The rapid advancement of deep learning has revolutionized artificial intelligence, enabling breakthroughs in fields ranging from natural language processing to computer vision. However, this progress comes at a significant environmental cost. The training and deployment of large-scale neural networks consume vast amounts of energy, contributing to carbon emissions and exacerbating climate change. This paper argues that the AI research community must prioritize sustainability to mitigate the environmental impact of deep learning. We highlight the growing energy demands of state-of-the-art models, such as GPT and large vision transformers, and discuss their ecological consequences. Furthermore, we propose actionable solutions, including the development of energy-efficient algorithms, the adoption of renewable energy for data centers, and the establishment of industry-wide sustainability standards. By adopting these practices, the AI community can continue to innovate while minimizing its environmental footprint. This paper calls for a collective effort to align the pursuit of technological advancement with the urgent need for environmental responsibility.

This paper proposes a single-objective filter pruning method based on constrained binary Particle Swarm Optimization, integrating L1-based search space reduction and adaptive weighting. The proposed method, AWL-BPSO, ensures efficient convergence while maintaining a diverse set of pruned models, providing multiple trade-offs between accuracy and computational efficiency. Unlike existing methods that rely on fixed pruning rates, our method dynamically explores optimal filter configurations, making it more flexible for real-world deployment on edge devices. Experimental results on popular deep convolutional neural networks demonstrate that our method achieves significant model compression while preserving accuracy, outperforming conventional multi-objective methods in both efficiency and adaptability. This work contributes to advancing scalable, efficient, and hardware-aware neural network pruning for deep learning applications.

Abstract Neural network quantization is a critical method for reducing memory usage and computational complexity in deep learning models, making them more suitable for deployment on resource‐constrained devices. In this article, we propose a method called BBPSO‐Quantizer, which utilizes an enhanced Bare‐Bones Particle Swarm Optimization algorithm, to address the challenging problem of mixed precision quantization of convolutional neural networks (CNNs). Our proposed algorithm leverages a new population initialization, a robust screening process, and a local search strategy to improve the search performance and guide the population towards a feasible region. Additionally, Deb's constraint handling method is incorporated to ensure that the optimized solutions satisfy the functional constraints. The effectiveness of our BBPSO‐Quantizer is evaluated on various state‐of‐the‐art CNN architectures, including VGG, DenseNet, ResNet, and MobileNetV2, using CIFAR‐10, CIFAR‐100, and Tiny ImageNet datasets. Comparative results demonstrate that our method delivers an excellent tradeoff between accuracy and computational efficiency.

This paper investigates the effect of modern lifestyles on physical activity and the rise of time-saving exercise protocols to improve health and performance. Such protocols, Bio Impedance Analysis (BIA), and Electrical Muscle Stimulation (EMS) are explored in detail. BIA is a non-invasive and painless technology that accurately evaluates the amount of water, proteins, minerals, and fat masses in the body. EMS involves applying electrical currents to muscles through electrodes to induce involuntary contractions and is being used for muscle rehabilitation, treating obesity, and improving body shape. The manufacturers of EMS devices claim that 20 minutes of electro-stimulation work is equivalent to 4 hours of traditional sports. The emergence of new technologies such as INTEGRAL-EMS, a combination of portable electro-stimulation with an integrated, wireless, high-performance generator, is making it increasingly accessible for people to achieve their health and fitness goals in a more efficient and convenient manner.

Text detection, recognition, and visualization from natural scenes are considered significant research topics in the field of information and communication technologies. Written text serves as a crucial source of information that humans rely on in their daily lives. However, detecting text poses several challenges, including variations in writing style, color, size, orientation, and complex backgrounds. In this study, we present a comprehensive application named ATDRV, which aims to address these challenges. For text detection and recognition, we propose a real-time multi-oriented text deep fully convolutional network system trained end-to-end. Additionally, we implement a visualization module based on Augmented Reality to display the results more clearly to users. The output of our application is a large and clear 3D object that enables easy reading of the text. Our approach improves the user experience and facilitates the reading of Arabic text regardless of color, orientation, writing style, or complex backgrounds.

This paper presents an automatic vision transformer pruning method that aims to alleviate the difficulty of deploying vision transformer models on resource-constrained devices. The proposed method aims to automatically search for the optimal pruned model by removing irrelevant units while maintaining the original accuracy. Specifically, the model pruning is formulated as an optimization problem using binary particle swarm optimization. To demonstrate its effectiveness, our method was tested on the DeiT Transformer model with CIFAR-10 and CIFAR-100 datasets. Experimental results demonstrate that our method achieves a significant reduction in computational cost with slight performance degradation.

In real-world applications such as emotion recognition from recorded brain activity, data are captured from electrodes over time. These signals constitute a multidimensional time series. In this article, Echo State Network (ESN), a recurrent neural network with great success in time series prediction and classification, is optimized with different neural plasticity rules for classification of emotions based on electroencephalogram (EEG) time series. The developed network could automatically extract valid features from EEG signals. We use the filtered signals as the network input and do not take any feature extraction methods. Evaluated on two well-known benchmarks, the DEAP dataset, and the SEED dataset, the performance of the ESN with intrinsic plasticity greatly outperforms the feature-based methods and shows certain advantages compared with other existing methods. Thus, the proposed network can form a more complete and efficient representation, whilst retaining the advantages such as faster learning speed and more reliable performance.

Deep learning (DL) has been expensively applied in multiple fields like computer vision, speech recognition and natural language processing. The field of Epileptic seizure prediction didn't receive the deserved attention by DL community, even though, deep neural networks can handle the challenging task of onsets prediction whilst achieving the highest rates of sensitivity, despite the complex nature of EEG signals. In the literature, this issue was addressed differently most of the time using handcrafted temporal and spectral features, machine learning techniques and rarely deep learning with extracted features. In this paper, we introduce an LSTM model designed to address the chaotic nature of an EEG signal in order to predict pre-ictal and inter-ictal states. Our model is evaluated on the publicly available CHBMIT database. We achieved an average sensitivity rate of 0.84 using a Raw EEG data segment as input to the LSTM model.

In this paper, deep EEG feature learning method is proposed for emotion recognition. It is well known that EEG signals dramatically vary from person to person, thereby making subject-independent emotion recognition very challenging. To address the above challenge, this work presents a deep echo state network (DeepESN) to learn temporal representation from raw EEG data. DeepESN as an input-driven discrete time non-linear dynamical system allows to process the temporal information at each time step in a deep temporal fashion by means of a hierarchical composition of multiple levels of recurrent neurons. To make the DeepESN robust, we pre-train the reservoir connections with an unsupervised intrinsic plasticity rule to generate activities following a desired Gaussian distribution. Then, we propose a hybrid learning algorithm for training the output weights which benefits from both the ridge regression and the online delta rule. Our leaky DeepESN achieved encouraging results when tested on the well-known affective benchmarks DEAP and DREAMER.

In this paper, we provide an improvement of the circuit implementation of a one-layer recurrent neural network for support vector machine learning in pattern classification and regression. Our goal is to reduce the complexity of this architecture. Numerical example with graphical illustration is given to illuminate our main results.