In recent years, the increasing complexity of deep network structures has hindered their application in small resource constrained hardware. Therefore, we urgently need to compress and accelerate deep network models. Channel pruning is an effective method to compress deep neural networks. However, most existing channel pruning methods are prone to falling into local optima. In this paper, we propose a channel pruning method via Improved Grey Wolf Optimizer Pruner which called IGWO-Pruner to prune redundant channels of convolutional neural networks. It identifies pruning ratio of each layer by using Improved Grey Wolf algorithm, and then fine-tuning the new pruned network model. In experimental section, we evaluate the proposed method in CIFAR datasets and ILSVRC-2012 with several classical networks, including VGGNet, GoogLeNet and ResNet-18/34/56/152, and experimental results demonstrate the proposed method is able to prune a large number of redundant channels and parameters with rare performance loss.

Convolutional neural network (CNN) is widely used in acoustic scene classification (ASC) tasks. In most cases, local convolution is utilized to gather time-frequency information between spectrum nodes. It is challenging to adequately express the non-local link between frequency domains in a finite convolution region. In this paper, we propose a dual-path convolutional neural network based on band interaction block (DCNN-bi) for ASC, with mel-spectrogram as the model’s input. We build two parallel CNN paths to learn the high-frequency and low-frequency components of the input feature. Additionally, we have created three band interaction blocks (bi-blocks) to explore the pertinent nodes between various frequency bands, which are connected between two paths. Combining the time-frequency information from two paths, the bi-blocks with three distinct designs acquire non-local information and send it back to the respective paths. The experimental results indicate that the utilization of the bi-block has the potential to improve the initial performance of the CNN substantially. Specifically, when applied to the DCASE 2018 and DCASE 2020 datasets, the CNN exhibited performance improvements of 1.79% and 3.06%, respectively.

Federated learning is a distributed machine learning approach in which clients train models locally with their own data and upload them to a server so that their trained results are shared between them without uploading raw data to the server. There are some challenges in federated learning, such as communication size reduction and client heterogeneity. The former can mitigate the communication overheads, and the latter can allow the clients to choose proper models depending on their available compute resources. To address these challenges, in this paper, we utilize Neural ODE based models for federated learning. The proposed flexible federated learning approach can reduce the communication size while aggregating models with different iteration counts or depths. Our contribution is that we experimentally demonstrate that the proposed federated learning can aggregate models with different iteration counts or depths. It is compared with a different federated learning approach in terms of the accuracy. Furthermore, we show that our approach can reduce communication size by up to 89.4% compared with a baseline ResNet model using CIFAR-10 dataset.

Wide deployment of artificial intelligence (AI) is inducing exponentially growing energy consumption. Traditional digital platforms are becoming difficult to fulfill such ever-growing demands on energy efficiency as well as computing latency, which necessitates the development of high efficiency analog hardware platforms for AI. Recently, optical and electrooptic hybrid computing is reactivated as a promising analog hardware alternative because it can accelerate the information processing in an energy-efficient way. Integrated photonic circuits offer such an analog hardware solution for implementing photonic AI and machine learning. For this purpose, we proposed a photonic analog of support vector machine and experimentally demonstrated low-latency and low-energy classification computing, which evidences the latency and energy advantages of optical analog computing over traditional digital computing. We also proposed an electrooptic Hopfield network for classifying and recognizing time-series data. This paper will review our work on implementing classification computing and Hopfield network by leveraging silicon photonic circuits.

Despite recent advancements in utilizing meta-learning for addressing the generalization challenges of graph neural networks (GNN), their performance in argumentation mining tasks, such as argument classifications, remains relatively limited. This is primarily due to the under-utilization of potential pattern knowledge intrinsic to argumentation structures. To address this issue, our study proposes a two-stage, pattern-based meta-GNN method in contrast to conventional pattern-free meta-GNN approaches. Initially, our method focuses on learning a high-level pattern representation to effectively capture the pattern knowledge within an argumentation structure and then predicts edge types. It then utilizes a meta-learning framework in the second stage, designed to train a meta-learner based on the predicted edge types. This feature allows for rapid generalization to novel argumentation graphs. Through experiments on real English discussion datasets spanning diverse topics, our results demonstrate that our proposed method substantially outperforms conventional pattern-free GNN approaches, signifying a significant stride forward in this domain.

In the field of machine learning security, as one of the attack surfaces especially for edge devices, the application of side-channel analysis such as correlation power/electromagnetic analysis (CPA/CEMA) is expanding. Aiming to evaluate the leakage resistance of neural network (NN) model parameters, i.e. weights and biases, we conducted a feasibility study of CPA/CEMA on floating-point (FP) operations, which are the basic operations of NNs. This paper proposes approaches to recover weights and biases using CPA/CEMA on multiplication and addition operations, respectively. It is essential to take into account the characteristics of the IEEE 754 representation in order to realize the recovery with high precision and efficiency. We show that CPA/CEMA on FP operations requires different approaches than traditional CPA/CEMA on cryptographic implementations such as the AES.

Emotional speech recognition is generally considered more difficult than non-emotional speech recognition. The acoustic characteristics of emotional speech differ from those of non-emotional speech. Additionally, acoustic characteristics vary significantly depending on the type and intensity of emotions. Regarding linguistic features, emotional and colloquial expressions are also observed in their utterances. To solve these problems, we aim to improve recognition performance by adapting acoustic and language models to emotional speech. We used Japanese Twitter-based Emotional Speech (JTES) as an emotional speech corpus. This corpus consisted of tweets and had an emotional label assigned to each utterance. Corpus adaptation is possible using the utterances contained in this corpus. However, regarding the language model, the amount of adaptation data is insufficient. To solve this problem, we propose an adaptation of the language model by using online tweet data downloaded from the internet. The sentences used for adaptation were extracted from the tweet data based on certain rules. We extracted the data of 25.86 M words and used them for adaptation. In the recognition experiments, the baseline word error rate was 36.11%, whereas that with the acoustic and language model adaptation was 17.77%. The results demonstrated the effectiveness of the proposed method.

This paper investigates the effects of backdoor attacks on graph neural networks (GNNs) trained through simple data augmentation by modifying the edges of the graph in graph classification. The numerical results show that GNNs trained with data augmentation remain vulnerable to backdoor attacks and may even be more vulnerable to such attacks than GNNs without data augmentation.

While deep image compression performs better than traditional codecs like JPEG on natural images, it faces a challenge as a learning-based approach: compression performance drastically decreases for out-of-domain images. To investigate this problem, we introduce a novel task that we call universal deep image compression, which involves compressing images in arbitrary domains, such as natural images, line drawings, and comics. Furthermore, we propose a content-adaptive optimization framework to tackle this task. This framework adapts a pre-trained compression model to each target image during testing for addressing the domain gap between pre-training and testing. For each input image, we insert adapters into the decoder of the model and optimize the latent representation extracted by the encoder and the adapter parameters in terms of rate-distortion, with the adapter parameters transmitted per image. To achieve the evaluation of the proposed universal deep compression, we constructed a benchmark dataset containing uncompressed images of four domains: natural images, line drawings, comics, and vector arts. We compare our proposed method with non-adaptive and existing adaptive compression methods, and the results show that our method outperforms them. Our code and dataset are publicly available at https://github.com/kktsubota/universal-dic.

We investigated the influence of horizontal shifts of the input images for one stage object detection method. We found that the object detector class scores drop when the target object center is at the grid boundary. Many approaches have focused on reducing the aliasing effect of down-sampling to achieve shift-invariance. However, down-sampling does not completely solve this problem at the grid boundary; it is necessary to suppress the dispersion of features in pixels close to the grid boundary into adjacent grid cells. Therefore, this paper proposes two approaches focused on the grid boundary to improve this weak point of current object detection methods. One is the Sub-Grid Feature Extraction Module, in which the sub-grid features are added to the input of the classification head. The other is Grid-Aware Data Augmentation, where augmented data are generated by the grid-level shifts and are used in training. The effectiveness of the proposed approaches is demonstrated using the COCO validation set after applying the proposed method to the FCOS architecture.

Convolutional Neural Networks (CNNs) and Transformers have achieved remarkable performance in detection and classification tasks. Nevertheless, their feature extraction cannot consider both local and global information, so the detection and classification performance can be further improved. In addition, more and more deep learning networks are designed as more and more complex, and the amount of computation and storage space required is also significantly increased. This paper proposes a combination of CNN and transformer, and designs a local feature enhancement module and global context modeling module to enhance the cascade network. While the local feature enhancement module increases the range of feature extraction, the global context modeling is used to capture the feature maps' global information. To decrease the model complexity, a shared sublayer is designed to realize the sharing of weight parameters between the adjacent convolutional layers or cross convolutional layers, thereby reducing the number of convolutional weight parameters. Moreover, to effectively improve the detection performance of neural networks without increasing network parameters, the optimal transport assignment approach is proposed to resolve the problem of label assignment. The classification loss and regression loss are the summations of the cost between the demander and supplier. The experiment results demonstrate that the proposed Combination of CNN and Transformer with Shared Sublayer (CCTSS) performs better than the state-of-the-art methods in various datasets and applications.

For massive multiple-input multiple-output (MIMO) communication systems, simple linear detectors such as zero forcing (ZF) and minimum mean square error (MMSE) can achieve near-optimal detection performance with reduced computational complexity. However, such linear detectors always involve complicated matrix inversion, which will suffer from high computational overhead in the practical implementation. Due to the massive parallel-processing and efficient hardware-implementation nature, the neural network has become a promising approach to signal processing for the future wireless communications. In this paper, we first propose an efficient neural network to calculate the pseudo-inverses for any type of matrices based on the improved Newton's method, termed as the PINN. Through detailed analysis and derivation, the linear massive MIMO detectors are mapped on PINNs, which can take full advantage of the research achievements of neural networks in both algorithms and hardwares. Furthermore, an improved limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) quasi-Newton method is studied as the learning algorithm of PINNs to achieve a better performance/complexity trade-off. Simulation results finally validate the efficiency of the proposed scheme.

In this paper, we describe a wavelength-division multiplexing visible-light communication (VLC) system using two colored light-emitting diodes (LEDs) with similar emission wavelengths. A multi-input multi-output signal-separation method using a neural network is proposed to cancel the optical cross chatter caused by the spectral overlap of LEDs. The experimental results demonstrate that signal separation using neural networks can be achieved in wavelength-multiplexed VLC systems with a bit error rate of less than 3.8×10-3 (forward error correction limit). Furthermore, the simulation results reveal that the carrier-to-noise ratio (CNR) is improved by 2dB for the successive interference canceller (SIC) compared to the zero-forcing method.

Existing simple routing protocols (e.g., OSPF, RIP) have some disadvantages of being inflexible and prone to congestion due to the concentration of packets on particular routers. To address these issues, packet routing methods using machine learning have been proposed recently. Compared to these algorithms, machine learning based methods can choose a routing path intelligently by learning efficient routes. However, machine learning based methods have a disadvantage of training time overhead. We thus focus on a lightweight machine learning algorithm, OS-ELM (Online Sequential Extreme Learning Machine), to reduce the training time. Although previous work on reinforcement learning using OS-ELM exists, it has a problem of low learning accuracy. In this paper, we propose OS-ELM QN (Q-Network) with a prioritized experience replay buffer to improve the learning performance. It is compared to a deep reinforcement learning based packet routing method using a network simulator. Experimental results show that introducing the experience replay buffer improves the learning performance. OS-ELM QN achieves a 2.33 times speedup than a DQN (Deep Q-Network) in terms of learning speed. Regarding the packet transfer latency, OS-ELM QN is comparable or slightly inferior to the DQN while they are better than OSPF in most cases since they can distribute congestions.

With a rapidly escalating number of sophisticated cyber-attacks, protecting Internet of Things (IoT) networks against unauthorized activity is a major concern. The detection of malicious attack traffic is thus crucial for IoT security to prevent unwanted traffic. However, existing traditional malicious traffic detection systems which relied on supervised machine learning approach need a considerable number of benign and malware traffic samples to train the machine learning models. Moreover, in the cases of zero-day attacks, only a few labeled traffic samples are accessible for analysis. To deal with this, we propose a few-shot malicious IoT traffic detection system with a prototypical graph neural network. The proposed approach does not require prior knowledge of network payload binaries or network traffic signatures. The model is trained on labeled traffic data and tested to evaluate its ability to detect new types of attacks when only a few labeled traffic samples are available. The proposed detection system first categorizes the network traffic as a bidirectional flow and visualizes the binary traffic flow as a color image. A neural network is then applied to the visualized traffic to extract important features. After that, using the proposed few-shot graph neural network approach, the model is trained on different few-shot tasks to generalize it to new unseen attacks. The proposed model is evaluated on a network traffic dataset consisting of benign traffic and traffic corresponding to six types of attacks. The results revealed that our proposed model achieved an F1 score of 0.91 and 0.94 in 5-shot and 10-shot classification, respectively, and outperformed the baseline models.

In machine learning, data augmentation (DA) is a technique for improving the generalization performance of models. In this paper, we mainly consider gradient descent of linear regression under DA using noisy copies of datasets, in which noise is injected into inputs. We analyze the situation where noisy copies are newly generated and injected into inputs at each epoch, i.e., the case of using on-line noisy copies. Therefore, this article can also be viewed as an analysis on a method using noise injection into a training process by DA. We considered the training process under three training situations which are the full-batch training under the sum of squared errors, and full-batch and mini-batch training under the mean squared error. We showed that, in all cases, training for DA with on-line copies is approximately equivalent to the l2 regularization training for which variance of injected noise is important, whereas the number of copies is not. Moreover, we showed that DA with on-line copies apparently leads to an increase of learning rate in full-batch condition under the sum of squared errors and the mini-batch condition under the mean squared error. The apparent increase in learning rate and regularization effect can be attributed to the original input and additive noise in noisy copies, respectively. These results are confirmed in a numerical experiment in which we found that our result can be applied to usual off-line DA in an under-parameterization scenario and can not in an over-parametrization scenario. Moreover, we experimentally investigated the training process of neural networks under DA with off-line noisy copies and found that our analysis on linear regression can be qualitatively applied to neural networks.

Convolutional Neural Networks (CNNs) have shown remarkable performance in image recognition tasks. In this letter, we propose a new CNN model called the EnsNet which is composed of one base CNN and multiple Fully Connected SubNetworks (FCSNs). In this model, the set of feature maps generated by the last convolutional layer in the base CNN is divided along channels into disjoint subsets, and these subsets are assigned to the FCSNs. Each of the FCSNs is trained independent of others so that it can predict the class label of each feature map in the subset assigned to it. The output of the overall model is determined by majority vote of the base CNN and the FCSNs. Experimental results using the MNIST, Fashion-MNIST and CIFAR-10 datasets show that the proposed approach further improves the performance of CNNs. In particular, an EnsNet achieves a state-of-the-art error rate of 0.16% on MNIST.

The original single-shot multibox detector (SSD) algorithm has good detection accuracy and speed for regular object recognition. However, the SSD is not suitable for detecting small objects for two reasons: 1) the relationships among different feature layers with various scales are not considered, 2) the predicted results are solely determined by several independent feature layers. To enhance its detection capability for small objects, this study proposes an improved SSD-based algorithm called proportional channels' fusion SSD (PCF-SSD). Three enhancements are provided by this novel PCF-SSD algorithm. First, a fusion feature pyramid model is proposed by concatenating channels of certain key feature layers in a given proportion for object detection. Second, the default box sizes are adjusted properly for small object detection. Third, an improved loss function is suggested to train the above-proposed fusion model, which can further improve object detection performance. A series of experiments are conducted on the public database Pascal VOC to validate the PCF-SSD. On comparing with the original SSD algorithm, our algorithm improves the mean average precision and detection accuracy for small objects by 3.3% and 3.9%, respectively, with a detection speed of 40FPS. Furthermore, the proposed PCF-SSD can achieve a better balance of detection accuracy and efficiency than the original SSD algorithm, as demonstrated by a series of experimental results.

Epileptic seizure prediction is an important research topic in the clinical epilepsy treatment, which can provide opportunities to take precautionary measures for epilepsy patients and medical staff. EEG is an commonly used tool for studying brain activity, which records the electrical discharge of brain. Many studies based on machine learning algorithms have been proposed to solve the task using EEG signal. In this study, we propose a novel seizure prediction models based on convolutional neural networks and scalp EEG for a binary classification between preictal and interictal states. The short-time Fourier transform has been used to translate raw EEG signals into STFT sepctrums, which is applied as input of the models. The fusion features have been obtained through the side-output constructions and used to train and test our models. The test results show that our models can achieve comparable results in both sensitivity and FPR upon fusion features. The proposed patient-specific model can be used in seizure prediction system for EEG classification.

Convolutional neural networks (CNNs) have become popular in medical image segmentation. The widely used deep CNNs are customized to extract multiple representative features for two-dimensional (2D) data, generally called 2D networks. However, 2D networks are inefficient in extracting three-dimensional (3D) spatial features from volumetric images. Although most 2D segmentation networks can be extended to 3D networks, the naively extended 3D methods are resource-intensive. In this paper, we propose an efficient and accurate network for fully automatic 3D segmentation. Specifically, we designed a 3D multiple-contextual extractor to capture rich global contextual dependencies from different feature levels. Then we leveraged an ROI-estimation strategy to crop the ROI bounding box. Meanwhile, we used a 3D ROI-attention module to improve the accuracy of in-region segmentation in the decoder path. Moreover, we used a hybrid Dice loss function to address the issues of class imbalance and blurry contour in medical images. By incorporating the above strategies, we realized a practical end-to-end 3D medical image segmentation with high efficiency and accuracy. To validate the 3D segmentation performance of our proposed method, we conducted extensive experiments on two datasets and demonstrated favorable results over the state-of-the-art methods.