Global Journal of Computer Science and Technology, D: Neural & Artificial Intelligence, Volume 22 Issue 1

As a next step, we need to create the inference engine. The core engine interface in Barracuda is called worker. IWorker breaks down the model into executable tasks and schedules them on GPU or CPU. The scripts are different depending on whether you use a GPU or a CPU. We can create a Worker from the WorkerFactory. We must specify a backend and a loaded model. // GPU WorkerFactory.CreateWorker(WorkerFactory.Type.ComputePrecompiled, NNmodel); // CPU Var worker= WorkerFactory. CreateWorker(WorkerFactory.Type.CSharpBurst, NNmodel); After loading the model and creating a worker, we move on to running the model: Tensor input = new Tensor (batch, height, width, channels); Worker. Execute(input); After executing the model, we can collect the outputs through the instruction: var output = worker.Peekoutput(outputName); If the results are not satisfactory, we can repeat the execution of the model by varying the number of epochs or by modifying the structure. In our case, to improve the accuracy of the network, we modified the structure by inserting an additional convolutional layer obtaining a value of 99%. VII. D iscussion VIII. C onclusion We have analyzed the most common visualization techniques used to represent neural networks, a valuable tool for their development. Still, we have found that some are sophisticated tools that are difficult to use for inexperienced designers. Therefore, we have highlighted how to apply Virtual Reality to construct deep learning models. This type of representation, in fact, through the visualization of technical information on the network structure and on the data flow, represents an effective tool both for improving the forecasting capacity of neural networks and for training activities for those who do not have a profound knowledge of deep learning systems. We have experimented with applying it to a pre-trained neural network to recognize human activities (HAR). Still, we believe that it can effectively design all neural networks for image classification or object detection. A cknowledgments This work was developed as an integral component of the project: "An integrated System for indoor people localization, tracking, and monitoring" supported by the Italian MIUR under GRANT PON Research and Innovation 2014-2020 Project Code C35E190000200001, AIM1839112-1: Technologies for living environments . R eferences R éférences R eferencias Neural Network Design using a Virtual Reality Platform var worker = Global Journal of Computer Science and Technology Volume XXII Issue I Version I 45 ( )D Year 2022 © 2022 Global Journals 1. Shai Shalev-Shwartz, Shai Ben-David, "Understanding Machine Learning: From Theory to Algorithms," Cambridge University Press, 2014. 2. L. Arnold, S. Rebecchi, S. Chevallier, H. Paugam- Moisy," An Introduction to Deep Learning," European Symposium on Artificial Neural Networks (ESANN) , Bruges, Belgium, 2011. 3. P. G. Benardos, G. C. Vosniakos," Optimizing feedforward artificial neural network architecture," Engineering Applications of Artificial Intelligence , 20(3), pp. 365-382, 2007. 4. I. Goodfellow, Y. Bengio, A. Courville, “Deep Learning ," MIT Press , p. 32, 2016. 5. G. Montavon, W. Samek, K. R. Muller, "Methods for interpreting and understanding deep neural networks," Digital Signal Processing , 2017. 6. J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, H. Lipson, "Understanding neural networks through deep visualization," arXiv:1506.06579 , 2015. Our approach shows how the VR technique can effectively design deep learning applications. Our platform represents a valuable tool for developers with little knowledge of machine learning, allowing rapid design. It is an interactive technique also valid as a learning tool. Compared to existing visualization techniques, Virtual Reality is a potential highly developed for neural network applications. Above all, it has proved to be suitable for classifying images in different scientific sectors. We can assume that it is applicable for any deep learning.