Changqing Zhang, Yu Geng, Zongbo Han et al. · IEEE Transactions on Neural Networks and Learning Systems · 2022
Modeling complex correlations on multiview data is still challenging, especially for high-dimensional features with possible noise. To address this issue, we propose a novel unsupervised multiview representation learning (UMRL) algorithm, termed autoencoder in autoencoder networks (AE2-Nets). The proposed framework effectively encodes information from high-dimensional heterogeneous data into a compact and informative representation with the proposed bidirectional encoding strategy. Specifically, the proposed AE2-Nets conduct encoding in two directions: the inner-AE-networks extract view-specific intrinsic information (forward encoding), while the outer-AE-networks integrate this view-specific intrinsic information from different views into a latent representation (backward encoding). For the nested architecture, we further provide a probabilistic explanation and extension from hierarchical variational autoencoder. The forward-backward strategy flexibly addresses high-dimensional (noisy) features within each view and encodes complementarity across multiple views in a unified framework. Extensive results on benchmark datasets validate the advantages compared to the state-of-the-art algorithms.
Cited by 44
Akihiro Suzuki (682968), Takashi Morie (4966699), Hakaru Tamukoh (4966702) · Figshare · 2018
<p>Reprinted from [<a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0194049#pone.0194049.ref041" target="_blank">41</a>] under a CC BY license, with permission from Springer Nature, original copyright 2016 (<a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0194049#pone.0194049.s001" target="_blank">S1 File</a>).</p>
Cited by 0Open access
Takeshi Okadome · 2025
Cited by 0
Shengwei Xu (6990542), Jijie Han (20439599), Jianbo Wang (184708) et al. · Figshare · 2025
<div> To solve the problems of existing encrypted traffic classification methods, such as the need for large-scale training data, high computational costs, and poor generalization ability, an encrypted traffic classification method based on autoencoders and convolutional neural networks was proposed. This method first utilizes an autoencoder to recon-struct the dataset, enabling it to work with smaller-scale datasets. The autoencoder allows shorter traffic flows to learn abstract feature representations from longer traffic flows of the same type, replacing zeros and mitigating the negative effects of zero-padding on traffic classification when using uniform flow lengths. After reconstruction, a convolutional neural network is used to classify the traffic. Due to its characteristics of parameter sharing and local connectivity, the CNN exhibits strong generalization ability when handling tasks, allowing it to better adapt to samples outside of the training data. Experimental results show that, compared to existing advanced methods, this method can achieve a classi-fication accuracy improvement of 2.86% to 18.13%, while also demonstrating greater robustness compared to other advanced methods. The code is available at <a href="https://github.com/han20011019/AECCN" target="_blank">https://github.com/han20011019/AECCN</a>. </div>
Cited by 0Open access
Victor Lempitsky · 2021
Cited by 0
Mostafa Eltager (17102176), Tamim Abdelaal (7111226), Mohammed Charrout (17102179) et al. · Figshare · 2023
<p>Both models are based on the encoder-decoder neural network structure to a learn latent space. A) An autoencoder is a deterministic model where <i>z</i> is a mapping of the input data. B) A variational autoencoder is a probabilistic model where the mapping <i>z</i> is generated by a probability distribution conditioned on the input data.</p>
Cited by 0Open access
Liu, Rongrong · Zenodo (CERN European Organization for Nuclear Research) · 2025
Galaxy Denoising Autoencoder (DAE) This repository accompanies Liu et al. 2025, ApJ. It contains a TensorFlow-based Denoising Autoencoder (DAE) to model galaxies from observed galaxy images and perform galaxy model subtraction. The code, configuration files, and dataset generation scripts allow full reproducibility of the experiments described in the paper. Prerequisites Before running the scripts, ensure you have the following installed: Python 3.7 or later TensorFlow 2.x NumPy Astropy GALFIT (for dataset generation) Python dependencies can be installed with: pip install -r requirements.txt File Structure dae_model.py: Defines the autoencoder model (DAE class). fits_data_generator.py: Loads and preprocesses FITS data in batches for training. train.py: Runs the training process using settings from train_config.json. test.py: Evaluates the trained model on test data, outputting denoised and residual images. gen_gal_input.py: Generates input parameter files for GALFIT. run_galfit.py: Uses GALFIT to create simulated galaxy images from parameter files. add_background.py: Adds realistic sky backgrounds to simulated galaxy images. train_config.json: Configuration file for training (paths, parameters, model settings). test_config.json: Configuration file for testing (paths, output directories, pretrained model). Dataset Generation The dataset for training and testing is created by generating simulated galaxy images and combining them with realistic sky backgrounds. Step 1: Generate Simulated Galaxies Use gen_gal_input.py to create GALFIT parameter files and run_galfit.py to generate galaxy images. python gen_gal_input.py python run_galfit.py Step 2: Add Sky Backgrounds Use add_background.py to add clean sky backgrounds to simulated galaxies. These pairs of noisy (galaxy + sky) and clean (galaxy only) images form the training dataset. python add_background.py Output: A dataset of paired images (noisy vs. clean) ready for use in training/testing the DAE. Training Hardware Requirements Training on Google Colab with an L4 GPU takes ~3 hours with the default configuration. Larger models will require more resources. Training Configuration Settings are specified in train_config.json, including: Input/output directories Model architecture parameters Optimizer settings Training options (epochs, batch size, validation split, early stopping, checkpointing, etc.) Example run: python train.py --config path/to/train_config.json Testing Test Configuration Settings are stored in test_config.json, including: Path to trained model (.h5 file) Input test dataset directory (FITS images) Output directories for denoised images and residuals Example run: python test.py --config path/to/test_config.json These scripts and configurations were used in the analysis presented in Liu et al. 2025, ApJ. They support reproducibility of the galaxy model subtraction experiments and can be adapted for related tasks in astronomical image processing.
Cited by 1Open access
Liu, Rongrong · Zenodo (CERN European Organization for Nuclear Research) · 2025
Galaxy Denoising Autoencoder (DAE) This repository accompanies Liu et al. 2025, ApJ. It contains a TensorFlow-based Denoising Autoencoder (DAE) to model galaxies from observed galaxy images and perform galaxy model subtraction. The code, configuration files, and dataset generation scripts allow full reproducibility of the experiments described in the paper. Prerequisites Before running the scripts, ensure you have the following installed: Python 3.7 or later TensorFlow 2.x NumPy Astropy GALFIT (for dataset generation) Python dependencies can be installed with: pip install -r requirements.txt File Structure dae_model.py: Defines the autoencoder model (DAE class). fits_data_generator.py: Loads and preprocesses FITS data in batches for training. train.py: Runs the training process using settings from train_config.json. test.py: Evaluates the trained model on test data, outputting denoised and residual images. gen_gal_input.py: Generates input parameter files for GALFIT. run_galfit.py: Uses GALFIT to create simulated galaxy images from parameter files. add_background.py: Adds realistic sky backgrounds to simulated galaxy images. train_config.json: Configuration file for training (paths, parameters, model settings). test_config.json: Configuration file for testing (paths, output directories, pretrained model). Dataset Generation The dataset for training and testing is created by generating simulated galaxy images and combining them with realistic sky backgrounds. Step 1: Generate Simulated Galaxies Use gen_gal_input.py to create GALFIT parameter files and run_galfit.py to generate galaxy images. python gen_gal_input.py python run_galfit.py Step 2: Add Sky Backgrounds Use add_background.py to add clean sky backgrounds to simulated galaxies. These pairs of noisy (galaxy + sky) and clean (galaxy only) images form the training dataset. python add_background.py Output: A dataset of paired images (noisy vs. clean) ready for use in training/testing the DAE. Training Hardware Requirements Training on Google Colab with an L4 GPU takes ~3 hours with the default configuration. Larger models will require more resources. Training Configuration Settings are specified in train_config.json, including: Input/output directories Model architecture parameters Optimizer settings Training options (epochs, batch size, validation split, early stopping, checkpointing, etc.) Example run: python train.py --config path/to/train_config.json Testing Test Configuration Settings are stored in test_config.json, including: Path to trained model (.h5 file) Input test dataset directory (FITS images) Output directories for denoised images and residuals Example run: python test.py --config path/to/test_config.json These scripts and configurations were used in the analysis presented in Liu et al. 2025, ApJ. They support reproducibility of the galaxy model subtraction experiments and can be adapted for related tasks in astronomical image processing.
Cited by 0Open access
Stephanie Lofton · Zenodo (CERN European Organization for Nuclear Research) · 2024
Cited by 0Open access
Ethem Alpaydın · 2016
We discuss an autoencoder model in which the encoding and decoding functions are implemented by decision trees. We use the soft decision tree where internal nodes realize soft multivariate splits given by a gating function and the overall output is the average of all leaves weighted by the gating values on their path. The encoder tree takes the input and generates a lower dimensional representation in the leaves and the decoder tree takes this and reconstructs the original input. Ex-ploiting the continuity of the trees, autoencoder trees are trained with stochastic gradient descent. On handwritten digit and news data, we see that the autoencoder trees yield good reconstruction error compared to traditional autoencoder percep-trons. We also see that the autoencoder tree captures hierarchical representations at different granularities of the data on its different levels and the leaves capture the localities in the input space. 1
Cited by 0
Syeda Nazia Ashraf (17541222), Raheel Siddiqi (19923944), Humera Farooq (19923947) · Figshare · 2024
<div><p>Convolutional Neural Network (CNN)-based models are prone to adversarial attacks, which present a significant hurdle to their reliability and robustness. The vulnerability of CNN-based models may be exploited by attackers to launch cyber-attacks. An attacker typically adds small, carefully crafted perturbations to original medical images. When a CNN-based model receives the perturbed medical image as input, it misclassifies the image, even though the added perturbation is often imperceptible to the human eye. The emergence of such attacks has raised security concerns regarding the implementation of deep learning-based medical image classification systems within clinical environments. To address this issue, a reliable defense mechanism is required to detect adversarial attacks on medical images. This study will focus on the robust detection of pneumonia in chest X-ray images through CNN-based models. Various adversarial attacks and defense strategies will be evaluated and analyzed in the context of CNN-based pneumonia detection. From earlier studies, it has been observed that a single defense mechanism is usually not effective against more than one type of adversarial attack. Therefore, this study will propose a defense mechanism that is effective against multiple attack types. A reliable defense framework for pneumonia detection models will ensure secure clinical deployment, facilitating radiologists and doctors in their diagnosis and treatment planning. It can also save time and money by automating routine tasks. The proposed defense mechanism includes a convolutional autoencoder to denoise perturbed Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) adversarial images, two state- of-the-art attacks carried out at five magnitudes, i.e., ε (epsilon) values. Two pre-trained models, VGG19 and VGG16, and our hybrid model of MobileNetV2 and DenseNet169, called Stack Model, have been used to compare their results. This study shows that the proposed defense mechanism outperforms state-of-the-art studies. The PGD attack using the VGG16 model shows a better attack success rate by reducing overall accuracy by up to 67%. The autoencoder improves accuracy by up to 16% against PGD attacks in both the VGG16 and VGG19 models.</p></div>
Cited by 0Open access
2025
Cited by 0
Li Yuan (102305), Yang Chen (9737) · Figshare · 2025
<div> The role played by communication technology in daily life is gradually increasing. However, there are problems such as complex types of signals, huge amount of data and noise interference, and the recognition accuracy of existing modulation classification methods is low. Therefore, the study proposes a signal automatic modulation classification model based on fixed K-mean algorithm and denoising autoencoder. The model uses fixed K-mean algorithm for feature classification and optimizes median filtering algorithm using dynamic thresholding. The classifier is used to improve the recognition accuracy of specific signals, and long short-term memory and data random corruption denoising are used to optimize the autoencoder. The experiments indicated that the signal classification accuracy of the model were 17.6% and 16.8% higher than the other two models, respectively. The computational complexity of the improved model decreased dramatically, but the average classification accuracy was only 1.6% lower than that before the improvement. The communication overhead and training efficiency were better than the other models, and the number of parameters of the model was 1/2 of that of the pre-improvement model. The memory occupancy and running time were reduced by 335KB and 33ms, respectively. Compared to the other two models, the model’s average classification accuracy at a signal-to-noise ratio of 0 was 18.4% and 19.7% higher, respectively. As a result, the improved model effectively increases signal recognition accuracy, enhances model robustness, and significantly reduces computational complexity while ensuring real-time signal processing for communication computing. </div>
Cited by 0Open access
arsdmgcl · Zenodo (CERN European Organization for Nuclear Research) · 2024
autoencoder pdf Rating: 4.3 / 5 (2119 votes) Downloads: 5210 = = = = = CLICK HERE TO DOWNLOAD = = = = = A vanilla VAE is essentially an autoencoder that is trained with the standard autoencoder reconstruction objec tive between the With the autoencoder paradigm in mind, we began an inquiry into the question of what can shape a good, useful representation. We were looking for unsupervised learning Richard Zemel COMS Lecture Variational Autoencoders 9/ Observation Model. Neural networks trained to attempt to copy its input to its output. Speci - PDF An autoencoder is a specific type of a neural network, which is mainlydesigned to encode the input into a compressed and meaningful Find, read and cite all the research you need on Note: this requires a VAE, not just an ordinary autoencoder. If the data lie on a nonlinear surface, it makes more sense to use a nonlinear autoencoder, e.g., one that looks like following: If the data is highly nonlinear, one could add more hidden layers to the network to have a deep autoencoder. Input. Finally, we build on this to derive a sparse autoencoder. Contain two parts: Encoder: map the input to a hidden representation. It sort-of is supervised learning because we give the system explicit goal data (the output should be the same as the input), and it sort-of isn't supervised learning because we don't have any manually determined labels or targets on the inputs Sparse autoencoderIntroduction Supervised learning is one of the most powerful tools of AI, and has led to automatic zip code recognition, speech recognition, self-driving cars, and a continually improving understanding of the human genome. oder: map the hidden representation to the output. Consider training a generator network with maximum likelihood. Abstract With the advent of science and technology, it has been observed that autoencoder plays a vital role in unsupervised learning and in deep architectures for many tasks like transfer learning and other tasks We say that an autoencoder is an example of semi-supervised or self-supervised learning. Here we will learn about the architecture of autoencoder, their implementation, types of autoencoder like autoencoders, and speci cally on the linear autoencoder over the real numbers and the unrestricted Boolean autoencoder with binary f0;1gvariablesGeneral Autoencoder Today, we'll cover thevariational autoencoder (VAE), a generative model that explicitly learns a low-dimensional representation. It sort-of is supervised learning because we give the system explicit goal data (the output Two Autoencoder Training MethodsAutoencoder is a feed-forward non-recurrent neural net •With an input layer, an output layer and one or more hidden layers •Can be trained learning about what actually autoencoder is and how are they used. p(x) = Z p(z)p(xjz)dz View PDF Abstract: In this article, we will look at autoencoders. Autoencoders belong to a class of learning algorithms known as unsupervised learning. Because OneツュClass Variational Autoencoder. Despite its sig-ni cant successes, supervised learning today is still severely limited. Roger Grosse and Jimmy Ba CSC/ For any A2Aand B2B, the autoencoder transforms an input vector x2Fn into an output vector A B(x) 2Fn (Figure1). h Hidden representation (the code) Reconstruction Encoder Two Autoencoder Training MethodsAutoencoder is a feed-forward non-recurrent neural net •With an input layer, an output layer and one or more hidden layers •Can be trained using the same techniques •Compute gradients using back-propagation •Followed by minibatchgradient descentUnlike feedforwardnetworks, can also be trained using For any A2Aand B2B, the autoencoder transforms an input vector x2Fn into an output vector A B(x) 2Fn (Figure1). We say that an autoencoder is an example of semi-supervised or self-supervised learning. Sushreeta Tripathy and Muskaan Tabasum. Learn abstract features in an unsupervised way so you can apply them to a supervised task Unlabled data can be much more plentiful than labeled data Learn a semantically meaningful representation where you can, e.g., interpolate between di erent images The corresponding autoencoder problem is to nd A2Aand Then, we show how this is used to construct an autoencoder, which is an unsupervised learning algorithm. Unlike Autoencoder. The corresponding autoencoder problem is to nd A2Aand B2Bthat minimize the overall distortion function: minE(A;B) = min A;B Xm t=1 E(x t) = min A;B Xm t=1 A B(x t);x t (1) In the non auto-associative case, when external targets y Autoencoder: An Unsupervised Deep Learning Approach. We will discuss what they are, what the limitations are, the typical use cases, and we will look at some examples linear surface. This article covers the mathematics and the fundamental concepts of autoencoders.
Cited by 0Open access
Maciej Mikulski, Jarosław Duda · arXiv (Cornell University) · 2019
Enforcing distributions of latent variables in neural networks is an active subject. It is vital in all kinds of generative models, where we want to be able to interpolate between points in the latent space, or sample from it. Modern generative AutoEncoders (AE) like WAE, SWAE, CWAE add a regularizer to the standard (deterministic) AE, which allows to enforce Gaussian distribution in the latent space. Enforcing different distributions, especially topologically nontrivial, might bring some new interesting possibilities, but this subject seems unexplored so far. This article proposes a new approach to enforce uniform distribution on d-dimensional torus. We introduce a circular spring loss, which enforces minibatch points to be equally spaced and satisfy cyclic boundary conditions. As example of application we propose multiple-path morphing. Minimal distance geodesic between two points in uniform distribution on latent space of angles becomes a line, however, torus topology allows us to choose such lines in alternative ways, going through different edges of $[-π,π]^d$. Further applications to explore can be for example trying to learn real-life topologically nontrivial spaces of features, like rotations to automatically recognize 2D rotation of an object in picture by training on relative angles, or even 3D rotations by additionally using spherical features - this way morphing should be close to object rotation.
Cited by 0Open access
Eugen-Richard Ardelean (14648699), Andreea Coporîie (14769084), Ana-Maria Ichim (14648702) et al. · Figshare · 2023
<p><b>(a)</b> Diagram of the autoencoder components <b>(b)</b> Original spike and the reconstruction made by an autoencoder.</p>
Cited by 0Open access
arsdmgcl · Zenodo (CERN European Organization for Nuclear Research) · 2024
&lt;pre&gt;&lt;code&gt;\n&lt;p&gt;&lt;strong&gt;autoencoder pdf&lt;/strong&gt;&lt;br&gt;&lt;/p&gt;\n&lt;p&gt;Rating: 4.3 / 5 (2119 votes)&lt;br&gt;&lt;/p&gt;\n&lt;p&gt;Downloads: 5210&lt;br&gt;&lt;br&gt;&lt;/p&gt;\n &lt;p&gt;= = = = = \n&lt;strong&gt;&lt;a href="https://tds11111.com/21Nr9y?keyword=autoencoder pdf" target="_blank"&gt;CLICK HERE TO DOWNLOAD&lt;/a&gt;&lt;/strong&gt;\n = = = = = &lt;br&gt;&lt;br&gt;&lt;/p&gt;\n&lt;p&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;/p&gt;\n&lt;p&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;br&gt;&lt;/p&gt;\n&lt;p&gt;&lt;br&gt;&lt;br&gt;A vanilla VAE is essentially an autoencoder that is trained with the standard autoencoder reconstruction objec tive between the With the autoencoder paradigm in mind, we began an inquiry into the question of what can shape a good, useful representation. We were looking for unsupervised learning Richard Zemel COMS Lecture Variational Autoencoders 9/ Observation Model. Neural networks trained to attempt to copy its input to its output. Speci - PDF An autoencoder is a specific type of a neural network, which is mainlydesigned to encode the input into a compressed and meaningful Find, read and cite all the research you need on Note: this requires a VAE, not just an ordinary autoencoder. If the data lie on a nonlinear surface, it makes more sense to use a nonlinear autoencoder, e.g., one that looks like following: If the data is highly nonlinear, one could add more hidden layers to the network to have a deep autoencoder. Input. Finally, we build on this to derive a sparse autoencoder. Contain two parts: Encoder: map the input to a hidden representation. It sort-of is supervised learning because we give the system explicit goal data (the output should be the same as the input), and it sort-of isn't supervised learning because we don't have any manually determined labels or targets on the inputs Sparse autoencoderIntroduction Supervised learning is one of the most powerful tools of AI, and has led to automatic zip code recognition, speech recognition, self-driving cars, and a continually improving understanding of the human genome. oder: map the hidden representation to the output. Consider training a generator network with maximum likelihood. Abstract With the advent of science and technology, it has been observed that autoencoder plays a vital role in unsupervised learning and in deep architectures for many tasks like transfer learning and other tasks We say that an autoencoder is an example of semi-supervised or self-supervised learning. Here we will learn about the architecture of autoencoder, their implementation, types of autoencoder like autoencoders, and speci cally on the linear autoencoder over the real numbers and the unrestricted Boolean autoencoder with binary f0;1gvariablesGeneral Autoencoder Today, we'll cover thevariational autoencoder (VAE), a generative model that explicitly learns a low-dimensional representation. It sort-of is supervised learning because we give the system explicit goal data (the output Two Autoencoder Training MethodsAutoencoder is a feed-forward non-recurrent neural net •With an input layer, an output layer and one or more hidden layers •Can be trained learning about what actually autoencoder is and how are they used. p(x) = Z p(z)p(xjz)dz View PDF Abstract: In this article, we will look at autoencoders. Autoencoders belong to a class of learning algorithms known as unsupervised learning. Because OneツュClass Variational Autoencoder. Despite its sig-ni cant successes, supervised learning today is still severely limited. Roger Grosse and Jimmy Ba CSC/ For any A2Aand B2B, the autoencoder transforms an input vector x2Fn into an output vector A B(x) 2Fn (Figure1). h Hidden representation (the code) Reconstruction Encoder Two Autoencoder Training MethodsAutoencoder is a feed-forward non-recurrent neural net •With an input layer, an output layer and one or more hidden layers •Can be trained using the same techniques •Compute gradients using back-propagation •Followed by minibatchgradient descentUnlike feedforwardnetworks, can also be trained using For any A2Aand B2B, the autoencoder transforms an input vector x2Fn into an output vector A B(x) 2Fn (Figure1). We say that an autoencoder is an example of semi-supervised or self-supervised learning. Sushreeta Tripathy and Muskaan Tabasum. Learn abstract features in an unsupervised way so you can apply them to a supervised task Unlabled data can be much more plentiful than labeled data Learn a semantically meaningful representation where you can, e.g., interpolate between di erent images The corresponding autoencoder problem is to nd A2Aand Then, we show how this is used to construct an autoencoder, which is an unsupervised learning algorithm. Unlike Autoencoder. The corresponding autoencoder problem is to nd A2Aand B2Bthat minimize the overall distortion function: minE(A;B) = min A;B Xm t=1 E(x t) = min A;B Xm t=1 A B(x t);x t (1) In the non auto-associative case, when external targets y Autoencoder: An Unsupervised Deep Learning Approach. We will discuss what they are, what the limitations are, the typical use cases, and we will look at some examples linear surface. This article covers the mathematics and the fundamental concepts of autoencoders.&lt;/p&gt;&lt;/code&gt;&lt;/pre&gt;
Cited by 0Open access
Ozan İrsoy, Ethem Alpaydın · arXiv (Cornell University) · 2014
We discuss an autoencoder model in which the encoding and decoding functions are implemented by decision trees. We use the soft decision tree where internal nodes realize soft multivariate splits given by a gating function and the overall output is the average of all leaves weighted by the gating values on their path. The encoder tree takes the input and generates a lower dimensional representation in the leaves and the decoder tree takes this and reconstructs the original input. Exploiting the continuity of the trees, autoencoder trees are trained with stochastic gradient descent. On handwritten digit and news data, we see that the autoencoder trees yield good reconstruction error compared to traditional autoencoder perceptrons. We also see that the autoencoder tree captures hierarchical representations at different granularities of the data on its different levels and the leaves capture the localities in the input space.
Cited by 0Open access
Stephanie Lofton · Zenodo (CERN European Organization for Nuclear Research) · 2024
Cited by 0Open access
Chih-Chieh Hung (11001770), Ying-Ju Chen (296144) · Figshare · 2021
Cited by 0Open access