks. Moreover, an efficient initialization method is explored to enhance training stability and avoid gradient explosion.

Weight pruning methods of deep neural networks (DNNs) have been demonstrated to achieve a good model pruning rate without loss of accuracy, thereby alleviating the significant computation/storage requirements of large-scale DNNs. Structured weight pruning methods have been proposed to overcome the limitation of irregular network structure and demonstrated actual GPU acceleration. However, in prior work, the pruning rate (degree of sparsity) and GPU acceleration are limited (to less than 50%) when accuracy needs to be maintained. In this work, we overcome these limitations by proposing a unified, systematic framework of structured weight pruning for DNNs. It is a framework that can be used to induce different types of structured sparsity, such as filterwise, channelwise, and shapewise sparsity, as well as nonstructured sparsity. The proposed framework incorporates stochastic gradient descent (SGD; or ADAM) with alternating direction method of multipliers (ADMM) and can be understood as a dynamic regularization method in which the regularization target is analytically updated in each iteration. Leveraging special characteristics of ADMM, we further propose a progressive, multistep weight pruning framework and a network purification and unused path removal procedure, in order to achieve higher pruning rate without accuracy loss. Without loss of accuracy on the AlexNet model, we achieve 2.58× and 3.65× average measured speedup on two GPUs, clearly outperforming the prior work. The average speedups reach 3.15× and 8.52× when allowing a moderate accuracy loss of 2%. In this case, the model compression for convolutional layers is 15.0× , corresponding to 11.93× measured CPU speedup. As another example, for the ResNet-18 model on the CIFAR-10 data set, we achieve an unprecedented 54.2× structured pruning rate on CONV layers. This is 32× higher pruning rate compared with recent work and can further translate into 7.6× inference time speedup on the Adreno 640 mobile GPU compared with the original, unpruned DNN model. We share our codes and models at the link http://bit.ly/2M0V7DO.

Federated learning (FL) is gradually becoming a key learning paradigm in Privacy-preserving Machine Learning (ML) systems. In FL, a large number of clients cooperate with a central server to learn a shared model without sharing their own data sets. However, since there is a great disparity between the client data sets, standard FL is often hard to tune and suffers from performance degradation due to the inharmony among local models. To this end, in this paper we propose a novel FL scheme, termed client reputation federated learning (CRFL), which dynamically assesses the reputation of the clients participating in FL. Our method leverages techniques from model explanation, and aims at precisely measure each client's impact to the global model. To be specific, we first calculate the saliency-weighted variance on pixelwise relevance scores as the quality factor of a single sample. Then we extract activation function values at the last hidden layer to compute the divergence factor of individual data set. Finally, the server integrates these two factors as an assessment of the client reputation. By leveraging such assessment, CRFL can dynamically adjust the weights of the clients in each aggregation round, thus leading to a significant improvement over the baseline method in terms of model accuracy and convergence rate. Intensive experiments are conducted on the MNIST and CIFAR-10 data sets, and experimental results demonstrate the efficacy of the proposed method.

Spiking neural networks (SNNs) are considered as a potential candidate to overcome current challenges, such as the high-power consumption encountered by artificial neural networks (ANNs); however, there is still a gap between them with respect to the recognition accuracy on various tasks. A conversion strategy was, thus, introduced recently to bridge this gap by mapping a trained ANN to an SNN. However, it is still unclear that to what extent this obtained SNN can benefit both the accuracy advantage from ANN and high efficiency from the spike-based paradigm of computation. In this article, we propose two new conversion methods, namely TerMapping and AugMapping. The TerMapping is a straightforward extension of a typical threshold-balancing method with a double-threshold scheme, while the AugMapping additionally incorporates a new scheme of augmented spike that employs a spike coefficient to carry the number of typical all-or-nothing spikes occurring at a time step. We examine the performance of our methods based on the MNIST, Fashion-MNIST, and CIFAR10 data sets. The results show that the proposed double-threshold scheme can effectively improve the accuracies of the converted SNNs. More importantly, the proposed AugMapping is more advantageous for constructing accurate, fast, and efficient deep SNNs compared with other state-of-the-art approaches. Our study, therefore, provides new approaches for further integration of advanced techniques in ANNs to improve the performance of SNNs, which could be of great merit to applied developments with spike-based neuromorphic computing.

In this paper, we propose a novel object detection framework named "Deep Regionlets" by establishing a bridge between deep neural networks and conventional detection schema for accurate generic object detection. Motivated by the abilities of regionlets for modeling object deformation and multiple aspect ratios, we incorporate regionlets into an end-to-end trainable deep learning framework. The deep regionlets framework consists of a region selection network and a deep regionlet learning module. Specifically, given a detection bounding box proposal, the region selection network provides guidance on where to select regions to learn the features from. The regionlet learning module focuses on local feature selection and transformation to alleviate local variations. To this end, we first realize non-rectangular region selection within the detection framework to accommodate variations in object appearance. Moreover, we design a "gating network" within the regionlet leaning module to enable soft regionlet selection and pooling. The Deep Regionlets framework is trained end-to-end without additional efforts. We perform ablation studies and conduct extensive experiments on the PASCAL VOC and Microsoft COCO datasets. The proposed framework outperforms state-of-the-art algorithms, such as RetinaNet and Mask R-CNN, even without additional segmentation labels.

Stochastic neurons are extremely efficient hardware for solving a large class of problems and usually come in two varieties - 'binary' where the neuronal state varies randomly between two values of ±1 and 'analog' where the neuronal state can randomly assume any value between -1 and +1. Both have their uses in neuromorphic computing and both can be implemented with low- or zero-energy-barrier nanomagnets whose random magnetization orientations in the presence of thermal noise encode the binary or analog state variables. In between these two classes isn-ary stochastic neurons, mainly ternary stochastic neurons (TSNs) whose state randomly assumes one of three values (-1,0,+1), which have proved to be efficient in pattern classification tasks such as recognizing handwritten digits from the MNIST data set or patterns from the CIFAR-10 data set. Here, we show how to implement a TSN with a zero-energy-barrier (shape isotropic) magnetostrictive nanomagnet subjected to uniaxial strain.

In the systems of industrial robotics and autonomous vehicles, instance segmentation is widely employed. However, manually labelling an object outline is time‐consuming. In order to reduce annotation costs, we present a weakly supervised instance segmentation method in this article. A deeply convolutional network is first used to construct multi‐scale feature maps for each object in the input image. After that, the encoder‐decoder framework with dynamic convolution is utilised to enhance model capacity and efficiency, while avoiding the issues of anchor design, proposal selection, and RoIAlign implementation. In particular, Dynamic Heads are used in the encoder to create dynamic convolution kernels, while Instance Heads are used in the decoder to provide the global feature map. With dynamic convolution, each instance can be segmented independently, reducing interference with other instances and improving segmentation accuracy. Under the supervision of projection loss and pixel point colour pairing loss, the contours of each object are finally outlined. On the PASCAL VOC and MS COCO datasets, the proposed method is competitive with more sophisticated approaches. In the VOC dataset, segmentation performance achieved 37.6% average precision with ResNet‐101 and FPN networks. The extensively visualised results demonstrate the effectiveness of the proposed encoder‐decoder framework with dynamic convolution.

Object detection is a fundamental task in computer vision, which is usually based on convolutional neural networks (CNNs). While it is difficult to be deployed in embedded devices due to the huge storage and computing consumptions, binary neural networks (BNNs) can execute object detection with limited resources. However, the extreme quantification in BNN causes diversity of feature representation loss, which eventually influences the object detection performance. In this paper, we propose a method balancing Information Retention and Deviation Control to achieve effective object detection, named IR-DC Net. On the one hand, we introduce the KL-Divergence to compose multiple entropy for maximizing the available information. On the other hand, we design a lightweight convolutional module to generate scale factors dynamically for minimizing the deviation between binary and real convolution. The experiments on PASCAL VOC, COCO2014, KITTI, and VisDrone datasets show that our method improved the accuracy in comparison with previous binary neural networks.

Convolutional neural network (CNN) is a state-of-the-art technique in machine learning and has achieved high accuracy in many computer vision tasks. However, the number of the parameters of the models is fast increasing for accuracy improvement; therefore, it requires more computation time and memory space for training and inference. Thus, compressing the model size and improving the inference speed has become an important issue. This paper focuses on filter pruning and NVIDIA sparse tensor core. Filter pruning is one of the model compression methods which uses a method that evaluates the importance of filters in the CNN model and removes the less important ones. NVIDIA sparse tensor core is the hardware support provided by NVIDIA Ampere GPU architecture. The sparse tensor core can speed up the matrix multiplication if the matrix has a structure that manifests as a 2:4 pattern. In this paper, we proposed a hybrid pruning metric to prune the CNN model. The hybrid pruning combines filter pruning and 2:4 pruning. We apply filter pruning to remove the redundant filters in convolutional layers to make the model smaller. Next, we use 2:4 pruning to prune the model according to a 2:4 pattern to utilize the sparse tensor core hardware for speedup. In this hybrid pruning situation, we have also proposed a hybrid ranking metric to decide the filter's importance during filter pruning. In hybrid ranking metric, we will preserve the filters that are important for both of the pruning steps. By considering both metrics, we can achieve higher accuracy than traditional filter prunings. We test our hybrid pruning algorithm on MNIST, SVHN, CIFAR-10 datasets using AlexNet. From our experiments, we concluded that our hybrid ranking method achieves better accuracy than the classic L1-norm metric and the output L1-norm metric. When we prune away 40 percent of filters in the model, our method has 2.8 %, 2.9 %, 2.7% higher accuracy than the classic L1-norm metric and the output L1-norm metric on these three datasets. Next, we evaluate the inference speed. We compare the hybrid pruning model with the models that result from either filter pruning or 2:4 pruning. We find that a hybrid pruning model can be 1.3x faster than the filter pruning model with similar accuracy.

Deep convolutional neural networks (CNNs) are widely used in many computer vision tasks. Since CNNs involve billions of computations, it is critical to reduce the resource /power consumption and improve parallelism. Compared with extensive researches on fixed point conversion for cost reduction, floating point customization has not been paid enough attention due to its higher cost than fixed point. This paper explores the customized floating point for both the training and inference of CNNs. 9-bit customized floating point is found sufficient for the training of ResNet-20 on CIFAR-10 dataset with less than 1% accuracy loss, which can also be applied to the inference of CNNs. With reduced bit-width, a computational unit (CU) based on Quad-Multiplier Packing is proposed to improve the resource efficiency of CNNs on FPGA. This design can save 87.5% DSP slices and 62.5% LUTs on Xilinx Kintex-7 platform compared to CU using 32-bit floating point. More CUs can be arranged on FPGA and higher throughput can be expected accordingly.