Simple Convolutional Network Research Articles

State-of-the-Art Results with the Fashion-MNIST Dataset

In September 2024, the Fashion-MNIST dataset will be 7 years old. Proposed as a replacement for the well-known MNIST dataset, it continues to be used to evaluate machine learning model architectures. This paper describes new results achieved with the Fashion-MNIST dataset using classical machine learning models and a relatively simple convolutional network. We present the state-of-the-art results obtained using the CNN-3-128 convolutional network and data augmentation. The developed CNN-3-128 model containing three convolutional layers achieved an accuracy of 99.65% in the Fashion-MNIST test image set. In addition, this paper presents the results of computational experiments demonstrating the dependence between the number of adjustable parameters of the convolutional network and the maximum acceptable classification quality, which allows us to optimise the computational cost of model training.

ЗАСТОСУВАННЯ ГЛИБОКИХ ШТУЧНИХ НЕЙРОННИХ МЕРЕЖ ДЛЯ КЛАСИФІКАЦІЇ МУЛЬТИМОДАЛЬНИХ ДАНИХ

Multimodal data analysis is gaining attention in recent research. Pu Liang et al. (2023) provide a comprehensive overview on multimodal machine learning, highlighting its founda-tions, challenges and achievements in recent years. More problem-oriented works propose new methods and applications for multimodal ML, such a Ngiam et al. (2011) propose to use joint audio and video data to improve speech recognition accuracy; Sun, Wand and Li (2018) describe application of multimodal classification for breast cancer prognosis prediction; Mao et al. (2014) propose an architecture of multimodal recurrent network to generate text de-scription of images and so on. However, such works usually focus on the task itself and meth-ods therein, and not on integrating multimodal data processing into other software systems. The goal of this research is to propose a way to conduct multimodal data processing, specifically as a part of a digital twin systems, thus efficiency and near-real-time operation are required. The paper presents an approach to conduct parallel multimodal data classification, adapting to available computing power. The method is modular and scalable and intended for in digital twin application as a part of analysis and modeling tools. Later, the detailed example of such a software module is discussed. It uses multimodal data from open datasets to detect and classify the behavior of pets using deep learning mod-els. Videos are processed using two artificial neural networks: YOLOv3 object detection net-work to process individual frames of the video and a relatively simple convolutional network to classify sounds based on their frequency spectra. Constructed module uses a producer-consumer parallel processing pattern and allows processing 5 frames per second of a video on available hardware, which can be sufficiently improved by using GPU acceleration or more paralleled processing threads.

Revitalizing Convolutional Network for Image Restoration.

Image restoration aims to reconstruct a high-quality image from its corrupted version, playing essential roles in many scenarios. Recent years have witnessed a paradigm shift in image restoration from convolutional neural networks (CNNs) to Transformerbased models due to their powerful ability to model long-range pixel interactions. In this paper, we explore the potential of CNNs for image restoration and show that the proposed simple convolutional network architecture, termed ConvIR, can perform on par with or better than the Transformer counterparts. By re-examing the characteristics of advanced image restoration algorithms, we discover several key factors leading to the performance improvement of restoration models. This motivates us to develop a novel network for image restoration based on cheap convolution operators. Comprehensive experiments demonstrate that our ConvIR delivers state-ofthe- art performance with low computation complexity among 20 benchmark datasets on five representative image restoration tasks, including image dehazing, image motion/defocus deblurring, image deraining, and image desnowing.

Efficient ship noise classification with positive incentive noise and fused features using a simple convolutional network

Ship noise analysis is a critical area of research in hydroacoustic remote sensing due to its practical implications in identifying ship direction, type, and even specific ship identities. However, the limited availability of data poses challenges in developing accurate ship noise classification models. Previous studies have mainly focused on small-sample learning approaches, resulting in complex network structures. Nonetheless, underwater robots often have limited computing power, making it essential to develop simpler recognition networks. In this paper, we address the issue of data scarcity by introducing positive incentive noise. We propose a CNN-based hydroacoustic signal recognition method that achieves comparable or superior performance to previous studies, using a simple network structure as a back-end decision system. We describe the feature extraction process using a dataset with added noise and compare the performance of various features. Additionally, we compare our proposed method with previous studies. Experimental results demonstrate that simple neural networks can achieve high performance and excellent generalizability without the need for complex network structures like adversarial learning models.

A Recognition Method for Soft Objects Based on the Fusion of Vision and Haptics.

For humans and animals to recognise an object, the integration of multiple sensing methods is essential when one sensing modality is only able to acquire limited information. Among the many sensing modalities, vision has been intensively studied and proven to have superior performance for many problems. Nevertheless, there are many problems which are difficult to solve by solitary vision, such as in a dark environment or for objects with a similar outlook but different inclusions. Haptic sensing is another commonly used means of perception, which can provide local contact information and physical features that are difficult to obtain by vision. Therefore, the fusion of vision and touch is beneficial to improve the robustness of object perception. To address this, an end-to-end visual-haptic fusion perceptual method has been proposed. In particular, the YOLO deep network is used to extract vision features, while haptic explorations are used to extract haptic features. Then, visual and haptic features are aggregated using a graph convolutional network, and the object is recognised based on a multi-layer perceptron. Experimental results show that the proposed method excels in distinguishing soft objects that have similar appearance but varied interior fillers, comparing a simple convolutional network and a Bayesian filter. The resultant average recognition accuracy was improved to 0.95 from vision only (mAP is 0.502). Moreover, the extracted physical features could be further used for manipulation tasks targeting soft objects.

Multi‐scale homography estimation based on dual feature aggregation transformer

AbstractThe accuracy of registration in image stitching task directly affects the performance of subsequent stages. Traditional registration methods rely heavily on the quality of the features when calculating the homography matrix, resulting in alignment failures in low‐texture or low‐overlap scenes due to extracting insufficient features. On the other hand, existing DNN‐based methods for homography estimation are more robust in multiple scenes but previous work usually employs an overly simple convolutional network structure to directly regress the homography, ignoring the redundant information contained in the feature maps so that their prediction accuracy is inferior to the traditional methods in simple scenarios. To overcome the disadvantages of the two methods, a Multi‐scale structure is proposed to extract feature maps at three scales and design two modules to handle the matrix prediction respectively. The DFA‐T module analyzes semantic information on the high‐level features to accomplish coarse‐grained alignment while the Contextual Correlation module on the bottom level to accomplish more accurate alignment. Experiments demonstrate that this method provides more accurate alignment results than the existing state‐of‐the‐art DNN‐based methods and outperforms traditional algorithms with more stable results in some extreme scenarios.

Privacy-preserved learning from non-i.i.d data in fog-assisted IoT: A federated learning approach

With the prevalence of the Internet of Things (IoT) systems, smart cities comprise complex networks, including sensors, actuators, appliances, and cyber services. The complexity and heterogeneity of smart cities have become vulnerable to sophisticated cyber-attacks, especially privacy-related attacks such as inference and data poisoning ones. Federated Learning (FL) has been regarded as a hopeful method to enable distributed learning with privacy-preserved intelligence in IoT applications. Even though the significance of developing privacy-preserving FL has drawn as a great research interest, the current research only concentrates on FL with independent identically distributed (i.i.d) data and few studies have addressed the non-i. i.d setting. FL is known to be vulnerable to Generative Adversarial Network (GAN) attacks, where an adversary can presume to act as a contributor participating in the training process to acquire the private data of other contributors. This paper proposes an innovative Privacy Protection-based Federated Deep Learning (PP-FDL) framework, which accomplishes data protection against privacy-related GAN attacks, along with high classification rates from non-i. i.d data. PP-FDL is designed to enable fog nodes to cooperate to train the FDL model in a way that ensures contributors have no access to the data of each other, where class probabilities are protected utilizing a private identifier generated for each class. The PP-FDL framework is evaluated for image classification using simple convolutional networks which are trained using MNIST and CIFAR-10 datasets. The empirical results have revealed that PF-DFL can achieve data protection and the framework outperforms the other three state-of-the-art models with 3%–8% as accuracy improvements.

Real world Gender and Age Prediction using CNN

The real world gender and age prediction is opined to develop to deal multitude of real world problems and scenarios where in there is an exceedingly great usage for the validation of a person and to authenticate the identity of a person. This project would be developed using a renowned deep learning algorithm – CNN ( Convolutional Neural Network ) to enhance the precision rate. To this end, we design a simple convolutional network architecture that can be used and deployed using adequate number of pooling layers so as to get sterling precision to the extent possible, training the model using publicly available dataset. Keywords deep learning, convolutional neural network, computer vision, prediction.

Automatic detection of subviral particles in fluorescence microscopy images of Marburg virus infected cells using a convolutional neural network

Abstract In collaboration with the Institute of Virology at the University of Marburg, a tracking algorithm has been developed in recent years. Thereby, for the tracking of subviral particles and their intracellular transport through the cell, the detection of these particles is the first major step. In this work a deep-learning based approach for automatic processing of fluorescence video sequences of virus infected cells is presented. The focus of this work was to generate synthetic data for training neuronal networks and to represent a particle using only a single delta peak. It was shown that a simple convolutional network is able to predict the particle positions with an almost identical accuracy as the algorithm developed by Kienzle, Rausch and colleagues [1-3], and that even a small dataset is sufficient for that purpose.

Fast and accurate dose predictions for novel radiotherapy treatments in heterogeneous phantoms using conditional 3D‐UNet generative adversarial networks

Novel radiotherapy techniques like synchrotron X-ray microbeam radiation therapy (MRT) require fast dose distribution predictions that are accurate at the sub-mm level, especially close to tissue/bone/air interfaces. Monte Carlo (MC) physics simulations are recognized to be one of the most accurate tools to predict the dose delivered in a target tissue but can be very time consuming and therefore prohibitive for treatment planning. Faster dose prediction algorithms are usually developed for clinically deployed treatments only. In this work, we explore a new approach for fast and accurate dose estimations suitable for novel treatments using digital phantoms used in preclinical development and modern machine learning techniques. We develop a generative adversarial network (GAN) model, which is able to emulate the equivalent Geant4 MC simulation with adequate accuracy and use it to predict the radiation dose delivered by a broad synchrotron beam to various phantoms. The energy depositions used for the training of the GAN are obtained using full Geant4 MC simulations of a synchrotron radiation broad beam passing through the phantoms. The energy deposition is scored and predicted in voxel matrices of size 140 × 18 × 18 with a voxel edge length of 1 mm. The GAN model consists of two competing 3D convolutional neural networks, which are conditioned on the photon beam and phantom properties. The generator network has a U-Net structure and is designed to predict the energy depositions of the photon beam inside three phantoms of variable geometry with increasing complexity. The critic network is a relatively simple convolutional network, which is trained to distinguish energy depositions predicted by the generator from the ones obtained with the full MC simulation. The energy deposition predictions inside all phantom geometries under investigation show deviations of less than 3% of the maximum deposited energy from the simulation for roughly 99% of the voxels in the field of the beam. Inside the most realistic phantom, a simple pediatric head, the model predictions deviate by less than 1% of the maximal energy deposition from the simulations in more than 96% of the in-field voxels. For all three phantoms, the model generalizes the energy deposition predictions well to phantom geometries, which have not been used for training the model but are interpolations of the training data in multiple dimensions. The computing time for a single prediction is reduced from several hundred hours using Geant4 simulation to less than a second using the GAN model. The proposed GAN model predicts dose distributions inside unknown phantoms with only small deviations from the full MC simulation with computations times of less than a second. It demonstrates good interpolation ability to unseen but similar phantom geometries and is flexible enough to be trained on data with different radiation scenarios without the need for optimization of the model parameter. This proof-of-concept encourages to apply and further develop the model for the use in MRT treatment planning, which requires fast and accurate predictions with sub-mm resolutions.

Polynomial Approximations for Several Neural Network Activation Functions

Active deployment of machine learning systems sets a task of their protection against different types of attacks that threaten confidentiality, integrity and accessibility of both processed data and trained models. One of the promising ways for such protection is the development of privacy-preserving machine learning systems, that use homomorphic encryption schemes to protect data and models. However, such schemes can only process polynomial functions, which means that we need to construct polynomial approximations for nonlinear functions used in neural models. The goal of this paper is the construction of precise approximations of several widely used neural network activation functions while limiting the degree of approximation polynomials as well as the evaluation of the impact of the approximation precision on the resulting value of the whole neural network. In contrast to the previous publications, in the current paper we study and compare different ways for polynomial approximation construction, introduce precision metrics, present exact formulas for approximation polynomials as well as exact values of corresponding precisions. We compare our results with the previously published ones. Finally, for a simple convolutional network we experimentally evaluate the impact of the approximation precision on the bias of the output neuron values of the network from the original ones. Our results show that the best approximation for ReLU could be obtained with the numeric method, and for the sigmoid and hyperbolic tangent – with Chebyshev polynomials. At the same time, the best approximation among the three functions could be obtained for ReLU. The results could be used for the construction of polynomial approximations of activation functions in privacy-preserving machine learning systems.

A deep descriptor for cross-tasking EEG-based recognition.

Due to the application of vital signs in expert systems, new approaches have emerged, and vital signals have been gaining space in biometrics. One of these signals is the electroencephalogram (EEG). The motor task in which a subject is doing, or even thinking, influences the pattern of brain waves and disturb the signal acquired. In this work, biometrics with the EEG signal from a cross-task perspective are explored. Based on deep convolutional networks (CNN) and Squeeze-and-Excitation Blocks, a novel method is developed to produce a deep EEG signal descriptor to assess the impact of the motor task in EEG signal on biometric verification. The Physionet EEG Motor Movement/Imagery Dataset is used here for method evaluation, which has 64 EEG channels from 109 subjects performing different tasks. Since the volume of data provided by the dataset is not large enough to effectively train a Deep CNN model, it is also proposed a data augmentation technique to achieve better performance. An evaluation protocol is proposed to assess the robustness regarding the number of EEG channels and also to enforce train and test sets without individual overlapping. A new state-of-the-art result is achieved for the cross-task scenario (EER of 0.1%) and the Squeeze-and-Excitation based networks overcome the simple CNN architecture in three out of four cross-individual scenarios.

Is Active Learning Always Beneficial? (Student Abstract)

This study highlights the limitations of automated curriculum learning, which may not be a viable strategy for tasks in which the benefits of the chosen curriculum are not apparent until much later. Using a simple convolutional network and a two-task training regime, we show that in some cases a network is not able to derive an optimal learning strategy using only the data available during a single training run.

Single-Stage Refinement CNN for Depth Estimation in Monocular Images

Depth reconstruction from single monocular images has been a challenging task due to the complexity and the quantity of depth cues that images have. Convolutional Neural Networks (CNN) have been successfully used to reconstruct depth of general object scenes; however, proposed works use several stages of training which make this process more complex and time consuming. As we aim to build a computational efficient model, we focus on single-stage training CNN. In this paper, we propose five different models for solving this task, ranging from a simple convolutional network, to one with residual, convolutional, refinement and upsampling layers. We compare our models with the current state of the art in depth reconstruction and measure depth reconstruction error for different datasets (KITTI, NYU), obtaining improvements in both global and local error measures.

Time-Series Forecasting of Mortality Rates using Deep Learning

The time-series nature of mortality rates lends itself to processing through neural networks that are specialized to deal with sequential data, such as recurrent and convolutional networks. Although appealing intuitively, a naive implementation of these networks does not lead to enhanced predictive performance. We show how the structure of the Lee Carter model can be generalized, and propose a relatively simple convolutional network model that can be interpreted as a generalization of the Lee Carter model, allowing for its components to be evaluated in familiar terms. The model produces highly accurate forecasts on the Human Mortality Database, and, without further modification, generalizes well to the United States Mortality Database.

T–S Fuzzy Model Based Multi-Branch Deep Network Architecture

In the traditional CNN design, the hyperparameters, such as the size of the convolutional kernel and stride, are difficult to determine. In this paper, a new convolutional network architecture, named multi-branch fuzzy architecture network (MBFAN), was proposed for this problem. In MBFAN, some branches with a certain convolutional neural network architecture are connected in parallel. In each branch, a different-sized convolutional kernel is applied. By data training and normalization, a weight is given to each branch. By these weights, the important features in the final output are strengthened. By normalization, the branches were interconnected together, making the training process more efficient. Due to overfitting, with the increase of branches, the MBFAN accuracy increases, and then decreases. The number of branches is optimized when the MBFAN accuracy is highest. On the other hand, the location of the convolutional kernel center in an image has a great influence on the convolutional results. This is also discussed in MBFAN. For the experiments, the proposed MBFAN was adopted and tested in a simple convolutional network and a VGG16 network.

Parallel Image Captioning Using 2D Masked Convolution

Automatically generating a novel description of an image is a challenging and important problem that brings together advanced research in both computer vision and natural language processing. In recent years, image captioning has significantly improved its performance by using long short-term memory (LSTM) as a decoder for the language model. However, despite this improvement, LSTM itself has its own shortcomings as a model because the structure is complicated and its nature is inherently sequential. This paper proposes a model using a simple convolutional network for both encoder and decoder functions of image captioning, instead of the current state-of-the-art approach. Our experiment with this model on a Microsoft Common Objects in Context (MSCOCO) captioning dataset yielded results that are competitive with the state-of-the-art image captioning model across different evaluation metrics, while having a much simpler model and enabling parallel graphics processing unit (GPU) computation during training, resulting in a faster training time.