Generative Network Research Articles

Decoding the molecular, cellular, and functional heterogeneity of zebrafish intracardiac nervous system

The proper functioning of the heart relies on the intricate interplay between the central nervous system and the local neuronal networks within the heart itself. While the central innervation of the heart has been extensively studied, the organization and functionality of the intracardiac nervous system (IcNS) remain largely unexplored. Here, we present a comprehensive taxonomy of the IcNS, utilizing single-cell RNA sequencing, anatomical studies, and electrophysiological techniques. Our findings reveal a diverse array of neuronal types within the IcNS, exceeding previous expectations. We identify a subset of neurons exhibiting characteristics akin to pacemaker/rhythmogenic neurons similar to those found in Central Pattern Generator networks of the central nervous system. Our results underscore the heterogeneity within the IcNS and its key role in regulating the heart’s rhythmic functionality. The classification and characterization of the IcNS presented here serve as a valuable resource for further exploration into the mechanisms underlying heart functionality and the pathophysiology of associated cardiac disorders.

CryptoGAN: a new frontier in generative adversarial network-driven image encryption

<span lang="EN-IN">There is a growing need for an image encryption scheme, for huge amount of social media data or even the medical data to secure the privacy of the patients or the user. This study introduces a ground-breaking deep learning architecture named crypto generative adversarial networks (CryptoGAN), a novel architecture for generating cipher images. This architecture has the ability to generate both encrypted and decrypted images. The CryptoGAN system consists of an initial encryption network, a generative network that verifies the output against the desired domain, and a subsequent decryption phase. The generative adversarial networks (GAN) are utilised as the learning network to generate cipher images. This is achieved by training the neural network using images encrypted from a conventional image encryption scheme such as advanced encryption standards (AES), and learning from the resulting losses. This enhances security measures when dealing with a large dataset of photos. The assessment of the performance metrics of the encrypted image, including entropy, histogram, correlation plot, and vulnerability to assaults, demonstrates that the suggested generative network may get a higher level of security.</span>

Micromechanics and microstructure based machine learning approach: Unveiling the role of porosity and hydrated phases on the tensile behaviour of cement pastes

The fracture process and the mechanical performance of the cement paste rely significantly on its tensile behaviour. Machine learning methods have been recently developed to predict this behaviour based on numerous experimental measurements. Due to multi-phase heterogeneous nature of cement paste microstructure, evaluation of the role of each chemical phase on the cracking of cement paste is necessary to develop optimized binder systems. Numerous experimental tests would be necessary to determine the best cement pastes in a considered environment. Therefore, in this study, a hybrid approach based on an ensemble machine learning model and a microstructure informed micromechanical model is proposed for predicting the tensile behaviour of hydrated cement pastes. The model can be used as a quick prediction tool for tensile strength of a given microstructure or for deep analysis on the optimization of binder system. A Deep Neural Networks (DNN) model was developed by training and validating a large dataset detailing the chemical phase composition, phase volume fractions and the tensile strength of the cement pastes. This dataset covers an extensive range of Portland cements, hydration ages and water-to-cement ratios. It was developed by simulating cement thermodynamics and tensile behaviour using a damage micromechanical model; which is otherwise almost impossible to obtain through experimental means. Additionally, two generative algorithms, tabular generative adversial networks (TGAN) and conditional generative adversial networks (CTGAN), were employed for dataset augmentation as a sampling method for the sensitivity analysis. This machine learning study provides an in-depth investigation to identify the role of volume fraction of solid phases (C-S-H, ettringite, portlandite and hydrogarnet) and micro porosity (dry and filled) in affecting the tensile behaviour of hydrated cement paste. Optimization assessment of the microstructure was then performed using micromechanical analysis and damage propagation in hydrated cement paste phases and porous network.

DeepTensor: Low-Rank Tensor Decomposition With Deep Network Priors.

DeepTensor is a computationally efficient framework for low-rank decomposition of matrices and tensors using deep generative networks. We decompose a tensor as the product of low-rank tensor factors (e.g., a matrix as the outer product of two vectors), where each low-rank tensor is generated by a deep network (DN) that is trained in a self-supervised manner to minimize the mean-square approximation error. Our key observation is that the implicit regularization inherent in DNs enables them to capture nonlinear signal structures (e.g., manifolds) that are out of the reach of classical linear methods like the singular value decomposition (SVD) and principal components analysis (PCA). Furthermore, in contrast to the SVD and PCA, whose performance deteriorates when the tensor's entries deviate from additive white Gaussian noise, we demonstrate that the performance of DeepTensor is robust to a wide range of distributions. We validate that DeepTensor is a robust and computationally efficient drop-in replacement for the SVD, PCA, nonnegative matrix factorization (NMF), and similar decompositions by exploring a range of real-world applications, including hyperspectral image denoising, 3D MRI tomography, and image classification. In particular, DeepTensor offers a 6 dB signal-to-noise ratio improvement over standard denoising methods for signal corrupted by Poisson noise and learns to decompose 3D tensors 60 times faster than a single DN equipped with 3D convolutions.

Geological type recognition for shield machine using a semi-supervised variational auto-encoder-based adversarial method

In the process of tunneling, accurate and timely recognition of the geological type is significant to optimize the control parameters of the tunneling machine, improving tunneling efficiency and avoiding accidents. The shield machine operator in shield tunneling machine cannot directly observe the geological environment due to the closed working environment, so the soft method that can indirectly recognize the geological type by the machine parameters has become a research hotspot. However, most current soft methods use only a small amount of labeled data for supervised learning, and large amounts of unlabeled data is wasted. In order to use all data to improve the recognition performance of the classifier, a semi-supervised variational auto-encoder-based adversarial method (VAE-EMGAN) is proposed. Firstly, 50 parameters associated with geological types are selected and pre-processed, then the Variational Auto-Encoder (VAE) is trained by unlabeled data, and the generated part of VAE is added to the structure of Enhanced Multi-Classification Adversarial Generative Network (EMGAN) as a generator. Finally, the recognition accuracy of classifier is improved through adversarial training with labeled data, unlabeled data and generated data. We used data from upper and lower tunnels in Singapore to create two tasks to verify the validity and generalization performance of VEVE-EMGAN. The results show that the proposed model not only achieves high accuracy of all test sets on both tasks, but also has much better generalization performance than other models. Mean accuracy is 10.82%, 17.68%, 11.05%, 17.72%, 17.45%, 12.68% and 5.27% higher than SVM, KNN, RF, XGBoost, MLP, DNN and CNN respectively of test set 2 on task A; Mean accuracy is 13.06%, 12.80%, 7.64%, 18.31%, 8.74%, 7.94% and 4.05% higher than SVM, KNN, RF, XGBoost, MLP, DNN and CNN respectively of test set 2 on task B. In particular, the performance of the adversarial trained classifier is better than that has the same structure but separately trained classifier. Therefore, this method can use unlabeled data for adversarial training to improve the classification accuracy and generalization performance of the classifier, which has important implications for engineering practice.

LAIgnd: Revolutionizing Drug Discovery with Advanced AI-Driven Molecule Generation

De novo molecular generation is crucial for advancing drug discovery and chemical research. This accelerates the search for new drug candidates and deepens our understanding of molecular diversity. The development of deep learning has propelled and expedited the de novo molecular generation. Generative networks, particularly Variational Autoencoders (VAEs), can randomly produce new molecules and modify molecular structures to enhance specific chemical properties, which are essential for advancing drug discovery. Although VAEs offer numerous advantages, they are hindered by limitations that affect their capacity to optimize properties and decode syntactically valid molecules. To address these challenges, we present LAIgnd, a de novo drug molecule generation model that implements a custom β-CVAE architecture conditioned on protein sequences and SELFIES input. Extensive experiments have shown that LAIgnd generates a wide variety of valid, novel, and effective molecules for complex and simple diseases, demonstrating its robustness and generalization capabilities. Additionally, by employing molecular docking, toxicity, similarity, and synthetic accessibility experiments, we demonstrated the drug-likeness and effectiveness of the generated molecules. The ability of our model to generate novel and diverse compounds was illustrated by a case study focusing on Lung Cancer. A total of four hundred (400) molecules were generated by LAIgnd, with a high number of molecules exhibiting strong inhibitory activity against the Epidermal Growth Factor receptor, as indicated by binding affinities. LAIgnd provides new insights into future directions to enhance therapeutics for complex and simple diseases by generating high-quality multi-target molecules for drug discovery. Key words: De novo molecular generation, Drug discovery, Variational Autoencoders (VAEs), SELFIES, Protein sequences. CISDI Journal Reference Format Obi, E.D., Yentumi, J.A., Mbatuegwu, D., Omotuyi, O.I., Ajayi, O.O. & Nwokoro, A. (2024): LAIgnd: Revolutionizing Drug Discovery with Advanced AI-Driven Molecule Generation. Computing, Information Systems, Development Informatics & Allied Research Journal. Vol 15 No 4, Pp 1-10. Available online at www.isteams.net/cisdijournal. dx.doi.org/10.22624/AIMS/CISDI/V15N3P4

A globally optimized fault diagnosis model based on generative flow model for imbalanced data

Abstract In the actual scenario of fault diagnosis based on deep learning, the diagnosis accuracy is often affected by the lack of fault state data, so the processing of imbalanced data is always a significant challenge. generative adversarial networks (GAN) and denoising diffusion probability models (DDPM) are widely used for data augmentation. However, GAN often shows sensitivity and instability in the training process, and the sample generation speed of DDPM is slow due to the steps requiring multiple iterations–both of which are limiting factors. To solve these problems, we introduce the generative flow network with invertible 1 × 1 convolutions (GLOW) into fault diagnosis. The GLOW model is optimized by maximum likelihood estimation and does not require multiple iterations to generate samples, avoiding the problems faced by GAN and DDPM. In order to generate balanced data explicitly, we propose a condition GLOW (CGLOW) to provide class-balanced samples in real time throughout the framework. On the other hand, using the reversibility of CGLOW, we design an end-to-end fault diagnosis framework that is globally optimized to mitigate the decline in diagnostic accuracy caused by the separation of generation and diagnosis and simplify the steps of fault diagnosis. In addition, to accommodate the non-stationary characteristics of fault signals, we propose a new data transformation method to improve the feature mining ability of the model and the diagnostic accuracy. Finally, we conduct extensive experiments to validate the superiority of the proposed approach. The experimental results demonstrate that our method outperforms existing ones.

Continuous Generative Neural Networks: A Wavelet-Based Architecture in Function Spaces

In this work, we present and study Continuous Generative Neural Networks (CGNNs), namely, generative models in the continuous setting: the output of a CGNN belongs to an infinite-dimensional function space. The architecture is inspired by DCGAN, with one fully connected layer, several convolutional layers and nonlinear activation functions. In the continuous L 2 setting, the dimensions of the spaces of each layer are replaced by the scales of a multiresolution analysis of a compactly supported wavelet. We present conditions on the convolutional filters and on the nonlinearity that guarantee that a CGNN is injective. This theory finds applications to inverse problems, and allows for deriving Lipschitz stability estimates for (possibly nonlinear) infinite-dimensional inverse problems with unknowns belonging to the manifold generated by a CGNN. Several numerical simulations, including signal deblurring, illustrate and validate this approach.

Universal scaling laws in quantum-probabilistic machine learning by tensor network towards interpreting representation and generalization powers

Abstract Interpreting the representation and generalization powers has been a long-standing issue in the field of machine learning (ML) and artificial intelligence. This work contributes to uncovering the emergence of universal scaling laws in quantum-probabilistic ML. We take the generative tensor network (GTN) in the form of a matrix product state as an example and show that with an untrained GTN (such as a random TN state), the negative logarithmic likelihood (NLL) $L$ generally increases linearly with the number of features $M$, i.e., $L \simeq k M + const$. This is a consequence of the so-called “catastrophe of orthogonality,” which states that quantum many-body states tend to become exponentially orthogonal to each other as $M$ increases. We reveal that while gaining information through training, the linear scaling law is suppressed by a negative quadratic correction, leading to $L \simeq \beta M - \alpha M^2 + const$. The scaling coefficients exhibit logarithmic relationships with the number of training samples and the number of quantum channels $\chi$. The emergence of the quadratic correction term in NLL for the testing (training) set can be regarded as evidence of the generalization (representation) power of GTN. Over-parameterization can be identified by the deviation in the values of $\alpha$ between training and testing sets while increasing $\chi$. We further investigate how orthogonality in the quantum feature map relates to the satisfaction of quantum probabilistic interpretation, as well as to the representation and generalization powers of GTN. The unveiling of universal scaling laws in quantum-probabilistic ML would be a valuable step toward establishing a white-box ML scheme interpreted within the quantum probabilistic framework.

Integrating symbolic reasoning into neural generative models for design generation

Design generation requires tight integration of neural and symbolic reasoning, as good design must meet explicit user needs and honor implicit rules for aesthetics, utility, and convenience. Current automated design tools driven by neural networks produce appealing designs, but cannot satisfy user specifications and utility requirements. Symbolic reasoning tools, such as constraint programming, cannot perceive low-level visual information in images or capture subtle aspects such as aesthetics. We introduce the Spatial Reasoning Integrated Generator (SPRING) for design generation. SPRING embeds a neural and symbolic integrated spatial reasoning module inside the deep generative network. The spatial reasoning module samples the set of locations of objects to be generated from a backtrack-free distribution. This distribution modifies the implicit preference distribution, which is learned by a recurrent neural network to capture utility and aesthetics. The sampling from the backtrack-free distribution is accomplished by a symbolic reasoning approach, SampleSearch, which zeros out the probability of sampling spatial locations violating explicit user specifications. Embedding symbolic reasoning into neural generation guarantees that the output of SPRING satisfies user requirements. Furthermore, SPRING offers interpretability, allowing users to visualize and diagnose the generation process through the bounding boxes. SPRING is also adept at managing novel user specifications not encountered during its training, thanks to its proficiency in zero-shot constraint transfer. Quantitative evaluations and a human study reveal that SPRING outperforms baseline generative models, excelling in delivering high design quality and better meeting user specifications.

SPFont: Stroke potential features embedded GAN for Chinese calligraphy font generation

Chinese calligraphy font generation is an extremely challenging problem. Firstly, Chinese calligraphy fonts have complex structures. The accuracy and artistic quality of the generated fonts will be affected by the order and layout of the strokes as well as the relationships between them. Secondly, the number of Chinese characters is large, but existing calligraphy works are scarce. Hence, it is difficult to establish a comprehensive and high-quality Chinese calligraphy dataset. In this paper, we propose an unsupervised calligraphy font generation network SPFont. It is based on a generative adversarial network (GAN) framework. The generator includes a style feature encoder, a content feature encoder, a stroke potential feature fusion module (SPFM) and a decoder. The SPFM module, by overlaying lower-level style and content features, better preserves fine details of the font such as stroke thickness, curve shapes and other characteristics. The SPFM module and the extracted style features are fused and then fed into the decoder, allowing it to consider the influence of style, content and stroke potential simultaneously during the generation process. Experimental results demonstrate that our model generates Chinese calligraphy fonts with higher quality compared to previous methods.

The Use of Deepfake Technology in China: Problems of Legal Regulation and Ways to Solve Them

The core of the deepfake technology is based on a generative-adversarial network built on a combination of two neural networks: a generative network (network G) creates samples, a discriminative network (network D) tries to distinguish correct («genuine») samples from incorrect ones. Networks G and D compete with each other thousands or even millions of times until network G improves its performance. Thus, the network D will no longer be able to distinguish real data from fake data. With the development of big data and machine learning technologies, the scenario for using deepfake technology has gradually changed from creating sound models and imitating text to deep video forgery. For a long time, images modified using traditional Photoshop and other technologies were easily recognized. Deepfake technology changed this situation, making it more difficult to identify fakes. As an important technological innovation in the field of artificial intelligence, deepfake technology is widely used in various areas of society, creating enormous applied value. However, any technology is a double-edged sword. The use of deepfake technology poses a great threat to personal privacy, property security and even national security. In order to find a balance between technological innovation and risk prevention and control, countries around the world are actively exploring various ways to manage. The paper describes the main risks posed by modern deepfake technology, provides an overview of legal regulation in this area in China and offers an effective way to solve problems.

AltGosling: Automatic Generation of Text Descriptions for Accessible Genomics Data Visualization.

Biomedical visualizations are key to accessing biomedical knowledge and detecting new patterns in large datasets. Interactive visualizations are essential for biomedical data scientists and are omnipresent in data analysis software and data portals. Without appropriate descriptions, these visualizations are not accessible to all people with blindness and low vision, who often rely on screen reader accessibility technologies to access visual information on digital devices. Screen readers require descriptions to convey image content. However, many images lack informative descriptions due to unawareness and difficulty writing such descriptions. Describing complex and interactive visualizations, like genomics data visualizations, is even more challenging. Automatic generation of descriptions could be beneficial, yet current alt text generating models are limited to basic visualizations and cannot be used for genomics. We present AltGosling, an automated description generation tool focused on interactive data visualizations of genome-mapped data, created with the grammar-based genomics toolkit Gosling. The logic-based algorithm of AltGosling creates various descriptions including a tree-structured navigable panel. We co-designed AltGosling with a blind screen reader user (co-author). We show that AltGosling outperforms state-of-the-art large language models and common image-based neural networks for alt text generation of genomics data visualizations. As a first of its kind in genomic research, we lay the groundwork to increase accessibility in the field. The source code, examples, and interactive demo are accessible under the MIT License at https://github.com/gosling-lang/altgosling. The package is available at https://www.npmjs.com/package/altgosling. Supplementary data are available at Bioinformatics online.

Radiomics-guided generative adversarial network for automatic primary target volume segmentation for nasopharyngeal carcinoma using computed tomography images.

Automatic primary gross tumor volume (GTVp) segmentation for nasopharyngeal carcinoma (NPC) is a quite challenging task because of the existence of similar visual characteristics between tumors and their surroundings, especially on computed tomography (CT) images with severe low contrast resolution. Therefore, most recently proposed methods based on radiomics or deep learning (DL) is difficult to achieve good results on CT datasets. A peritumoral radiomics-guided generative adversarial network (PRG-GAN) was proposed to address this challenge. A total of 157 NPC patients with CT images was collected and divided into training, validation, and testing cohorts of 108, 9, and 30 patients, respectively. The proposed model was based on a standard GAN consisting of a generator network and a discriminator network. Morphological dilation on the initial segmentation results from GAN was first conducted to delineate annular peritumoral region, in which radiomics features were extracted as priori guide knowledge. Then, radiomics features were fused with semantic features by the discriminator's fully connected layer to achieve the voxel-level classification and segmentation. The dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and average symmetric surface distance (ASSD) were used to evaluate the segmentation performance using a paired samples t-test with Bonferroni correction and Cohen's d (d) effect sizes. A two-sided p-value of less than 0.05 was considered statistically significant. The model-generated predictions had a high overlap ratio with the ground truth. The average DSC, HD95, and ASSD were significantly improved from 0.80±0.12, 4.65±4.71mm, and 1.35±1.15mm of GAN to 0.85±0.18 (p=0.001, d=0.71), 4.15±7.56mm (p=0.002, d=0.67), and 1.11±1.65mm (p<0.001, d=0.46) of PRG-GAN, respectively. Integrating radiomics features into GAN is promising to solve unclear border limitations and increase the delineation accuracy of GTVp for patients with NPC.

GAN-based pseudo random number generation optimized through genetic algorithms

Pseudo-random number generators (PRNGs) are deterministic algorithms that generate sequences of numbers approximating the properties of random numbers, which are widely utilized in various fields. In this paper, we present a Genetic Algorithm Optimized Generative Adversarial Network (hereinafter referred to as GAGAN), which is designed for the effective training of discrete generative adversarial networks. In situations where non-differentiable activation functions, such as the modulo operation, are employed and traditional gradient-based backpropagation methods are inapplicable, genetic algorithms are utilized to optimize the parameters of the generator network. Based on this framework, we propose a novel recursive PRNG. Given that a PRNG can be constructed from one-way functions and their associated hardcore predicates, our proposed generator consists of two neural networks that simulate these functions and serve as the state transition function and the output function, respectively. The proposed PRNG has been rigorously tested using stringent benchmarks such as the NIST Statistical Test Suite (SP800-22) and the Chinese standard for random number generation (GM/T 0005-2021). Additionally, it has demonstrated outstanding performance in terms of Hamming distance. The results indicate that the proposed GAN-based PRNG has achieved a high degree of randomness and is highly sensitive to variations in the input.

Development of a cerebellar ataxia diagnosis model using conditional GAN-based synthetic data generation for visuomotor adaptation task

This study proposes a synthetic data generation model to create a classification framework for cerebellar ataxia patients using trajectory data from the visuomotor adaptation task. The classification objectives include patients with cerebellar ataxia, age-matched normal individuals, and young healthy subjects. Synthetic data for the three classes is generated based on class conditions and random noise by leveraging a combination of conditional adversarial generative neural networks and reconstruction networks. This synthetic data, alongside real data, is utilized as training data for the patient classification model to enhance classification accuracy. The fidelity of the synthetic data is assessed visually to measure the validity and diversity of the generated data qualitatively while quantitatively evaluating distribution similarity to real data. Furthermore, the clinical efficacy of the patient classification model employing synthetic data is demonstrated by showcasing improved classification accuracy through a comparative analysis between results obtained using solely real data and those obtained when both real and synthetic data are utilized. This methodological approach holds promise in addressing data insufficiency in the digital healthcare domain, employing deep learning methodologies, and developing early disease diagnosis tools.

WGAN-based multi-structure segmentation of vertebral cross-section MRI using ResU-Net and clustered transformer

The vertebral foramen, lamina, and vertebral body are three critical components of the spine structure, essential for maintaining spinal connectivity and stability. Accurately segmenting lumbar structures such as the vertebral body, vertebral foramen, and lamina in MRI cross-sections helps doctors better understand and diagnose the pathological causes of spine-related diseases. This study presents a multi-structure semantic segmentation method for vertebral transverse section MRI slices using WGAN with a residual U-Net and clustered Transformer. The generator network was replaced with a combination of a residual U-Net and a clustered Transformer-based segmentation network. The enhanced U-Net encoder, utilizing dilated convolutions and residual structures, improved multi-scale feature extraction capabilities. Meanwhile, the clustered Transformer structure, with reduced progressive linear complexity, ensured the extraction of global positional information. The results of multiple experiments show that the Dice coefficient for vertebral body segmentation increased by 3.1%, the Hausdorff distance decreased by 0.6 mm, mIOU improved by 4.1–96.2%, and PPV increased by 2.0–98.8% compared to mainstream segmentation models. These improvements are statistically significant (p < 0.05).Ablation experiments further validated the effectiveness of the proposed enhanced modules in improving segmentation accuracy for the three target structures.

Inpainting of cloud-occlusion sea surface temperature image from a novel generative network using multi-scale physical constraints

Inpainting of cloud-occlusion sea surface temperature image from a novel generative network using multi-scale physical constraints

Mitigating coarse spatial reconstruction to generate missing bands for the HLS dataset

Abstract. The Harmonized LandSat-Sentinel (HLS) dataset has significantly advanced Earth Observation by integrating data from Landsat and Sentinel satellites. However, challenges persist in achieving spectral band parity between LandSat and Sentinel-derived HLS products. This paper presents an extended investigation aimed at enhancing spatial reconstruction accuracy to enable spectral band parity within HLS products. Building upon our previous work, which utilized generative neural networks to address partial feature mismatches between S30 and L30 products, we introduce a refined approach that fully integrates a Self-Supervised Learning (SSL)-pretrained encoder into a U-Net architecture. This method aims to access multi-scale features and improve spatial reconstruction accuracy, addressing the limitations in spatial resolution observed in our earlier study. Our methodology incorporates a comprehensive ablation study to assess various SSL-pretrained backbone architectures. Preliminary results demonstrate significant improvements in spatial reconstruction accuracy compared to our previous work. The adapted U-Net architecture, leveraging SSL-pretrained encoders, shows enhanced capability in capturing intricate spatial features within the HLS dataset. Our experiments demonstrate a substantial improvement in spatial resolution and feature reconstruction for L30 products, particularly in bands not natively present in Landsat data, paving the way for more accurate multi-sensor analyses.

Multi-channel neural audio decorrelation using generative adversarial networks

The degree of correlation between the sounds received by the ears significantly influences the spatial perception of a sound image. Audio signal decorrelation is, therefore, a commonly used tool in various spatial audio rendering applications. In this paper, we propose a multi-channel extension of a previously proposed decorrelation method based on generative adversarial networks. A separate generator network is employed for each output channel. All generator networks are optimized jointly to obtain a multi-channel output signal with the desired properties. The training objective includes a number of individual loss terms to control both the input-output and the inter-channel correlation as well as the quality of the individual output channels. The proposed approach is trained on music signals and evaluated both objectively and through formal listening tests. Thereby, a comparison with two classical signal processing-based multi-channel decorrelators is performed. Additionally, the influence of the number of output channels, the individual loss term weightings, and the employed training data on the proposed method’s performance is investigated.