Traditional Deep Neural Networks Research Articles

Reliable data transmission for a VANET-IoIT architecture: A DNN approach

The challenges and resilience of vehicular ad hoc network (VANET) and deep neural network (DNN) hybrid architectures in terms of reliability in smart cities have attracted much global interest stemming from the rollout of the next generation of intelligent networks. In this paper, we propose a novel distributed DNN (D-DNN) scheme with blockchain to support the Internet of Intelligent Things (IoIT) infrastructure in the VANET environment of the future. In particular, because the communication links between edge nodes are very unstable in VANETs, a new neuro-fuzzy server that serves the dual roles of finding reliable links between edge nodes and performing optimal routing path selection is proposed. Next, a blockchain layer is employed at the edge nodes, which are initially scrutinized before establishing communication links to ensure reliability during data transfer. Then, the proposed D-DNN (PD-DNN) scheme is applied to enhance the performance of the VANET-IoIT architecture by improving the data flow and convergence rate and mitigating erratic variations in output. To address reliability concerns, the coverage probability (CP) metric is investigated as a measure of network connectivity. Furthermore, we present an analysis of the PD-DNN scheme in comparison with the traditional DNN (T-DNN) scheme. Finally, simulation results for VANET-IoIT scenarios show that, subject to data protection and privacy constraints, the CP values corresponding to different communication links are improved to a greater extent under our scheme than under the traditional scheme, demonstrating the feasibility of the proposed scheme.

Physics-informed neural network integrate with unclosed mechanism model for turbulent mass transfer

Turbulent mass transfer is widespread in chemical engineering processes including separation, reaction, and others. Traditionally, turbulent mass transfer processes can be simulated by computational fluid dynamics (CFD) methods. However, CFD methods require adequate turbulent models for the fluid flow and mass (and heat) transfer, which are normally difficult to develop for reliable solutions. Moreover, the CFD simulation is computationally intensive and hard to use in the cases like process optimization or analysis where the simulation needs to be called repeatedly. In this paper, physics-informed neural network (PINN) is developed and trained by unclosed mechanism model and sparse observation data of turbulent mass transfer process. The PINN method shows stronger generalization ability in solving the velocity, pressure and concentration fields in a turbulent mass transfer process than traditional deep neural network (DNN) method and turbulent Schmidt model. Under different boundary conditions, the PINN developed in the present paper can instantly predict the concentration distributions with sufficient accuracy and be used for inverse computation for estimating the turbulent viscosity and mass diffusivity as output results. The PINN is also capable of handling data noise by adjusting parameters, suggesting its potential in integrating experimental data.

Joint MCS Adaptation and RB Allocation in Cellular Networks Based on Deep Reinforcement Learning With Stable Matching

Joint modulation-coding scheme (MCS) adaptation and resource block (RB) allocation is an effective approach to guarantee different quality of service (QoS) requirements of all UEs under dynamic network environments. In this paper, we consider a fifth generation (5 G) cellular network with time-varying wireless channels, in which the BS serves multiple user equipments (UEs) under limited available RBs. We aim to minimize the total RB consumption subject to the rigorous constraints of each UE's QoS requirement. To attain this objective, this paper puts forth an online learning technique, referred to as integrated Deep Reinforcement learning and stable Matching (DeepRM), in the sense that the MCS adaptation and RB allocation decisions are conducted without acquiring the real-time channel quality indicator (CQI) feedback. DeepRM is a closed-loop framework, where the output of deep reinforcement learning (DRL) is imported into the stable matching to guide optimal RB allocation whilst the output of stable matching is fed into the DRL framework to assist efficient MCS decision-making. Specifically, in DeepRM, we first develop a powerful DRL algorithm, termed as Action-and-Reward Branching Deep Q-network (ARBDQ), by incorporating the action branch architecture into conventional DRL and modifying the traditional deep neural network training mechanism, to perform judicious MCS decisions on different links in parallel. Then, a new many-to-one stable matching algorithm, called adaptive deferred acceptance, is exploited to dynamically adjust the RB quota of each UE in a computationally efficient fashion. Simulation results demonstrate that compared with ACO-HM, OLLA-ADA, and ARBDQ-Random algorithms, DeepRM induces much less RB consumption while guaranteeing the QoS requirements of all UEs in various network scenarios. Furthermore, under miscellaneous QoS requirement, number of UEs, and CQI reporting period setups, DeepRM is more robust than other baselines.

A feature transfer model with Mixup and contrastive loss in domain generalization

When domains, which represent underlying data distributions, differ between training and test datasets, traditional deep neural networks suffer from a substantial drop in their performance. Domain generalization methods aim to boost generalizability on an unseen target domain by using only training data from source domains. Mainstream domain generalization algorithms usually make modifications on some popular feature extraction networks such as ResNet, or add more complex parameter modules after the feature extraction networks. Popular feature extraction networks are usually well pre-trained on large-scale datasets, so they have strong feature extraction abilities, while modifications can weaken such abilities. Adding more complex parameter modules results in a deeper network and is much more computationally demanding. In this paper, we propose a novel feature transfer model based on popular feature extraction networks in domain generalization, without making any changes or adding any module. The generalizability of this feature transfer model is boosted by incorporating a contrastive loss and a data augmentation strategy (i.e., Mixup), and a new sample selection strategy is proposed to coordinate Mixup and contrastive loss. Experiments on the benchmarks PACS and Domainnet demonstrate the superiority of our proposed method against conventional domain generalization methods.

Advancing task recognition towards artificial limbs control with ReliefF-based deep neural network extreme learning

In the rapidly advancing field of biomedical engineering, effective real-time control of artificial limbs is a pressing research concern. Addressing this, the current study introduces a pioneering method for augmenting task recognition in prosthetic control systems, combining a ReliefF-based Deep Neural Networks (DNNs) approach. This paper has leveraged the MILimbEEG dataset, a comprehensive rich source collection of EEG signals, to calculate statistical features of Arithmetic Mean (AM), Standard Deviation (SD), and Skewness (S) across various motor activities. Supreme Feature Selection (SFS), of the adopted time-domain features, was performed using the ReliefF algorithm. The highest scored DNN-ReliefF developed model demonstrated remarkable performance, achieving accuracy, precision, and recall rates of 97.4 %, 97.3 %, and 97.4 %, respectively. In contrast, a traditional DNN model yielded accuracy, precision, and recall rates of 50.8 %, 51.1 %, and 50.8 %, highlighting the significant improvements made possible by incorporating SFS. This stark contrast underscores the transformative potential of incorporating ReliefF, situating the DNN-ReliefF model as a robust platform for forthcoming advancements in real-time prosthetic control systems.

A novel multi-scale loss function for classification problems in machine learning

We introduce two-scale loss functions for use in various gradient descent algorithms applied to classification problems via deep neural networks. This new method is generic in the sense that it can be applied to a wide range of machine learning architectures, from deep neural networks to support vector machines for example. These two-scale loss functions allow to focus the training onto objects in the training set which are not well classified. This leads to an increase in several measures of performance for appropriately-defined two-scale loss functions with respect to the more classical cross-entropy when tested on traditional deep neural networks on the MNIST, CIFAR10, and CIFAR100 data-sets.

Non-intrusive personal thermal comfort modeling: A machine learning approach using infrared face recognition

At present, non-intrusive personal thermal comfort models are receiving more and more attention. Non-intrusive sensing technology is used to accurately capture the real-time thermal state of occupants indoors so as to construct personal thermal comfort models by data-driven methods. This study developed a non-intrusive personal thermal comfort model using machine learning techniques combined with infrared facial recognition. Firstly, the Charlotte-ThermalFace database was used to extract the temperatures from six regions of interest on the face using infrared face recognition and key point extraction algorithms. Subsequently, the feature importance of the variables was calculated by random forest (RF) and gradient boosting decision tree (GBDT) respectively to explore the key parameters influencing the prediction performance of personal thermal preferences. Finally, the performance of 12 machine learning models was systematically compared, including 6 traditional models, 5 ensemble models, and 1 broad model, based on precision, recall, F1 score and macro-F1 score. The results show that the ensemble learning models and the broad learning (BL) model perform better than the traditional models by using the full training dataset size. Secondly, the BL model is applied for the first time as an alternative to deep network models for thermal preference prediction, with a prediction precision of 90.44%. Compared with traditional deep neural networks (DNN) model, it has lower computational complexity and faster training speed. Furthermore, BL and deep cascade forest (DCF) have significant advantages over other models in predicting thermal preference with different data subsets. Overall, the results of this study provide a reference for non-intrusive personal thermal comfort modeling that can be used to optimize building thermal environments.

EDKSANet: An Efficient Dual-Kernel Split Attention Neural Network for the Classification of Tibetan Medicinal Materials

Tibetan medicine has received wide acclaim for its unique diagnosis and treatment methods. The identification of Tibetan medicinal materials, which are a vital component of Tibetan medicine, is a key research area in this field. However, traditional deep learning-based visual neural networks face significant challenges in efficiently and accurately identifying Tibetan medicinal materials due to their large number, complex morphology, and the scarcity of public visual datasets. To address this issue, we constructed a computer vision dataset with 300 Tibetan medicinal materials and proposed a lightweight and efficient cross-dimensional attention mechanism, the Dual-Kernel Split Attention (DKSA) module, which can adaptively share parameters of the kernel in both spatial and channel dimensions. Based on the DKSA module, we achieve efficient unification of convolution and self-attention under the CNN architecture and develop a new lightweight backbone architecture, EDKSANet, to provide enhanced performance for various computer vision tasks. As compared to RedNet, the top-1 accuracy is improved by 1.2% on an ImageNet dataset, and a larger margin of +1.5 box AP for object detection and an improvement of +1.3 mask AP for instance segmentation on MS-COCO dataset are obtained. Moreover, EDKSANet achieved excellent classification performance on the Tibetan medicinal materials dataset, with an accuracy of up to 96.85%.

Infrared Fault Classification Based on the Siamese Network

The rapid development of solar energy technology has led to significant progress in recent years, but the daily maintenance of solar panels faces significant challenges. The diagnosis of solar panel failures by infrared detection devices can improve the efficiency of maintenance personnel. Currently, due to the scarcity of infrared solar panel failure samples and the problem of unclear image effective features, traditional deep neural network models can easily encounter overfitting and poor generalization performance under small sample conditions. To address these problems, this paper proposes a solar panel failure diagnosis method based on an improved Siamese network. Firstly, two types of solar panel samples of the same category are constructed. Secondly, the images of the samples are input into the feature model combining convolution, adaptive coordinate attention (ACA), and the feature fusion module (FFM) to extract features, learning the similarities between different types of solar panel samples. Finally, the trained model is used to determine the similarity of the input solar image, obtaining the failure diagnosis results. In this case, adaptive coordinate attention can effectively obtain interested effective feature information, and the feature fusion module can integrate the different effective information obtained, further enriching the feature information. The ACA-FFM Siamese network method can alleviate the problem of insufficient sample quantity and effectively improve the classification accuracy, achieving a classification accuracy rate of 83.9% on an open-accessed infrared failure dataset with high similarity.

Rgfc-Forest: An enhanced deep forest method towards small-sample fault diagnosis of electromechanical system

Deep forest models offer a promising alternative to traditional deep neural networks by demanding fewer training samples and hyperparameters. However, existing deep forest fault diagnosis models encounter persistent challenges such as insufficient representation of multi-grained spatial information and redundancy of cascaded forest features. To address the above challenges, an enhanced deep forest method called random multi-grained fusion cascade forest (rgfc-Forest) is presented for fault diagnosis of electromechanical systems with limited training samples. First, a random multi-grained scanning module is designed to improve feature information learning. Subsequently, a feature fusion cascade forest module is constructed to improve the representativeness of features in multi-grained scanning and cascade forest delivery while ensuring data diversity. Finally, a decision tree self-growth strategy is combined to refine the classification capability of the high-level forest. To evaluate the effectiveness of our proposed method, we applied it to experimental data related to motor system and gearbox faults. Our results demonstrate significant improvements over existing methods: With just 20 samples per class, our method achieved an average accuracy of 84.41% for motor System Diagnosis. Similarly, for the gearbox system, we attained an impressive accuracy of up to 92.72% with the same limited dataset. These outcomes underscore the superior feature representation and fault classification capabilities of our approach compared to both benchmark deep forest models and mainstream deep learning methods when confronted with small training datasets.

Deep transfer learning for groundwater flow in heterogeneous aquifers using a simple analytical model

Deep learning models exhibit good interpolating ability, but their performance is often hindered by the scarcity of data in groundwater problems. Analytical models (solutions) provide a first-order physical principle for groundwater flow, but they are only applicable under specific conditions, such as when the aquifer is homogeneous. This study introduces a novel framework for deep transfer learning that integrates the strengths of both methods and overcomes their limitations. Specifically, we propose a deep learning model guided by a simple analytical model to predict groundwater flow in heterogeneous aquifers. It differs from previous deep learning model by incorporating the knowledge from the simple analytical model and utilizing transfer learning technique to improve the prediction in relatively complicated problems where the analytical model is not applicable. The model is tested against the traditional deep learning model Deep Back Propagation Neural Network (DBPNN) in the scenarios with unknown hydraulic conductivity fields. The results show that the proposed model significantly improve the accuracy of hydraulic head predictions by fusing analytical knowledge with neural networks. The hydraulic conductivity mainly affects the parameters of the shallow layers in the neural network, which enables the use of transfer learning techniques in more complex problems. In all test scenarios, the prediction errors of the proposed model are much smaller than those of the DBPNN. Additionally, the proposed model performs satisfactorily even with limited training data.

SpinalNet: Deep Neural Network With Gradual Input

Deep neural networks (DNNs) have achieved the state of the art performance in numerous fields. However, DNNs need high computation times, and people always expect better performance in a lower computation. Therefore, we study the human somatosensory system and design a neural network (SpinalNet) to achieve higher accuracy with fewer computations. Hidden layers in traditional NNs receive inputs in the previous layer, apply activation function, and then transfer the outcomes to the next layer. In the proposed SpinalNet, each layer is split into three splits: 1) input split, 2) intermediate split, and 3) output split. Input split of each layer receives a part of the inputs. The intermediate split of each layer receives outputs of the intermediate split of the previous layer and outputs of the input split of the current layer. The number of incoming weights becomes significantly lower than traditional DNNs. The SpinalNet can also be used as the fully connected or classification layer of DNN and supports both traditional learning and transfer learning. We observe significant error reductions with lower computational costs in most of the DNNs. Traditional learning on the VGG-5 network with SpinalNet classification layers provided the state-of-the-art (SOTA) performance on QMNIST, Kuzushiji-MNIST, and EMNIST (Letters, Digits, and Balanced) datasets. Traditional learning with ImageNet pre-trained initial weights and SpinalNet classification layers provided the SOTA performance on STL-10, Fruits 360, Bird225, and Caltech-101 datasets. The scripts of the proposed SpinalNet training are available at the following link: <uri>https://github.com/dipuk0506/SpinalNet</uri> <i>Impact Statement</i>&#x2014;Research in deep neural networks (DNNs) has gained significant attention from industries and academia due to their eye-catching performance. DNNs have enabled machines to perform myriad tasks with high accuracy that once only humans could do. Several researchers have recently proposed different types of NNs and have achieved high accuracy. The recent success of biologically inspired convolutional neural networks and the miraculous spinal architecture of humans has motivated us to develop a neural network with gradual inputs. We have achieved superior performance in several datasets. After the first online appearance of the initial version of this paper, several researchers have applied the proposed neural network in several new datasets and reported promising results. We may observe numerous novel applications of SpinalNet in upcoming years.

A Homomorphic Encryption Framework for Privacy-Preserving Spiking Neural Networks

Machine learning (ML) is widely used today, especially through deep neural networks (DNNs); however, increasing computational load and resource requirements have led to cloud-based solutions. To address this problem, a new generation of networks has emerged called spiking neural networks (SNNs), which mimic the behavior of the human brain to improve efficiency and reduce energy consumption. These networks often process large amounts of sensitive information, such as confidential data, and thus privacy issues arise. Homomorphic encryption (HE) offers a solution, allowing calculations to be performed on encrypted data without decrypting them. This research compares traditional DNNs and SNNs using the Brakerski/Fan-Vercauteren (BFV) encryption scheme. The LeNet-5 and AlexNet models, widely-used convolutional architectures, are used for both DNN and SNN models based on their respective architectures, and the networks are trained and compared using the FashionMNIST dataset. The results show that SNNs using HE achieve up to 40% higher accuracy than DNNs for low values of the plaintext modulus t, although their execution time is longer due to their time-coding nature with multiple time steps.

SECP-Net: SE-Connection Pyramid Network for Segmentation of Organs at Risk with Nasopharyngeal Carcinoma.

Nasopharyngeal carcinoma (NPC) is a kind of malignant tumor. The accurate and automatic segmentation of computed tomography (CT) images of organs at risk (OAR) is clinically significant. In recent years, deep learning models represented by U-Net have been widely applied in medical image segmentation tasks, which can help to reduce doctors' workload. In the OAR segmentation of NPC, the sizes of the OAR are variable, and some of their volumes are small. Traditional deep neural networks underperform in segmentation due to the insufficient use of global and multi-size information. Therefore, a new SE-Connection Pyramid Network (SECP-Net) is proposed. For extracting global and multi-size information, the SECP-Net designs an SE-connection module and a pyramid structure for improving the segmentation performance, especially that of small organs. SECP-Net also uses an auto-context cascaded structure to further refine the segmentation results. Comparative experiments are conducted between SECP-Net and other recent methods on a private dataset with CT images of the head and neck and a public liver dataset. Five-fold cross-validation is used to evaluate the performance based on two metrics; i.e., Dice and Jaccard similarity. The experimental results show that SECP-Net can achieve SOTA performance in these two challenging tasks.

Exploring Trade-Offs in Spiking Neural Networks.

Spiking neural networks (SNNs) have emerged as a promising alternative to traditional deep neural networks for low-power computing. However, the effectiveness of SNNs is not solely determined by their performance but also by their energy consumption, prediction speed, and robustness to noise. The recent method Fast & Deep, along with others, achieves fast and energy-efficient computation by constraining neurons to fire at most once. Known as time-to-first-spike (TTFS), this constraint, however, restricts the capabilities of SNNs in many aspects. In this work, we explore the relationships of performance, energy consumption, speed, and stability when using this constraint. More precisely, we highlight the existence of trade-offs where performance and robustness are gained at the cost of sparsity and prediction latency. To improve these trade-offs, we propose a relaxed version of Fast & Deep that allows for multiple spikes per neuron. Our experiments show that relaxing the spike constraint provides higher performance while also benefiting from faster convergence, similar sparsity, comparable prediction latency, and better robustness to noise compared to TTFS SNNs. By highlighting the limitations of TTFS and demonstrating the advantages of unconstrained SNNs, we provide valuable insight for the development of effective learning strategies for neuromorphic computing.

Improved Transformer-Based Target Matching of Terahertz Broadband Reflective Metamaterials With Monolayer Graphene

On-demand metamaterial designs aided by artificial intelligence have hitherto received tremendous attention recently. However, the traditional deep neural networks (DNNs) still show the limited generalization ability in the inverse design of tunable graphene-based terahertz (THz) metamaterial. In this article, we propose two kinds of DNNs based on the self-attention mechanism to implement the inverse design of tunable broadband reflectors working in the THz band. Moreover, the proposed networks have been improved, so that they could adapt to different types of input vector or matrix in terms of different kinds of on-demand design requirements. Besides, adaptive batch normalization (BN) layers are introduced in our improved networks to accelerate the converging speed with low computational consumption. It could be shown in experiments that the proposed networks exhibit higher accuracy and faster convergence speed than the traditional neural networks, such as multilayer perceptron (MLP) and convolutional neural network (CNN). Finally, this work may provide a key guide for developing THz metamaterials with 2-D materials employing DNNs based on self-attention mechanism.

X-Former: In-Memory Acceleration of Transformers

Transformers have achieved great success in a wide variety of natural language processing (NLP) tasks due to the self-attention mechanism, which assigns an importance score for every word relative to other words in a sequence. However, these models are very large, often reaching hundreds of billions of parameters, and therefore require a large number of dynamic random access memory (DRAM) accesses. Hence, traditional deep neural network (DNN) accelerators such as graphical processing units (GPUs) and tensor processing units (TPUs) face limitations in processing Transformers efficiently. In-memory accelerators based on nonvolatile memory (NVM) promise to be an effective solution to this challenge, since they provide high storage density while performing massively parallel matrix–vector multiplications (MVMs) within memory arrays. However, attention score computations, which are frequently used in Transformers unlike convolutional neural networks (CNNs) and recurrent neural network (RNNs), require MVMs where both the operands change dynamically for each input. As a result, conventional NVM-based accelerators incur high write latency and write energy when used for Transformers and further suffer from the low endurance of most NVM technologies. To address these challenges, we present, a hybrid in-memory hardware accelerator that consists of both NVM and CMOS processing elements to execute transformer workloads efficiently. To improve the hardware utilization of, we also propose a sequence blocking dataflow, which overlaps the computations of the two processing elements and reduces execution time. Across several benchmarks, we show that achieves up to 69.8 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> and 13 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> improvements in latency and energy over a NVIDIA GeForce GTX 1060 GPU and up to 24.1 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> and 7.95 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> improvements in latency and energy over a state-of-the-art in-memory NVM accelerator.

MINT: Detecting Fraudulent Behaviors from Time-Series Relational Data

The e-commerce platforms, such as Shopee, have accumulated a huge volume of time-series relational data, which contains useful information on differentiating fraud users from benign users. Existing fraud behavior detection approaches typically model the time-series data with a vanilla Recurrent Neural Network (RNN) or combine the whole sequence as a single intention without considering the temporal behavioral patterns, row-level interactions, and different view intentions. In this paper, we present MINT, a M ultiview row- IN teractive T ime-aware framework to detect fraudulent behaviors from time-series structured data. The key idea of MINT is to build a time-aware behavior graph for each user's time-series relational data with each row represented as an action node. We utilize the user's temporal information to construct three different graph convolutional matrices for hierarchically learning the user's intentions from different views, that is, short-term, medium-term, and long-term intentions. To capture more meaningful row-level interactions and alleviate the over-smoothing issue in a vanilla time-aware behavior graph, we propose a novel gated neighbor interaction mechanism to calibrate the aggregated information by each action node. Since the receptive fields of the three graph convolutional layers are designed to grow nearly exponentially, our MINT requires many fewer layers than traditional deep graph neural networks (GNNs) to capture multi-hop neighboring information, and avoids recurrent feedforward propagation, thus leading to higher training efficiency and scalability. Our extensive experiments on the large-scale e-commerce datasets from Shopee with up to 4.6 billion records and a public dataset from Amazon show that MINT achieves superior performance over 10 state-of-the-art models and provides better interpretability and scalability.

Engine gas path component fault diagnosis based on a sparse deep stacking network

Accurate engine gas path component fault diagnosis methods are key to ensuring the reliability and safety of engine operations. At present, the effectiveness of the data-driven gas path component fault diagnosis methods has been widely verified in engineering applications. The deep stack neural network (DSN), as a common deep learning neural network, has been gaining more attention in gas path fault diagnosis studies. However, various gas path component faults with strong coupling effects could occur simultaneously, resulting the DSN method less effective for engine gas path fault diagnosis. In order to improve the prediction performance of the DSN handling multiple gas path component fault diagnosis, a sparse regularization and representation method was proposed. The sparse regularization term is used to expand the traditional deep stacking neural network in the sparse representation, and the predicted output tag is close to the target output tag through this term. The diagnosis performance of six different neural network methods were compared by various engine gas path component fault diagnosis types. The results show that the proposed sparse regularization method significantly improves the prediction performance of the DSN, with an accuracy rate 99.9% under various gas path component fault conditions, which is higher than other methods. The proposed engine gas path component fault diagnosis method can handle multiple coupling gas path faults, and help engine operators to develop maintenance plans for the purpose of engine health management.

Minimalist DCT-based Depthwise Separable Convolutional Neural Network Approach for Tangut Script

The Tangut script, a lesser-explored dead script comprising numerous characters, has received limited attention in deep learning research, particularly in the field of optical character recognition (OCR). Existing OCR studies primarily focus on widely-used characters like Chinese characters and employ deep convolutional neural networks (CNNs) or combinations with recurrent neural networks (RNNs) to enhance accuracy in character recognition. In contrast, this study takes a counterintuitive approach to develop an OCR model specifically for the Tangut script. We utilize shorter layers with slimmer filters using a depthwise separable convolutional neural network (DSCNN) architecture. Furthermore, we preprocess the dataset using a frequency-based transformation, namely the Discrete Cosine Transform (DCT). The results demonstrate successful training of the model, showcasing faster convergence and higher accuracy compared to traditional deep neural networks commonly used in OCR applications.