Optical Memory and Neural Networks最新文献

This paper improved the back-propagation neural network (BPNN) algorithm for recommending ideological and political courses by a squeeze-and-excitation network (SEnet) and a multi-head attention mechanism. Simulation experiments were conducted to compare the improved algorithm with two other recommendation algorithms, followed by ablation experiments. Moreover, the effectiveness of the recommendation algorithm was tested in actual teaching of ideological and political courses. The results demonstrated that the improved BPNN algorithm outperformed others and the SEnet and the multi-head attention mechanism significantly enhanced the accuracy of recommendations. The algorithm effectively improved students’ performance in ideological and political courses and was satisfied by the majority of students.

Future computing systems will likely require reversible logic gates since, in the best of circumstances; they are known to provide zero power dissipation. Reversible logic gates also have the advantage of minimizing quantum costs and unused outputs. In this work, reversible 2 : 1 multiplexer has been implemented using all-optical microring resonator. Reversible multiplexer is simulated in MATLAB at 260 Gbps. Some performance indicating factors such as “extinction ratio”, “contrast ratio”, “relative eye opening”, etc are analyzed and developed. The chosen optimal parameters can be validated practically.

This paper presents a description of an improved algorithm for computing the weights of a convolutional neural network, as well as the results of comparative experiments on training a convolutional neural network on a handwritten digit dataset MNIST using both precomputed and randomly initialized weights. The experimental results demonstrate the advantages of preliminary analytical computation of neural network weight values compared to random weight initialization. The weights of the convolutional neural network were computed using only 10 digit images, randomly selected from the dataset MNIST. Experiments showed that the time spent on the analytical computation of weight values was negligible. The testing results on the test dataset with the computed but yet untrained neural network showed an accuracy of more than 50% in correctly recognizing the images from the test dataset MNIST. The results of numerous training experiments on the same convolutional neural network, using both computed and random weights, showed that training with precomputed weights yields better results and requires less training time.

Adaptive extrapolation-based algorithms for upscaling multidimensional digital arrays are investigated. For each array element an atomic extrpolator is chosen from a set of computationally simple atomic extrapolators. The choise relies on the local variation ratio in different directions. The adaptation suggests the efficient automatic selection of the local variation ratio limit at which one atomic extrapolator is replaced by another. The computational efficiency of the self-adjustment algorithm is determined by the use of preceeding (neighboring) values of the extrapolator accuracy factor in calculating the local variation ratio limit. Higher accuracy of the extrapolator is due to the use of the downscaled version of the input multidimensional data array for adjusting the extrapolator. The higher efficiency of adaptive extrapolation when scaling multidimensional data arrays has been experimentally proven.

Worldwide, lung disease is a serious health problem, affecting a large percentage of the global population. Accurate diagnosis of lung diseases can be challenging as many of these conditions present with similar symptoms. The goal of this research is to design a robust deep learning-based technique capable of accurately detecting four types of lung diseases, such as COVID-19, bacterial pneumonia, viral pneumonia, and mycoplasma pneumonia, using computed tomography (CT) images. In this study, a publicly available CT image dataset is used for designing the lung disease detection system. First, CT images are preprocessed to enhance their quality, then a GAN (Generative Adversarial Network)-based augmentation technique is used to expand the dataset, doubling its size. A hybrid deep learning model named LungNet, consisting of a Convolutional Neural Network (CNN) module with a dual attention mechanism, a multi-head attention module, and the proposed RPLSTM (revamped potent Long Short-Term Memory) network module, is designed for lung disease detection. The CNN module extracts useful features from CT images. The multi-head attention module helps to focus on the most significant features extracted by the CNN module. Lastly, the proposed RPLSTM module is used to diagnose lung disease effectively. The designed LungNet model achieves an average detection accuracy of 99.30%. The key innovations of this research are: (1) the design of a novel deep learning architecture called LungNet for different lung disease detection, (2) the utilization of channel attention and spatial attention in the CNN module of the proposed model for robust feature extraction, (3) employing a multi-head attention layer in the designed model to enhance its efficacy, (4) proposing an advanced architecture of potent long short-term memory (PLSTM) called RPLSTM and utilizing it for lung disease detection, and (5) utilization of GAN to increase the dataset size and thus solve the dataset scarcity problem.

Efficient traffic signal control plays a critical role in reducing urban congestion and improving transportation efficiency. This paper presents an adaptive traffic signal control approach based on reinforcement learning that employs the Soft Actor-Critic (SAC) algorithm to optimize traffic signal phase durations. By focusing on phase duration optimization rather than discrete phase selection, the proposed method ensures a more predictable and adaptive traffic management system. Unlike traditional methods that select discrete signal phases, our approach continuously adjusts phase duration based on real-time traffic conditions, providing a more flexible and responsive control strategy. The proposed framework uses connected vehicle data, including position, speed, and acceleration, to predict vehicle arrival times at intersections. These predictions, combined with aggregated traffic characteristics such as queue length and waiting time, form the state representation for the reinforcement learning model. The SAC algorithm is then used to determine optimal phase durations. We evaluated the proposed approach using the SUMO traffic simulator in three different urban scenarios: a single intersection, a three-intersection arterial road, and a small road network. Experimental results demonstrate that the proposed method outperforms baseline approaches, including Deep Q-Network and a heuristic-based method, in terms of average travel time, time loss, and waiting time. Specifically, the SAC-based algorithm achieves reductions of up to 1.5% in average travel time and up to 13% in average waiting time across various simulation scenarios compared to the baseline methods. Furthermore, training convergence analysis and visualizations confirm the stability and effectiveness of the learned policy.

In this study, a combination of a gated recurrent unit (GRU) model and a Transformer model was used to predict soybean prices in the agricultural sector. Simulation experiments were conducted. The soybean price prediction algorithm was initially compared with two other algorithms, random forest (RF) and back-propagation neural network (BPNN). Then, ablation experiments were performed on the proposed prediction algorithm. The importance of the feature indicators used in predicting soybean prices was tested. The results indicated that, compared to the RF and BPNN algorithms, the GRU-Transformer model demonstrated a superior performance. Additionally, the results of the ablation experiments revealed that both GRU and Transformer models significantly contributed to the accuracy of soybean price prediction. Moreover, the importance of feature indicators such as soybean imports, soybean exports, soybean oil prices, exchange rates, and soybean meal prices was found to be high.

Emotions reflect the mental state of a person. Individual changes in physiological, physical, mental and behavioral factors reflect different types of emotions. Studies on emotion recognition always have attention among researchers. Signal processing and feature handling techniques are developed for accurate emotions recognition from the biological brain signals. Through electroencephalography (EEG) channels, physiological signals are obtained and the essential features are extracted for analysis. However, the detection or recognition performance of traditional methods provides room for improvement due to poor accuracy or improper feature handling performances. The EEG signals for emotion recognition are predicted from two data sources that are preprocessed through filtering method for reducing artifacts of the EEG signals. Then, the preprocessed signals are given to the feature extraction phase such as Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM). The CNN model is applied for the extraction of statistical related features and the LSTM model is applied to extract non-linear related features. These features are fused at the final stage of the feature extraction phase and ResNet152 model is implemented in this paper for classifying types of emotions in the EEG signals according to the extracted features. The comprehensive analyses are performed through different performance evaluation measures and the proposed model attained better performances of 0.9867 and 0.9646 from accuracy and mathew’s correlation coefficient respectively. From this experimental validation, the proposed model achieved better outcome than other compared existing approaches.

A complicated neurodevelopmental disorder, autism spectrum disorder (ASD) is represented by difficulties with cognition and behavior. Early and accurate diagnosis is crucial for effective intervention. However, existing machine learning methods for ASD detection face limitations, including inefficiencies in EEG signal noise removal, challenges in feature extraction, and difficulties in stage-wise classification. To address these challenges, the SilverHowl-QDecomp Framework is proposed to enhance EEG-based ASD classification through advanced signal processing and feature extraction techniques. The LaplaZ Filter effectively minimizes noise while preserving critical signal components and normalization techniques ensure data consistency. Furthermore, the proposed feature extraction method captures nonlinear and dynamic EEG characteristics, improving classification accuracy by isolating essential features and reducing computational complexity. To enhance ASD stage classification, the SilverHowl Classifier was introduced, implementing the BCIAUT-P300 dataset and leveraging optimized hyperparameters to achieve better discrimination between ASD stages. With an accuracy of 0.985 and a precision of 0.98572, this method performs better than conventional techniques, thereby offering a more reliable and precise classification framework. The proposed method contributes to personalized ASD interventions by enabling more accurate and stage-specific diagnoses.

Large Language Models (LLMs) require a lot of computational resources for inference. That is why the latest advancements in hardware design may offer many possibilities for speeding the LLM up. For example, TPU optimize calculations on data, transformed into the Coordinate sparse tensor format. The SparseCore processing unit that performs the calculations is heavily tailored for the extremely sparse embeddings of Deep Learning Recommendation Models. The other example of the enhanced hardware is Sparse Tensor Cores, that offer support for (n:m) data structure ((n) zeroes out of every subsequent (m) elements), that allows to drastically reduce the calculations by compressing the original matrix into a dense one. Methods like Wanda and SliceGPT prepare LLM weights to harness the power of the latter. However, as the weights are the most crucial assets of any model, it appears to be a good idea to modify the activations instead. This article introduces a novel dynamic sparsification algorithm called KurSparse , which proposes fine-grained n : m sparsity pattern, that affects only a portion of channels. This portion is selected with kurtosis threshold (zeta ). The proposed method shows significant reduction in MAC operations by 3.1x with average quality drop for LLaMA-3.1-8B model less than 2%.