Syukron Abu Ishaq Alfarozi

Background: Most lip-reading studies primarily utilize image-based representations of lip movements, offering extensive visual data while imposing significant computational burdens. Lip landmark-based representations are still not well understood, even though they could be a better way to describe lip dynamics in a smaller and more efficient way. This limitation is even more apparent in Bahasa lip-reading research, where there are few studies and computationally efficient solutions remain essential. Objective: This study examines the shortcomings of image-based lip-reading methods by leveraging lip landmarks as a concise and computationally efficient input representation. The proposed method is tested on the IndoLR open dataset, which contains video data of lip-reading in Bahasa. Methods: In this study, video sequences were transformed into coordinate-based landmark data to minimize computational demands while preserving critical information regarding lip dynamics. An attention-based BiLSTM model was trained to group10 word classes and 4 phrase classes using this dataset. Results: The model achieved accuracies of 93.03% for word classification and 95.18% for phrase classification. The approach also maintained high efficiency, with average inference times of 0.000530 and 0.011366 s per sample and computational costs of only 0.01 and 0.14 GFLOPs, respectively. Conclusion: These results show how well lip landmarks can be combined with a lightweight deep learning model with very few resources. This study makes a significant contribution to research on lip-reading in Bahasa and lays the groundwork for future studies that will use larger and more diverse datasets. Keywords: Attention, BiLSTM, Bahasa, Landmark, Lip-Reading

The classification of cognitive load is crucial to evaluate mental effort in various tasks. Compared to physiological measures such as electroencephalography (EEG), electrocardiography (ECG), or galvanic skin response (GSR), eye tracking offers a less intrusive method for the classification of cognitive load. However, previous studies that used traditional machine learning have faced limitations in the accuracy of multiclass classification and relied on proprietary feature extraction methods. Hence, it hinders reproducibility with noncommercial software. To address these challenges, we propose a novel approach by implementing three time series deep learning models—Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Temporal Convolutional Network (TCN)—to enhance cognitive load classification in multiclass scenarios. By introducing various temporal windows for novel features and optimizing hyperparameters during training, we improved the accuracy of the cognitive load classification. We benchmarked these deep learning models against traditional machine learning techniques on the COLET dataset. The experimental results revealed that deep learning models significantly outperformed conventional machine learning methods, with BiLSTM achieving the highest accuracies of 0.8780 and 0.8836 for the classification of cognitive load and activity tasks in multiclass scenarios, respectively. This study highlights the potential of deep learning to revolutionize cognitive workload assessments. Using eye movement indices, our method facilitates accurate, scalable, and cost-effective solutions suitable for low-cost eye tracking devices, broadening accessibility for applications in education, healthcare, and beyond.

Time series classification (TSC) is essential in various application domains for understanding system dynamics. The adoption of deep learning has advanced TSC, but model performance is sensitive to hyperparameter configurations. Manual tuning of high-dimensional hyperparameters can be labor-intensive, leading to a preference for automatic hyperparameter optimization (HPO) methods. To the best of our knowledge, survey paper covering various studies on automatic hyperparameters optimization (HPO) of deep learning for TSC is scarce and even none. To address this gap, we present a systematic literature review to assist researchers in addressing the HPO problem for deep learning in TSC. It analyzes studies published between 2018 and June 2024. This review examines HPO methods, hyperparameters, and tools utilized in this context, based on a total of 77 primary studies sourced from academic databases. The findings indicate that Metaheuristic algorithms and Bayesian Optimization are commonly employed approaches, with a focus on hyperparameters related to deep learning architecture. This review provides insights that can inform the design and implementation of HPO strategies for deep learning models in time series analysis.

Counterfeit products have become a pervasive problem in the global marketplace, necessitating effective strategies to protect both consumers and brands. This study examines the role of cybersecurity in addressing counterfeiting issues, specifically focusing on a multi-level grayscale watermark-based authentication system. The system comprises a generator responsible for creating a secure 2D code, and an authenticator designed to extract watermark information and verify product authenticity. To authenticate the secure 2D code, we propose various features, including the analysis of the spatial domain, frequency domain, and grayscale watermark distribution. Furthermore, we emphasize the importance of selecting appropriate interpolation methods to enhance counterfeit detection. Our proposed approach demonstrates remarkable performance, achieving precision, recall, and specificities surpassing 84.8%, 83.33%, and 84.5%, respectively, across different datasets.

Learning management systems (LMSs) have been used massively due to the growing utilization of distance learning. This advancement has led to increased educational data that can be analyzed to improve the quality of the learning process. Learning analytics (LA) is one of the most important methods that can be used to analyze student performance. In this paper, we proposed an LA method based on deep learning, i.e., transformer encoder, to sequentially predict the student’s final performance based on log activities provided by an LMS. The objective is to predict at-risk students of failing so that they can be mitigated as soon as possible. The proposed model was evaluated on the Open University LA Dataset (OULAD) for daily or weekly prediction. The results show that the model could predict at the early stage with an accuracy of 83.17% on withdrawn versus pass-distinction classes. Meanwhile, for other tasks, i.e., withdrawn-fail versus pass-distinction and fail versus pass-distinction tasks, the accuracy was at least 76% at the early stage. The proposed model was compared to the LSTM model. We found that the transformer encoder performed better than the LSTM, with the average difference values from 1% to 3% in terms of accuracy and from 3% to 7% in terms of F1-score for all tasks, based on the statistical testing. Furthermore, the ablation study using positional encoding, different feature aggregation methods, and weighted loss function for the imbalanced class problem was done. In OULAD, we found that model without positional encoding was better in all cases. In contrast, weekly feature aggregation and the use of the weighted loss function performed better only for some cases.

Counterfeiting has been one of the crimes of the 21st century. One of the methods to overcome product counterfeiting is a copy detection pattern (CDP) stamped on the product. CDP is a copy-sensitive pattern that leads to quality degradation of the pattern after the print and scan process. The amount of information loss is used to distinguish between original and fake CDPs. This paper proposed a CDP estimation model based on the convolutional neural network (CNN), namely, CDP-CNN. The CDP-CNN addresses the spatial dependency of the image patch. Thus, it should be better than the state-of-the-art model that uses a multi-layer perceptron (MLP) architecture. The proposed model had an estimation bit error rate (BER) of 9.91% on the batch estimation method. The error rate was 9% lower than the previous method that used an autoencoder MLP model. The proposed model also had a lower number of parameters compared to the previous method. The effect of preprocessing, namely the use of an unsharp mask, was tested using a statistical testing method. The effect of preprocessing had no significant difference except in the batch estimation scheme where the unsharp mask filter reduced the error rate by at least 0.5%. In addition, the proposed model was also used for the authentication method. The authentication using the estimation model had a good separation distribution to distinguish the fake and original CDPs. Thus, the CDP can still be used as the authentication method with reliable performance. It helps anti-counterfeiting on product distribution and reduces negative impacts on various sectors of the economy.

A non-iterative learning algorithm for artificial neural networks is an alternative to optimize the neural network parameters with extremely fast convergence time. Extreme learning machine (ELM) is one of the fastest learning algorithms based on a non-iterative method for a single hidden layer feedforward neural network (SLFN) model. ELM uses a randomization technique that requires a large number of hidden nodes to achieve the high accuracy. This leads to a large and complex model, which is slow at the inference time. Previously, we reported analytical incremental learning (AIL) algorithm, which is a compact model and a non-iterative deterministic learning algorithm, to be used as an alternative. However, AIL cannot grow its set of hidden nodes, due to the node saturation problem. Here, we describe a local sigmoid method (LSM) that is also a sufficiently compact model and a non-iterative deterministic learning algorithm to overcome both the ELM randomization and AIL node saturation problems. The LSM algorithm is based on “divide and conquer” method that divides the dataset into several subsets which are easier to optimize separately. Each subset can be associated with a local segment represented as a hidden node that preserves local information of the subset. This technique helps us to understand the function of each hidden node of the network built. Moreover, we can use such a technique to explain the function of hidden nodes learned by backpropagation, the iterative algorithm. Based on our experimental results, LSM is more accurate than other non-iterative learning algorithms and one of the most compact models.