Information Coding and Hardware Architecture of Spiking Neural Networks

Introduction. The article considers the solution of the problem of detecting fires at technological facilities in automatic mode. To solve this problem, an approach was chosen to create a convolutional neural network capable of operating on a real-time video stream. Aims and Purposes. The aim of this work is to create a neural network capable of detecting flames and smoke in the image from CCTV cameras. The purposes of the work include: selection of the optimal architecture of the neural network in accordance with the latest research in this field; speeding up the chosen architecture using quantization and filter thinning techniques. Methods. Different architectures of convolutional neural networks performing the task of detecting objects in an image are considered. Their performance and quality of work are compared. The YOLOv5 architecture, its target function, training methods and ways of speeding up work are considered. Results and discu ssion. The paper shows the results of training a convolutional neural network of the YOLOv5 architecture for the task of flame and smoke detection, as well as how the results change when applying neural network acceleration methods. It was determined that the use of such acceleration methods such as quantization and filter cleaning can significantly increase up the speed of the neural network, while almost no loss in accuracy of operation. Conclusions. As a result of the conducted work, the architecture of the neural network was determined for performing the task of detecting a fire source. Based on the chosen architecture, a neural network was trained to detect flames and smoke in the image. The speed of its work allows to process video stream in real-time without using graphic accelerator.

The design of neural network models involves numerous complexities, including the determination of input vectors, choosing the number of hidden layers and their computational units, and specifying activation functions for the latter. The combinatoric possibilities are daunting, yet experience has yielded informal guidelines that can be useful. Alternatively, current research on genetic algorithms (GA) suggests that they might be of practical use as a formal method of determining ‘good’ architectures for neural networks. In this paper, we use a genetic algorithm to find effective architectures for backpropagation neural networks (BP). We compare the performance of heuristically designed BP networks with that of GA-designed BP networks. Our test domains are sets of problems having compensatory, conjunctive, and mixed-decision structures. The results of our experiment suggest that heuristic methods produce architectures that are simpler and yet perform comparatively well. © 1998 John Wiley & Sons, Ltd.

Deep neural networks (DNNs) are advantageous for representing complex nonlinear relationships. This paper applies DNNs to source localization in a shallow water environment. Two methods are proposed to estimate the range and depth of a broadband source through different neural network architectures. The first adopts the classical two-stage scheme, in which feature extraction and DNN analysis are independent steps. The eigenvectors associated with the modal signal space are extracted as the input feature. Then, the time delay neural network is exploited to model the long term feature representation and constructs the regression model. The second concerns a convolutional neural network-feed-forward neural network (CNN-FNN) architecture, which trains the network directly by taking the raw multi-channel waveforms as input. The CNNs are expected to perform spatial filtering for multi-channel signals, in an operation analogous to time domain filters. The outputs of CNNs are summed as the input to FNN. Several experiments are conducted on the simulated and experimental data to evaluate the performance of the proposed methods. The results demonstrate that DNNs are effective for source localization in complex and varied water environments, especially when there is little precise environmental information.

Multilevel image segmentation with adaptive image context based thresholding

Aim. To study various network architecture options for implementing an electronic differential system in an electric vehicle. Materials and Methods. The study primarily used comparative analysis to identify the most rational neural network (NN) architecture for processing numerical data structured as arrays. Results. The analysis revealed that a deep learning neural network is the most effective choice. For future developments, and after experimental confirmation, a recurrent neural network could also be a viable option. Conclusion. The study confirmed that achieving the desired goal is not feasible using convolutional neural networks, large language models, random vector functional communication networks and radial-basis functional NNs.

• First application of SNN for HSI regression analysis. • Spectra-Spiking Neural Network (S 2 NN) is proposed for wheat protein estimation. • The model built for protein content ranges from 9.5 % to 17.25 % by weight. • The S 2 NN reduces approximately 15 to 108 times energy consumption and uses approximately 5 to 34 times fewer FLOPs compared to ANN, CNN, and Transformer. Hyperspectral imaging (HSI) combined with traditional machine learning techniques has demonstrated an accurate assessment of seed chemical composition. However, their high computational demands and energy consumption make them unsuitable for real-time applications. Spiking neural networks (SNNs) are recognized as an energy-efficient alternative due to their spike-based mechanism. We explore SNN in the field of HSI for the regression analysis to estimate the wheat protein content. We proposed a spectra-spiking neural network (S 2 NN) over a single timestep that uses direct input encoding to analyze spectral data more efficiently. S 2 NN achieved an R² value of 0.9969, RMSE, and RPD of 0.0813 and 17.85, respectively. The S 2 NN was compared with other neural network models (i.e., ANN, CNN, and Transformer) in terms of accuracy, training time, floating-point operations (FLOPs), and energy consumption. We first compared the performance of the S 2 NN with ANN while maintaining a similar network architecture for both models. In this comparison, S 2 NN achieved a 5.08 % higher R² value and was approximately 2.06 times more energy-efficient. Next, we train the ANN, CNN, and Transformer to reach their maximum predictive accuracy, yielding results comparable to those of S 2 NN. In this case, the S 2 NN reduces energy consumption by approximately 15 to 108 times, requires approximately 5 to 34 times fewer FLOPs, and demonstrates significantly shorter training times compared to other neural network architectures. These findings show the potential of S 2 NNs as an efficient alternative to other neural networks, particularly for resource-constrained real-time agricultural applications where high-end computational devices may not be available. The code is available at https://github.com/Apurva-Git-Hub/Spectra-SNN.git .

Nowadays with the deployment of a large and dense heterogeneous networks more sophisticated algorithms for resource scheduling are needed. Implementing hard coded scheduling algorithms without taking into account the very specific dynamic of the traffic generated by the mobile users can lead to a network performance quite far from the optimal. By using novel machine learning (ML) algorithms we can store not only the raw traffic data and its variations but also build the so-called heat maps, reflecting the changes of the traffic over time, space and per user. Using neural network (NN) architectures, trained by the raw data statistics, we can store the network traffic model at minimum data storage without the need of keeping and looking up at the raw data. Using such NN architecture the network state in next time intervals could be predicted and this prediction used for decision making about how the network resources to be scheduled among the active mobile users. To implement adaptive resource scheduling named “AdaptSch” a neural network architecture with two main blocks is proposed. The simulation results show that by incorporating a neural classifier for adapting the resource scheduler we can utilize the advantages and the effectiveness of multiple scheduler algorithms and improve overall throughput and packet delay.

The aim of this paper is to investigate new neural network algorithms, architectures and models which are inspired by the image processing capability of the visual system in the human brain, and subsequently to apply these models to medical imaging(MI) applications. Realizing the neural network architecture on FPGA that supports parallelism and reconfigurability. A sincere attempt is taken up to implement the complex and massive parallel architecture on FPGA optimising speed, area, and power with out compromising on image quality and SNR that is a very important criteria for medical image compression. Multi layered neural network (NN) architecture is proposed for compression of high-resolution image, the architecture is implemented on FPGA as it supports reconfigurability. The architecture considered has N-M-N (64-4-64) multi-layered NN structure, which achieves compression ratio of (CR) 93.75. Compression ratio is reconfigurable with change in M. The architecture is generalized and can achieve compression ratios from 2 to 99, which is reconfigurable. Performance of any neural network architecture for compression depends on training; this architecture considers general back propagation training. The training is performed offline, with known set of image samples consisting most of the properties of any standard image. The Mean Square Error (MSE) computed during every iteration of training is scaled and fed back into the network to update the weight matrix at specific points, this reduces the training time. As the weight matrix occupies more space for storage, the redundancies in the weight matrix are exploited and a storage space is created with minimum memory requirement on FPGA. Compression ratios obtained demonstrate performance superiority of the network as compared with JPEG compression standard. Inserting noise on the compressed sets of data, tests the network performance. The hardware complexity, area requirement, speed is compared and discussed; there is a saving of 22% of space on FPGA, with 40% increase in speed, and reducing the power by 12%. The major advantage of this architecture is its reconfigurability on the architecture size that achieves different compression ratios.

Predicting the next activity of a running execution trace of a business process represents a challenging task in process mining. The problem has been already tackled by using different machine learning approaches. Among them, deep artificial neural networks architectures suited for sequential data, such as recurrent neural networks (RNNs), recently achieved the state of the art results. However, convolutional neural networks (CNNs) architectures can outperform RNNs on tasks for sequence modeling, such as machine translation. In this paper we investigate the use of stacked inception CNN modules for the next-activity prediction problem. The proposed neural network architecture leads to better results when compared to RNNs architectures both in terms of computational efficiency and prediction accuracy on different real-world datasets.

Model quantization is a widely used technique to compress and accelerate deep neural network (DNN) inference.Emergent DNN hardware accelerators begin to support mixed precision (1-8 bits) to further improve the computation efficiency, which raises a great challenge to find the optimal bitwidth for each layer: it requires domain experts to explore the vast design space trading off among accuracy, latency, energy, and model size, which is both timeconsuming and sub-optimal.There are plenty of specialized hardware for neural networks, but little research has been done for specialized neural network optimization for a particular hardware architecture.Conventional quantization algorithm ignores the different hardware architectures and quantizes all the layers in a uniform way.In this paper, we introduce the Hardware-Aware Automated Quantization (HAQ) framework which leverages the reinforcement learning to automatically determine the quantization policy, and we take the hardware accelerator's feedback in the design loop.Rather than relying on proxy signals such as FLOPs and model size, we employ a hardware simulator to generate direct feedback signals (latency and energy) to the RL agent.Compared with conventional methods, our framework is fully automated and can specialize the quantization policy for different neural network architectures and hardware architectures.Our framework effectively reduced the latency by 1.4-1.95×and the energy consumption by 1.9× with negligible loss of accuracy compared with the fixed bitwidth (8 bits) quantization.Our framework reveals that the optimal policies on different hardware architectures (i.e., edge and cloud architectures) under different resource constraints (i.e., latency, energy and model size) are drastically different.We interpreted the implication of different quantization policies, which offer insights for both neural network architecture design and hardware architecture design.* indicates equal contributions.68 69 70 71 72 73 25 44 63 82 101 120 MobileNets (fixed 8-bit quantization) MobileNets (our flexible-bit quantization) Latency (ms) Top-1 Accuracy (%) 1MB 2MB 3MB Model Size:Figure 1: We need mixed precision for different layers.We quantize MobileNets [12] to different number of bits (both weights and activations), and it lies on a better pareto curve (yellow) than fixed bit quantization (blue).The reason is that different layers have different redundancy and have different arithmetic intensity (OPs/byte) on the hardware, which advocates for using mixed precision for different layers.

Exploiting photoplethysmography signals (PPG) for non-invasive blood pressure (BP) measurement is interesting for various reasons. First, PPG can easily be measured using fingerclip sensors. Second, camera based approaches allow to derive remote PPG (rPPG) signals similar to PPG and therefore provide the opportunity for non-invasive measurements of BP. Various methods relying on machine learning techniques have recently been published. Performances are often reported as the mean average error (MAE) on the data which is problematic. This work aims to analyze the PPG- and rPPG based BP prediction error with respect to the underlying data distribution. First, we train established neural network (NN) architectures and derive an appropriate parameterization of input segments drawn from continuous PPG signals. Second, we use this parameterization to train NNs with a larger PPG dataset and carry out a systematic evaluation of the predicted blood pressure. The analysis revealed a strong systematic increase of the prediction error towards less frequent BP values across NN architectures. Moreover, we tested different train/test set split configurations which underpin the importance of a careful subject-aware dataset assignment to prevent overly optimistic results. Third, we use transfer learning to train the NNs for rPPG based BP prediction. The resulting performances are similar to the PPG-only case. Finally, we apply different personalization techniques and retrain our NNs with subject-specific data for both the PPG-only and rPPG case. Whilst the particular technique is less important, personalization reduces the prediction errors significantly.

The study presents the results of the development and testing of deep learning neural network architectures, which demonstrate high accuracy rates in classifying neurophysiological data, in particular hemodynamic brain activation patterns obtained by functional near-infrared spectroscopy, during solving mathematical problems on spatial-numerical associations. The analyzed signal represents a multidimensional time series of oxyhemoglobin and deoxyhemoglobin dynamics. Taking the specificity of the fNIRS signal into account, a comparative analysis of 2 types of neural network architectures was carried out: (1) architectures based on recurrent neural networks: recurrent neural network with long short-term memory, recurrent neural network with long short-term memory with fully connected layers, bidirectional recurrent neural network with long short-term memory, convolutional recurrent neural network with long short-term memory; (2) architectures based on convolutional neural networks with 1D convolutions: convolutional neural network, fully convolutional neural network, residual neural network. Trained long short-term memory recurrent neural network architectures showed worse results in accuracy in comparison with 1D convolutional neural network architectures. Residual neural network (model_Resnet) demonstrated the highest accuracy rates in three experimental conditions more than 88% in detecting age-related differences in brain activation during spatial-numerical association tasks considering the individual characteristics of the respondents’ signal.

Determining the best model for the neural network architecture and how to optimize the architecture with the metaheuristic Protis Approach is a subject of the study. A comprehensive investigation and utilization of metaheuristic methods are necessary. These methods aim to solve problems and adapt from the lifestyle of the amoeba protis. In this study, the proposed method modifies the life cycle of the amoeba, which consists of four phases: prophase, metaphase, anaphase, and telophase. These four phases are modified in the neural network architecture to optimize the appropriate number of hidden layers and produce an efficient architecture model. The results show that the protis approach optimized the neural network architecture, especially in generating hidden layers to improve the neural network model. Distinctive features of the results obtained are that the average range of degenerate neurons in the hidden layer is 0 to 35 neurons in each layer. The standard number of neurons makes it possible to solve the problem of determining the best model on the neural network architecture. The protis algorithm embedded in the protis recurrent neural network for categorical data measurements produces an average RMSE value, representing the difference between actual measurements and predictions, equal to 0.066. Consequently, the developed model surpasses the current classical neural network model in terms of performance. Regarding accuracy, the protis algorithm embedded in the neural network for categorical and time series data achieves an average precision of 0.952 and a recall of 0.950. The protis convolutional neural network achieves an accuracy of 95.9 %. Therefore, from the three tested datasets, the protis convolutional neural network exhibits the highest accuracy value

This paper presents fits for neural network model , and comparative resulting forecasts with those obtained from Box-Jenkins Method.We use time series data of Tigris's monthly flow into Mosul city from 1950-1995.To perform a comparative .forecasting work through the Box-Jenkins and neural network doesn't mean working with two different or competing aspect ; on the contrary choosing a proper architecture of neural net works requires using the skills of statistical modeling .As for application , Box-Jenkins Method has given more appropriate forecasts than those given by feed forward artificial neural network .We used Minitab and SPSS programs in the statistical aspect and Alyuda program in the neural network aspect.

In this paper, an integrated architecture of spiking neural network is used for learning of 2 DOF robotic arm. Brain architecture is consisting of 6 different areas that have different tasks. Two sensors are used to detect the target position and send the signals to sensory neurons. As an integrated architecture, all of the sensory neurons are connected to all motor neurons at the beginning of the process. The neural network is trained to learn a specific task using spike timing-dependent plasticity. Simultaneous sensing of target and movement of robot toward target result in unsupervised learning which is useful for learning of robots in unknown environments. To remove the effects of random inputs, the experiment is repeated six times. Through the learning process, the synaptic weights are changed. The connection between left sensory neurons and Flexor motor neurons also the connection between right sensory neurons and Extensor motor neurons are increased while other connections are weakened. The results show that the spiking neural network is effective in controlling the robot's motion.