A Security Concern About Deep Learning Models

In computer vision, significant advances have been made in recent years on object recognition and detection with the rapid development of deep learning, especially deep convolutional neural networks (CNN). The majority of deep learning methods for object detection have been developed for large objects and their performances on small-object detection are not very good. This paper contributes to research in low-resolution small-object detection by evaluating the performances of leading deep learning methods for object detection using a common dataset, which is a new dataset for bird detection, called Little Birds in Aerial Imagery (LBAI), created from real-life aerial imagery data. LBAI contains birds with sizes ranging from 10px to 40px. In our experiments, some of the best deep learning architectures were implemented and applied to LBAI, which include object detection techniques such as YOLOv2, SSH, and Tiny Face, in addition to small instance segmentation techniques including U-Net and Mask R-CNN. Among the object detection methods, experimental results demonstrated that SSH performed the best for easy cases, whereas Tiny Face performed the best for hard cases, i.e. where a cluttered background makes detecting birds difficult. Among small instance segmentation methods, experimental results revealed U-Net achieved slightly better performance than Mask R-CNN.

Steering angle prediction plays a crucial role in the control of Autonomous Vehicles (AVs) and has garnered the interest of researchers, manufacturers, and insurance companies within the automotive industry. Various Deep Learning (DL) architectures have been utilized to forecast the steering angle of AVs across different scenarios. An examination focusing on steering angle prediction through deep learning algorithms can assist seasoned researchers in pinpointing areas that necessitate further advancement. Additionally, newcomers in the field can utilize this study as a foundation. This paper offers an extensive analysis of the recent progress achieved in DL architectures concerning the steering angle prediction of AVs. A comprehensive taxonomy outlining the application of DL in steering angle prediction of AVs has been developed. The survey provides a succinct synthesis, summary, and analysis of research findings. It is evident that Convolutional Neural Network (CNN) is the preferred choice among researchers for predicting the steering angle of autonomous vehicles, compared to other DL architectures. Furthermore, existing research challenges have been identified. The primary obstacle encountered in DL-based steering angle prediction of AVs is the scarcity of real-world datasets, leading researchers to heavily rely on data generated from simulated environments. Lastly, potential alternative approaches to address the identified research challenges have been suggested, indicating promising avenues for future research endeavours.

This study proposes a compact deep learning (DL) architecture and a highly parallelized computing hardware platform to reconstruct the blood flow index (BFi) in diffuse correlation spectroscopy (DCS). We leveraged a rigorous analytical model to generate autocorrelation functions (ACFs) to train the DL network. We assessed the accuracy of the proposed DL using simulated and milk phantom data. Compared to convolutional neural networks (CNN), our lightweight DL architecture achieves 66.7% and 18.5% improvement in MSE for BFi and the coherence factor β, using synthetic data evaluation. The accuracy of rBFi over different algorithms was also investigated. We further simplified the DL computing primitives using subtraction for feature extraction, considering further hardware implementation. We extensively explored computing parallelism and fixed-point quantization within the DL architecture. With the DL model's compact size, we employed unrolling and pipelining optimizations for computation-intensive for-loops in the DL model while storing all learned parameters in on-chip BRAMs. We also achieved pixel-wise parallelism, enabling simultaneous, real-time processing of 10 and 15 autocorrelation functions on Zynq-7000 and Zynq-UltraScale+ field programmable gate array (FPGA), respectively. Unlike existing FPGA accelerators that produce BFi and the β from autocorrelation functions on standalone hardware, our approach is an encapsulated, end-to-end on-chip conversion process from intensity photon data to the temporal intensity ACF and subsequently reconstructing BFi and β. This hardware platform achieves an on-chip solution to replace post-processing and miniaturize modern DCS systems that use single-photon cameras. We also comprehensively compared the computational efficiency of our FPGA accelerator to CPU and GPU solutions.

Deep learning based classification of breast tumors with shear-wave elastography

The aims of this work are (1) to explore deep learning (DL) architectures, spectroscopic input types, and learning designs toward optimal quantification in MR spectroscopy of simulated pathological spectra; and (2) to demonstrate accuracy and precision of DL predictions in view of inherent bias toward the training distribution. Simulated 1D spectra and 2D spectrograms that mimic an extensive range of pathological in vivo conditions are used to train and test 24 different DL architectures. Active learning through altered training and testing data distributions is probed to optimize quantification performance. Ensembles of networks are explored to improve DL robustness and reduce the variance of estimates. A set of scores compares performances of DL predictions and traditional model fitting (MF). Ensembles of heterogeneous networks that combine 1D frequency-domain and 2D time-frequency domain spectrograms as input perform best. Dataset augmentation with active learning can improve performance, but gains are limited. MF is more accurate, although DL appears to be more precise at low SNR. However, this overall improved precision originates from a strong bias for cases with high uncertainty toward the dataset the network has been trained with, tending toward its average value. MF mostly performs better compared to the faster DL approach. Potential intrinsic biases on training sets are dangerous in a clinical context that requires the algorithm to be unbiased to outliers (i.e., pathological data). Active learning and ensemble of networks are good strategies to improve prediction performances. However, data quality (sufficient SNR) has proven as a bottleneck for adequate unbiased performance-like in the case of MF.

Deteriorated Inverted-T patching can lead to uneven settlement, dip, or reflective transverse cracking on the asphalt overlay. This paper demonstrates a hybrid method to automatically detect deteriorated Inverted-T patching for an efficient maintenance schedule. First, hundreds of 2D/3D pavement images with deteriorated Inverted-T patching were manually identified and labelled from more than 400 miles of field data in Indiana. All data were collected through a high-speed 3D laser imaging system. Afterward, three deep learning architectures, including the Single Shot Detector network (SSD300), an advanced Region-based Convolutional Neural Network (Mask R-CNN), and a fast and precise convolutional network (U-Net), were applied to develop artificial intelligence models to identify deteriorated Inverted-T patching from 3D images. The results indicate that the Mask R-CNN model can achieve good detection accuracy only on the prepared testing images. Further, a hybrid deep learning model was developed to combine International Roughness Index (IRI) values and the corresponding 3D images to detect deteriorated Inverted-T patching. The hybrid method was promising and significantly improved the efficiency of locating deteriorated Inverted-T patching from network screening.

AI-based emotion recognition has emerged as a critical component of affect-aware educational technologies, particularly in online, large-scale, and technology-mediated learning environments where direct observation of learners’ emotional states is limited. By leveraging facial expressions, speech, physiological signals, and interaction data, emotion-aware systems aim to support adaptive instruction, learner engagement, and instructional decision-making. Despite rapid technical progress, the integration of emotion recognition into authentic educational practice remains constrained by dataset limitations, cultural bias, weak pedagogical alignment, and unresolved ethical concerns. This article presents a critical and theory-informed review of recent advances in AI-based emotion recognition in education, with particular emphasis on deep learning architectures and multimodal fusion approaches, while highlighting persistent challenges and future research directions.Purpose:This study aims to critically examine the current state of AI-based emotion recognition in educational contexts by analyzing dominant modeling paradigms, sensing modalities, application domains, and ethical implications, in order to identify key gaps between technical development and pedagogical practice.Methods/Study design/approach:The study adopts a narrative literature review approach grounded in educational and affective computing theories. Representative and influential studies were analyzed across facial, speech, physiological, and multimodal emotion recognition systems, with emphasis on deep learning, attention-based, and transformer architectures. The review synthesizes findings related to system performance, educational applications, dataset characteristics, and ethical considerations.Result/Findings:The findings indicate that multimodal emotion recognition systems generally outperform unimodal approaches in detecting learning-relevant affective states such as engagement, frustration, and confusion. Advances in deep learning, particularly attention-based and transformer models, improve temporal modeling and robustness. However, most systems rely on non-educational datasets, demonstrate limited cultural generalizability, and prioritize recognition accuracy over pedagogical interpretability and long-term educational impact. Ethical challenges related to privacy, consent, bias, and emotional surveillance remain insufficiently addressed.Novelty/Originality/Value:This study provides an integrative and pedagogically grounded synthesis that bridges affective computing research with educational theory and governance perspectives. By reframing AI-based emotion recognition as a decision-support mechanism rather than an automated control system, the article offers a conceptual foundation for responsible, culturally inclusive, and ethically aligned deployment of emotion-aware AI in education, serving as a reference for future interdisciplinary research and policy development.

Detection of smoke from videos captured by surveillance cameras in outdoor environments is one of the useful outcome of Internet of Things (IoT) applications. The potential benefit increases when deep learning (DL) architectures are involved. However, an inherent difficulty is to detect smoke while natural events like fog exists. The effectiveness of color spaces in detection performance has not yet fully evaluated in those architectures. Moreover, the energy and memory requirements of DL architectures may not be applicable for handling IoT implementation demands. Therefore, in this work, a DL architecture with a suitable color space model, applicable for IoT implementations is proposed to detect smoke from videos in foggy environment. By collecting several videos including smoke samples, the performance comparison of popular and the state-of-the-art DL architectures denoted the outperforming result according to both accuracy and memory usage.

Deep learning (also known as deep structured learning, hierarchical learning, or deep machine learning) is a branch of machine learning based on a set of algorithms that attempt to model high-level abstractions in data by using a deep graph with multiple processing layers, composed of multiple linear and nonlinear transformations. Deep learning has been characterized as a class of machine learning algorithms with the following characteristics [257]: • They use a cascade of many layers of nonlinear processing units for feature extraction and transformation. Each successive layer uses the output from the previous layer as input. The algorithms may be supervised or unsupervised and applications include pattern analysis (unsupervised) and classification (supervised). • They are based on the (unsupervised) learning of multiple levels of features or representations of the data. Higher-level features are derived from lower-level features to form a hierarchical representation. • They are part of the broader machine learning field of learning representations of data. • They learn multiple levels of representations that correspond to different levels of abstraction; the levels form a hierarchy of concepts. These definitions have in common: multiple layers of nonlinear processing units and the supervised or unsupervised learning of feature representations in each layer, with the layers forming a hierarchy from low-level to high-level features. Various deep learning architectures such as deep neural networks, convolutional deep neural networks, deep belief networks (DBN), and recurrent neural networks have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition, and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. In this chapter we start in Section 7.1 with an introduction, giving a brief history of this field, the relevant literature, and its applications. Then we study some basic concepts of deep learning such as convolutional neural networks, recurrent neural networks, backpropagation algorithm, restricted Boltzmann machines, and deep learning networks in Section 7.2. Then we illustrate three examples for Apache Spark implementation for mobile big data (MBD), user moving pattern extraction, and combination with nonparametric Bayesian learning, respectively, in Sections 7.3 through 7.5. Finally, we have summary in Section 7.6.

Deep learning and neural networks have revolutionized artificial intelligence, transforming industries and daily life with applications ranging from voice assistants to medical diagnostics. Despite their ubiquity, these technologies remain enigmatic to many enthusiasts and practitioners. This article demystifies the fundamental concepts underlying neural networks, exploring their biological inspiration, architectural components, and learning mechanisms. Various deep learning architectures are examined, including convolutional neural networks, recurrent neural networks, transformers, and generative adversarial networks, elucidating their distinctive features and applications. The discussion extends to practical considerations in training neural networks, highlighting data requirements, optimization challenges, and regularization techniques. By exploring applications across computer vision, natural language processing, speech recognition, and recommendation systems, the transformative impact of these technologies is illustrated. The article concludes by addressing limitations and ethical considerations, emphasizing the importance of interpretability, fairness, resource efficiency, and environmental sustainability as the field continues to advance.

Natural language processing (NLP) techniques have evolved substantially in recent years, with transformer-based architectures emerging as powerful tools for textual analysis across various fields. This paper investigates the underexplored area of Arabic legal text classification through the implementation of transformer-based language models integrated with deep learning (DL) architectures. Even though transformer models are widely used in legal NLP for high-resource languages, their application to Arabic legal texts remains limited due to linguistic complexities and dataset scarcity. We introduce a classification model for Arabic legal text and present a new dataset of over 27,000 Moroccan legal cases with seven categories. Our methodology implements a segment-based approach to overcome the token limitations of transformer models, processing document segments individually before aggregating embeddings to create complete document representations. Four deep learning architectures are then used to classify those documents. Experimental evaluations demonstrate that MARBERTv2 combined with Gated Recurrent Unit (GRU) architecture achieves superior performance with 95.01% accuracy and Matthews Correlation Coefficient (MCC) of 0.9090.

This study compared popular Deep Learning (DL) architectures to classify machining surface roughness using sound and force data. The DL architectures considered in this study include Multi-Layer Perceptron (MLP), Convolution Neural Network (CNN), Long Short-Term Memory (LSTM), and transformer. The classification was performed on the sound and force data generated during machining aluminum sheets for different levels of spindle speed, feed rate, depth of cut, and end-mill diameter, and it was trained on 30 s machining data (10–40 s) of the machining experiments. Since a raw audio waveform is seldom used in DL models, Mel-Spectrogram and Mel Frequency Cepstral Coefficients (MFCCs) audio feature extraction techniques were used in the DL models. The results of DL models were compared for the training–validation accuracy, training epochs, and training parameters of each model. Although the roughness classification by all the DL models was satisfactory (except for CNN with Mel-Spectrogram), the transformer-based modes had the highest training (>96%) and validation accuracies (≈90%). The CNN model with Mel-Spectrogram exhibited the worst training and inference accuracy, which is influenced by limited training data. Confusion matrices were plotted to observe the classification accuracy visually. The confusion matrices showed that the transformer model trained on Mel-Spectrogram and the transformer model trained on MFCCs correctly predicted 366 (or 91.5%) and 371 (or 92.7%) out of 400 test samples. This study also highlights the suitability and superiority of the transformer model for time series sound and force data and over other DL models.

Question classification is a primary essential study for automatic question answering implementations. Linguistic features take a significant role to develop an accurate question classifier. Recently, deep learning systems have achieved remarkable success in various text-mining problems such as sentiment analysis, document classification, spam filtering, document summarization, and web mining. In this study, we explain our study on investigating some deep learning architectures for a question classification task in a highly inflectional language Turkish that is an agglutinative language where word structure is produced by adding suffixes (morphemes) to root word. As a non-Indo-European language, languages like Turkish have some unique features, which make it challenging for natural language processing. For instance, Turkish has no grammatical gender and noun classes. In this study, user questions in Turkish are used to train and test the deep learning architectures. In addition to this, the details of the deep learning architectures are compared in terms of test and 10-cross fold validation accuracy. We use two major deep learning models in our paper: long short-term memory (LSTM), Convolutional Neural Networks (CNN), and we also implemented the combination of CNN-LSTM, CNN-SVM structures and a number of various those architectures by changing vector sizes and the embedding types. As well as this, we have built word embeddings using the Word2vec method with a CBOW and skip gram models with different vector sizes on a large corpus composed of user questions. Our another investigation is the effect of using different Word2vec pre-trained word embeddings on these deep learning architectures. Experiment results show that the use of different Word2vec models has a significant impact on the accuracy rate on different deep learning models. Additionally, there is no Turkish question dataset labeled and so another contribution in this study is that we introduce new Turkish question dataset which is translated from UIUC English question dataset. By using these techniques, we have reached an accuracy of 94% on the question dataset.

Supervised object detection models are trained to recognize certain objects. These models are classified into two types: single-stage detectors and two-stage detectors. The single-stage detectors just need one pass through the model to anticipate all the bounding boxes, whereas the two-stage detectors require to first estimate the image portions where the object could be located. Due to their speed and simplicity, single-stage anchor-based models are used in many industrial settings. Training such models require bounding boxes that describe the spatial location of an object, which are usually drawn by an expert. However, the question remains, how much area should be considered when drawing the bounding boxes? In this paper, we demonstrate the effects that the size and placement of a rectangular bounding box can have on the performance of the anchor-based models. For this, we first perform experiments on a synthetically generated binary dataset and then on a real-world object detection dataset. Our results show that fixing the size of the bounding boxes can help in improving the performance of the model in the case of single class object detection (approximately 50% improvement in mAP@[.5:.95] for real world dataset). Furthermore, we also demonstrate how freely available tools can be combined for obtaining the best possible semi automated object labeling pipeline.

Recent advancements in deep learning techniques tailored for autonomous driving in Indian road conditions are crucial for revolutionizing transportation systems. Indian roads present unique challenges, including unpredictable traffic patterns, diverse road infrastructures, and challenging weather conditions. Deep learning is pivotal in addressing these challenges, focusing on perception, decision-making, and control. Various architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Deep Reinforcement Learning (DRL), are analyzed for their efficacy in autonomous driving. Sensor fusion techniques enhance perception capabilities, such as lidar, radar, and camera data integration. Additionally, advancements in semantic segmentation and object detection algorithms improve scene understanding and obstacle recognition. Collaborative efforts between academia, industry, and government agencies are essential for accelerating the deployment of autonomous driving technology in Indian road conditions. In conclusion, leveraging cutting-edge deep learning methodologies enables safe and efficient navigation of autonomous vehicles, ushering in a transformative future in transportation systems.