Geforce gtx 1060

Firstly, a deep feed forward neural network was trained with the idealized simulated dataset_train,sim. To find an optimal network architecture for our problem of predicting end-effector wrenches, an autonomous hyperparameter optimization was done using the optuna toolbox (Akiba et al., 2019 ) and the LAMB optimizer (You et al., 2019 ). In Table 3 the static as well as the varied neural network parameters and their ranges are listed. The loss function for training a particular neural network is an optimization parameter and can vary between mean squared error and mean absolute error (see Table 3). In order to compare the different tested network architectures with an equal metric to find the neural network with the highest prediction accuracy, we calculated a separate evaluation error using the root mean squared error. Training was done on four GPUs of a DGX-2. In total 700 different architectures were tried, which needed a computing time of approximately around 19 days. The resulting optimal neural network model with the highest prediction accuracy is named NN_sim. The trained model was used on a regular desktop PC, with an Intel core i7 CPU and a nVidia GeForce GTX 1060. One evaluation step takes 21.5 ms.

A convolutional neural network (CNN) for semantic image segmentation composed of an encoder and corresponding decoder subnetwork was set up [2 (link)]. The network was pre-initialized with layers and weights from a pre-trained VGG 16 model [32 ].
The network used a pixel classification layer to predict the categorical label for every pixel in the input images. Class frequency of CSF (8.6%), brain (22.1%), tissue (14.3%), and background (55.0%) was obtained. Since the class “CSF” was underrepresented in the training data, a class weighting was carried out to balance classes.
A stochastic gradient descent with momentum (0.9) optimizer was used and a regularization term for the weights to the loss function was added with a weight decay of 0.0005. Cross-entropy was used as a loss function for optimizing the classification model. The initial learning rate was set to 0.001. Furthermore, the learning rate was reduced by a factor of 0.3 every 10 epochs. The network was tested against the validation data set every epoch to stop training when the validation accuracy converged. This prevented the network from overfitting on the training data set. The training was conducted on a single GPU (NVIDIA GeForce GTX 1060). The validation accuracy converged after 6000 repetitions.

The experiment was performed using a PC equipped with an NVidia GeForce GTX 1060 graphics card. The setup included the Oculus Rift (oculus.com/rift) head-mounted display (HMD) with 2160 × 1200 resolution (1080 × 1200 per eye), 110° field of view, and 90 Hz refresh rate for 3D immersive viewing, head rotational, and positional tracking, and providing spatialized audio. The application was created using the Unity (version 2018.2.1) game engine and the environment using Autodesk Maya and Adobe Photoshop. The virtual characters were designed and rigged using Autodesk Character Generator. For the lip synchronization feature, the SALSA plug-in for Unity was used.