Neural Ordinary Differential Equations (ODEs) have been successful in various applications due to their continuous nature and parameter-sharing efficiency. However, these unique characteristics also introduce challenges in training, particularly with respect to gradient computation accuracy and convergence analysis. In this paper, we address these challenges by investigating the impact of activation functions. We demonstrate that the properties of activation functions—specifically smoothness and nonlinearity—are critical to the training dynamics. Smooth activation functions guarantee globally unique solutions for both forward and backward ODEs, while sufficient nonlinearity is essential for maintaining the spectral properties of the Neural Tangent Kernel (NTK) during training. Together, these properties enable us to establish the global convergence of Neural ODEs under gradient descent in overparameterized regimes. Our theoretical findings are validated by numerical experiments, which not only support our analysis but also provide practical guidelines for scaling Neural ODEs, potentially leading to faster training and improved performance in real-world applications.

Graph Neural Networks (GNNs) have shown remarkable success in molecular tasks, yet their interpretability remains challenging. Traditional model-level explanation methods like XGNN and GNNInterpreter often fail to identify valid substructures like rings, leading to questionable interpretability. This limitation stems from XGNN's atom-by-atom approach and GNNInterpreter's reliance on average graph embeddings, which overlook the essential structural elements crucial for molecules. To address these gaps, we introduce an innovative **M**otif-b**A**sed **G**NN **E**xplainer (MAGE) that uses motifs as fundamental units for generating explanations. Our approach begins with extracting potential motifs through a motif decomposition technique. Then, we utilize an attention-based learning method to identify class-specific motifs. Finally, we employ a motif-based graph generator for each class to create molecular graph explanations based on these class-specific motifs. This novel method not only incorporates critical substructures into the explanations but also guarantees their validity, yielding results that are human-understandable. Our proposed method's effectiveness is demonstrated through quantitative and qualitative assessments conducted on six real-world molecular datasets.

Deep learning models are trained with certain assumptions about the data during the development stage and then used for prediction in the deployment stage. It is important to reason about the trustworthiness of the model's predictions with unseen data during deployment. Existing methods for specifying and verifying traditional software are insufficient for this task, as they cannot handle the complexity of DNN model architecture and expected outcomes. In this work, we propose a novel technique that uses rules derived from neural network computations to infer data preconditions for a DNN model to determine the trustworthiness of its predictions. Our approach, DeepInfer involves introducing a novel abstraction for a trained DNN model that enables weakest precondition reasoning using Dijkstra's Predicate Transformer Semantics. By deriving rules over the inductive type of neural network abstract representation, we can overcome the matrix dimensionality issues that arise from the backward non-linear computation from the output layer to the input layer. We utilize the weakest precondition computation using rules of each kind of activation function to compute layer-wise precondition from the given postcondition on the final output of a deep neural network. We extensively evaluated DeepInfer on 29 real-world DNN models using four different datasets collected from five different sources and demonstrated the utility, effectiveness, and performance improvement over closely related work. DeepInfer efficiently detects correct and incorrect predictions of high-accuracy models with high recall (0.98) and high F-1 score (0.84) and has significantly improved over the prior technique, SelfChecker. The average runtime overhead of DeepInfer is low, 0.22 sec for all the unseen datasets. We also compared runtime overhead using the same hardware settings and found that DeepInfer is 3.27 times faster than SelfChecker, the state-of-the-art in this area.

Deep learning-based vulnerability detection has shown great performance and, in some studies, outperformed static analysis tools. However, the highest-performing approaches use token-based transformer models, which are not the most efficient to capture code semantics required for vulnerability detection. Classical program analysis techniques such as dataflow analysis can detect many types of bugs based on their root causes. In this paper, we propose to combine such causal-based vulnerability detection algorithms with deep learning, aiming to achieve more efficient and effective vulnerability detection. Specifically, we designed DeepDFA, a dataflow analysis-inspired graph learning framework and an embedding technique that enables graph learning to simulate dataflow computation. We show that DeepDFA is both performant and efficient. DeepDFA outperformed all non-transformer baselines. It was trained in 9 minutes, 75x faster than the highest-performing baseline model. When using only 50+ vulnerable and several hundreds of total examples as training data, the model retained the same performance as 100% of the dataset. DeepDFA also generalized to real-world vulnerabilities in DbgBench; it detected 8.7 out of 17 vulnerabilities on average across folds and was able to distinguish between patched and buggy versions, while the highest-performing baseline models did not detect any vulnerabilities. By combining DeepDFA with a large language model, we surpassed the state-of-the-art vulnerability detection performance on the Big-Vul dataset with 96.46 F1 score, 97.82 precision, and 95.14 recall. Our replication package is located at https://doi.org/10.6084/m9.figshare.21225413.

Numerous studies have shown that long non-coding RNAs (lncRNAs) can be the key factors in causing a variety of human diseases, such as cancers. Predicting the potential associations between lncRNAs and diseases can be highly beneficial for pathogenesis investigation and understanding the functionalities of lncRNAs in disease development. Existing computational methods for lncRNAs-disease association (LDA) prediction focuses on integrating heterogeneous information from given data including associations of various biological elements. However, many of them do not fully utilize the topological information obtained from graph-structured frameworks, such as heterogeneous graphs consisting of different biological variables including lncRNA, disease, miRNA, and others. We proposed a novel method, named HSELDA, that utilized a relational graph convolutional neural network based on a link prediction technique via subgraphs extraction to predict the potential lncRNAs-disease associations. In this method, we first constructed a heterogeneous graph consisting of three types of nodes: lncRNAs, miRNAs, and disease, with their known associations between each node pair as edges. Then, we extracted a list of local subgraphs around each target link, where the subgraphs preserve rich heuristic information related to link existence with the neighbor nodes directly connected to the target links. We employed a relational graph convolutional network to identify the most promising lncRNAs-diseases associations based on local subgraphs. The experimental results showed that HSELDA was superior to several state-of-the-art methods. Of note, our proposed model shows significant superiority over other compared baseline approaches in AUPRC.

Neural networks with wide layers have attracted significant attention due to their equivalence to Gaussian processes, enabling perfect fitting of training data while maintaining generalization performance, known as benign overfitting. However, existing results mainly focus on shallow or finite-depth networks, necessitating a comprehensive analysis of wide neural networks with infinite-depth layers, such as neural ordinary differential equations (ODEs) and deep equilibrium models (DEQs). In this paper, we specifically investigate the deep equilibrium model (DEQ), an infinite-depth neural network with shared weight matrices across layers. Our analysis reveals that as the width of DEQ layers approaches infinity, it converges to a Gaussian process, establishing what is known as the Neural Network and Gaussian Process (NNGP) correspondence. Remarkably, this convergence holds even when the limits of depth and width are interchanged, which is not observed in typical infinite-depth Multilayer Perceptron (MLP) networks. Furthermore, we demonstrate that the associated Gaussian vector remains non-degenerate for any pairwise distinct input data, ensuring a strictly positive smallest eigenvalue of the corresponding kernel matrix using the NNGP kernel. These findings serve as fundamental elements for studying the training and generalization of DEQs, laying the groundwork for future research in this area.

We introduce Meta-AdaM, a meta-learned adaptive optimizer with momentum, designed for few-shot learning tasks that pose significant challenges to deep learning models due to the limited number of labeled examples. Meta-learning has been successfully employed to address these challenges by transferring meta-learned prior knowledge to new tasks. Most existing works focus on meta-learning an optimal model initialization or an adaptive learning rate learner for rapid convergence. However, these approaches either neglect to consider weight-update history for the adaptive learning rate learner or fail to effectively integrate momentum for fast convergence, as seen in many-shot learning settings. To tackle these limitations, we propose a meta-learned learning rate learner that utilizes weight-update history as input to predict more appropriate learning rates for rapid convergence. Furthermore, for the first time, our approach incorporates momentum into the optimization process of few-shot learning via a double look-ahead mechanism, enabling rapid convergence similar to many-shot settings. Extensive experimental results on benchmark datasets demonstrate the effectiveness of the proposed Meta-AdaM.

It is desirable to develop efficient image rescaling methods to transmit digital images with different resolutions between devices and assure visual quality. In image downscaling, the inevitable loss of high-frequency information makes the reverse upscaling highly ill-posed. Recent approaches focus on joint learning of image downscaling and upscaling (e.g., rescaling). However, existing methods still fail to recover satisfactory high-frequency signals when upscaling. To solve it, we propose high-frequency flow (HfFlow), which learns the distribution of high-frequency signals during rescaling. HfFlow is an overall invertible framework with a conditional flow on the high-frequency space to compensate for the information lost during downscaling. To facilitate finding the optimal upscaling solution, we introduce a reference low-resolution (LR) manifold and propose a cross-entropy Gaussian loss (CGloss) to force the downscaled manifold closer to the reference LR manifold and simultaneously fulfill recovering missing details. HfFlow can be generalized to other scale transformation tasks such as image colorization with its excellent rescaling capacity. Qualitative and quantitative experimental evaluations demonstrate that HfFlow restores rich high-frequency details and outperforms state-of-the-art rescaling methods in PSNR, SSIM, and perceptual quality metrics.

We consider feature representation learning problem of molecular graphs. Graph Neural Networks have been widely used in feature representation learning of molecular graphs. However, most existing methods deal with molecular graphs individually while neglecting their connections, such as motif-level relationships. We propose a novel molecular graph representation learning method by constructing a heterogeneous motif graph to address this issue. In particular, we build a heterogeneous motif graph that contains motif nodes and molecular nodes. Each motif node corresponds to a motif extracted from molecules. Then, we propose a Heterogeneous Motif Graph Neural Network (HM-GNN) to learn feature representations for each node in the heterogeneous motif graph. Our heterogeneous motif graph also enables effective multi-task learning, especially for small molecular datasets. To address the potential efficiency issue, we propose to use an edge sampler, which can significantly reduce computational resources usage. The experimental results show that our model consistently outperforms previous state-of-the-art models. Under multi-task settings, the promising performances of our methods on combined datasets shed light on a new learning paradigm for small molecular datasets. Finally, we show that our model achieves similar performances with significantly less computational resources by using our edge sampler.

We consider the problem of representation learning for graph data. Given images are special cases of graphs with nodes lie on 2D lattices, graph embedding tasks have a natural correspondence with image pixel-wise prediction tasks such as segmentation. While encoder-decoder architectures like U-Nets have been successfully applied to image pixel-wise prediction tasks, similar methods are lacking for graph data. This is because pooling and up-sampling operations are not natural on graph data. To address these challenges, we propose novel graph pooling and unpooling operations. The gPool layer adaptively selects some nodes to form a smaller graph based on their scalar projection values. We further propose the gUnpool layer as the inverse operation of the gPool layer. Based on our proposed methods, we develop an encoder-decoder model, known as the graph U-Nets. Experimental results on node classification and graph classification tasks demonstrate that our methods achieve consistently better performance than previous models. Along this direction, we extend our methods by integrating attention mechanisms. Based on attention operators, we proposed attention-based pooling and unpooling layers, which can better capture graph topology information. The empirical results on graph classification tasks demonstrate the promising capability of our methods.

Pooling operations have shown to be effective on computer vision and natural language processing tasks. One challenge of performing pooling operations on graph data is the lack of locality that is not well-defined on graphs. Previous studies used global ranking methods to sample some of the important nodes, but most of them are not able to incorporate graph topology. In this work, we propose the topology-aware pooling (TAP) layer that explicitly considers graph topology. Our TAP layer is a two-stage voting process that selects more important nodes in a graph. It first performs local voting to generate scores for each node by attending each node to its neighboring nodes. The scores are generated locally such that topology information is explicitly considered. In addition, graph topology is incorporated in global voting to compute the importance score of each node globally in the entire graph. Altogether, the final ranking score for each node is computed by combining its local and global voting scores. To encourage better graph connectivity in the sampled graph, we propose to add a graph connectivity term to the computation of ranking scores. Results on graph classification tasks demonstrate that our methods achieve consistently better performance than previous methods.

Convolutional neural networks (CNNs) have shown great capability of solving various artificial intelligence tasks. However, the increasing model size has raised challenges in employing them in resource-limited applications. In this work, we propose to compress deep models by using channel-wise convolutions, which replace dense connections among feature maps with sparse ones in CNNs. Based on this novel operation, we build light-weight CNNs known as ChannelNets. ChannelNets use three instances of channel-wise convolutions; namely group channel-wise convolutions, depth-wise separable channel-wise convolutions, and the convolutional classification layer. Compared to prior CNNs designed for mobile devices, ChannelNets achieve a significant reduction in terms of the number of parameters and computational cost without loss in accuracy. Notably, our work represents an attempt to compress the fully-connected classification layer, which usually accounts for about 25 percent of total parameters in compact CNNs. Along this new direction, we investigate the behavior of our proposed convolutional classification layer and conduct detailed analysis. Based on our in-depth analysis, we further propose convolutional classification layers without weight-sharing. This new classification layer achieves a good trade-off between fully-connected classification layers and the convolutional classification layer. Experimental results on the ImageNet dataset demonstrate that ChannelNets achieve consistently better performance compared to prior methods.

Rectified linear units (ReLUs) are currently the most popular activation function used in neural networks. Although ReLUs can solve the gradient vanishing problem and accelerate training convergence, it suffers from the dying ReLU problem in which some neurons are never activated if the weights are not updated properly. In this work, we propose a novel activation function, known as the adaptive convolutional ReLU (ConvReLU), that can better mimic brain neuron activation behaviors and overcome the dying ReLU problem. With our novel parameter sharing scheme, ConvReLUs can be applied to convolution layers that allow each input neuron to be activated by different trainable thresholds without involving a large number of extra parameters. We employ the zero initialization scheme in ConvReLU to encourage trainable thresholds to be close to zero. Finally, we develop a partial replacement strategy that only replaces the ReLUs in the early layers of the network. This resolves the dying ReLU problem and retains sparse representations for linear classifiers. Experimental results demonstrate that our proposed ConvReLU has consistently better performance compared to ReLU, LeakyReLU, and PReLU. In addition, the partial replacement strategy is shown to be effective not only for our ConvReLU but also for LeakyReLU and PReLU.

Attention operators have been applied on both 1-D data like texts and higher-order data such as images and videos. Use of attention operators on high-order data requires flattening of the spatial or spatial-temporal dimensions into a vector, which is assumed to follow a multivariate normal distribution. This not only incurs excessive requirements on computational resources, but also fails to preserve structures in data. In this work, we propose to avoid flattening by assuming the data follow matrix-variate normal distributions. Based on this new view, we develop Kronecker attention operators (KAOs) that operate on high-order tensor data directly. More importantly, the proposed KAOs lead to dramatic reductions in computational resources. Experimental results show that our methods reduce the amount of required computational resources by a factor of hundreds, with larger factors for higher-dimensional and higher-order data. Results also show that networks with KAOs outperform models without attention, while achieving competitive performance as those with original attention operators.

Graph neural networks have shown significant success in the field of graph representation learning. Graph convolutions perform neighborhood aggregation and represent one of the most important graph operations. Nevertheless, one layer of these neighborhood aggregation methods only consider immediate neighbors, and the performance decreases when going deeper to enable larger receptive fields. Several recent studies attribute this performance deterioration to the over-smoothing issue, which states that repeated propagation makes node representations of different classes indistinguishable. In this work, we study this observation systematically and develop new insights towards deeper graph neural networks. First, we provide a systematical analysis on this issue and argue that the key factor compromising the performance significantly is the entanglement of representation transformation and propagation in current graph convolution operations. After decoupling these two operations, deeper graph neural networks can be used to learn graph node representations from larger receptive fields. We further provide a theoretical analysis of the above observation when building very deep models, which can serve as a rigorous and gentle description of the over-smoothing issue. Based on our theoretical and empirical analysis, we propose Deep Adaptive Graph Neural Network (DAGNN) to adaptively incorporate information from large receptive fields. A set of experiments on citation, co-authorship, and co-purchase datasets have confirmed our analysis and insights and demonstrated the superiority of our proposed methods.

Transposed convolutional layers have been widely used in a variety of deep models for up-sampling, including encoder-decoder networks for semantic segmentation and deep generative models for unsupervised learning. One of the key limitations of transposed convolutional operations is that they result in the so-called checkerboard problem. This is caused by the fact that no direct relationship exists among adjacent pixels on the output feature map. To address this problem, we propose the pixel transposed convolutional layer (PixelTCL) to establish direct relationships among adjacent pixels on the up-sampled feature map. Our method is based on a fresh interpretation of the regular transposed convolutional operation. The resulting PixelTCL can be used to replace any transposed convolutional layer in a plug-and-play manner without compromising the fully trainable capabilities of original models. The proposed PixelTCL may result in slight decrease in efficiency, but this can be overcome by an implementation trick. Experimental results on semantic segmentation demonstrate that PixelTCL can consider spatial features such as edges and shapes and yields more accurate segmentation outputs than transposed convolutional layers. When used in image generation tasks, our PixelTCL can largely overcome the checkerboard problem suffered by regular transposed convolutional operations.

We consider the problem of representation learning for graph data. Convolutional neural networks can naturally operate on images, but have significant challenges in dealing with graph data. Given images are special cases of graphs with nodes lie on 2D lattices, graph embedding tasks have a natural correspondence with image pixel-wise prediction tasks such as segmentation. While encoder-decoder architectures like U-Nets have been successfully applied on many image pixel-wise prediction tasks, similar methods are lacking for graph data. This is due to the fact that pooling and up-sampling operations are not natural on graph data. To address these challenges, we propose novel graph pooling (gPool) and unpooling (gUnpool) operations in this work. The gPool layer adaptively selects some nodes to form a smaller graph based on their scalar projection values on a trainable projection vector. We further propose the gUnpool layer as the inverse operation of the gPool layer. The gUnpool layer restores the graph into its original structure using the position information of nodes selected in the corresponding gPool layer. Based on our proposed gPool and gUnpool layers, we develop an encoder-decoder model on graph, known as the graph U-Nets. Our experimental results on node classification and graph classification tasks demonstrate that our methods achieve consistently better performance than previous models.

Attention operators have been widely applied in various fields, including computer vision, natural language processing, and network embedding learning. Attention operators on graph data enables learnable weights when aggregating information from neighboring nodes. However, graph attention operators (GAOs) consume excessive computational resources, preventing their applications on large graphs. In addition, GAOs belong to the family of soft attention, instead of hard attention, which has been shown to yield better performance. In this work, we propose novel hard graph attention operator~(hGAO) and channel-wise graph attention operator~(cGAO). hGAO uses the hard attention mechanism by attending to only important nodes. Compared to GAO, hGAO improves performance and saves computational cost by only attending to important nodes. To further reduce the requirements on computational resources, we propose the cGAO that performs attention operations along channels. cGAO avoids the dependency on the adjacency matrix, leading to dramatic reductions in computational resource requirements. Experimental results demonstrate that our proposed deep models with the new operators achieve consistently better performance. Comparison results also indicates that hGAO achieves significantly better performance than GAO on both node and graph embedding tasks. Efficiency comparison shows that our cGAO leads to dramatic savings in computational resources, making them applicable to large graphs.

Variational auto-encoder (VAE) is a powerful unsupervised learning framework for image generation. One drawback of VAE is that it generates blurry images due to its Gaussianity assumption and thus ℓ2 loss. To allow the generation of high quality images by VAE, we increase the capacity of decoder network by employing residual blocks and skip connections, which also enable efficient optimization. To overcome the limitation of ℓ2 loss, we propose to generate images in a multi-stage manner from coarse to fine. In the simplest case, the proposed multi-stage VAE divides the decoder into two components in which the second component generates refined images based on the course images generated by the first component. Since the second component is independent of the VAE model, it can employ other loss functions beyond the ℓ2 loss and different model architectures. The proposed framework can be easily generalized to contain more than two components. Experiment results on the MNIST and CelebA datasets demonstrate that the proposed multi-stage VAE can generate sharper images as compared to those from the original VAE.

With the development of graph convolutional networks (GCN), deep learning methods have started to be used on graph data. In additional to convolutional layers, pooling layers are another important components of deep learning. However, no effective pooling methods have been developed for graphs currently. In this work, we propose the graph pooling (gPool) layer, which employs a trainable projection vector to measure the importance of nodes in graphs. By selecting the k-most important nodes to form the new graph, gPool achieves the same objective as regular max pooling layers operation on images and texts. Another limitation of GCN when used on graph-based text representation tasks is that, GCNs do not consider the order information of nodes in graph. To address this limitation, we propose the hybrid convolutional (hConv) layer that combines GCN and regular convolutional operations. The hConv layer is capable of increasing receptive fields quickly and computing features automatically. Based on the proposed gPool and hConv layers, we develop new deep networks for text categorization tasks. Our experimental results show that the networks based on gPool and hConv layers achieves new state-of-the-art performance as compared to baseline methods.

Multi-modality data are widely used in clinical applications, such as tumor detection and brain disease diagnosis. Different modalities can usually provide complementary information, which commonly leads to improved performance. However, some modalities are commonly missing for some subjects due to various technical and practical reasons. As a result, multi-modality data are usually incomplete, raising the multi-modality missing data completion problem. In this work, we formulate the problem as a conditional image generation task and propose an encoder-decoder deep neural network to tackle this problem. Specifically, the model takes the existing modality as input and generates the missing modality. By employing an auxiliary adversarial loss, our model is able to generate high-quality missing modality images. At the same time, we propose to incorporate the available category information of subjects in training to enable the model to generate more informative images. We evaluate our method on the Alzheimer's Disease Neuroimaging Initiative~(ADNI) database, where positron emission tomography~(PET) modalities are missing. Experimental results show that the trained network can generate high-quality PET modalities based on existing magnetic resonance imaging~(MRI) modalities, and provide complementary information to improve the detection and tracking of the Alzheimer's disease. Our results also show that the proposed methods generate higher quality images than baseline methods as measured by various image quality statistics.

Deep learning methods have shown great success in pixel-wise prediction tasks. One of the most popular methods employs an encoder-decoder network in which deconvolutional layers are used for up-sampling feature maps. However, a key limitation of the deconvolutional layer is that it suffers from the checkerboard artifact problem, which harms the prediction accuracy. This is caused by the independency among adjacent pixels on the output feature maps. Previous work only solved the checkerboard artifact issue of deconvolutional layers in the 2D space. Since the number of intermediate feature maps needed to generate a deconvolutional layer grows exponentially with dimensionality, it is more challenging to solve this issue in higher dimensions. In this work, we propose the voxel deconvolutional layer (VoxelDCL) to solve the checkerboard artifact problem of deconvolutional layers in 3D space. We also provide an efficient approach to implement VoxelDCL. To demonstrate the effectiveness of VoxelDCL, we build four variations of voxel deconvolutional networks (VoxelDCN) based on the U-Net architecture with VoxelDCL. We apply our networks to address volumetric brain images labeling tasks using the ADNI and LONI LPBA40 datasets. The experimental results show that the proposed iVoxelDCNa achieves improved performance in all experiments. It reaches 83.34% in terms of dice ratio on the ADNI dataset and 79.12% on the LONI LPBA40 dataset, which increases 1.39% and 2.21% respectively compared with the baseline. In addition, all the variations of VoxelDCN we proposed outperform the baseline methods on the above datasets, which demonstrates the effectiveness of our methods.

Convolutional neural networks (CNNs) have achieved great success on grid-like data such as images, but face tremendous challenges in learning from more generic data such as graphs. In CNNs, the trainable local filters enable the automatic extraction of high-level features. The computation with filters requires a fixed number of ordered units in the receptive fields. However, the number of neighboring units is neither fixed nor are they ordered in generic graphs, thereby hindering the applications of convolutional operations. Here, we address these challenges by proposing the learnable graph convolutional layer (LGCL). LGCL automatically selects a fixed number of neighboring nodes for each feature based on value ranking in order to transform graph data into grid-like structures in 1-D format, thereby enabling the use of regular convolutional operations on generic graphs. To enable model training on large-scale graphs, we propose a sub-graph training method to reduce the excessive memory and computational resource requirements suffered by prior methods on graph convolutions. Our experimental results on node classification tasks in both transductive and inductive learning settings demonstrate that our methods can achieve consistently better performance on the Cora, Citeseer, Pubmed citation network, and protein-protein interaction network datasets. Our results also indicate that the proposed methods using sub-graph training strategy are more efficient as compared to prior approaches.

Convolutional neural networks (CNNs) have shown great capability of solving various artificial intelligence tasks. However, the increasing model size has raised challenges in employing them in resource-limited applications. In this work, we propose to compress deep models by using channel-wise convolutions, which replace dense connections among feature maps with sparse ones in CNNs. Based on this novel operation, we build light-weight CNNs known as ChannelNets. ChannelNets use three instances of channel-wise convolutions; namely group channel-wise convolutions, depth-wise separable channel-wise convolutions, and the convolutional classification layer. Compared to prior CNNs designed for mobile devices, ChannelNets achieve a significant reduction in terms of the number of parameters and computational cost without loss in accuracy. Notably, our work represents the first attempt to compress the fully-connected classification layer, which usually accounts for about 25% of total parameters in compact CNNs. Experimental results on the ImageNet dataset demonstrate that ChannelNets achieve consistently better performance compared to prior methods.

Convolutional neural networks have shown great success on feature extraction from raw input data such as images. Although convolutional neural networks are invariant to translations on the inputs, they are not invariant to other transformations, including rotation and flip. Recent attempts have been made to incorporate more invariance in image recognition applications, but they are not applicable to dense prediction tasks, such as image segmentation. In this paper, we propose a set of methods based on kernel rotation and flip to enable rotation and flip invariance in convolutional neural networks. The kernel rotation can be achieved on kernels of 3 × 3, while kernel flip can be applied on kernels of any size. By rotating in eight or four angles, the convolutional layers could produce the corresponding number of feature maps based on eight or four different kernels. By using flip, the convolution layer can produce three feature maps. By combining produced feature maps using maxout, the resource requirement could be significantly reduced while still retain the invariance properties. Experimental results demonstrate that the proposed methods can achieve various invariance at reasonable resource requirements in terms of both memory and time.