Proposed Method Clause Samples

Proposed Method. Training: The trainining process (see top half of flowchart in Figure 1) is di- vided in four stages: registration, preprocessing of unlabeled data, feature extrac- tion and learning. The first step is to coarsely align all the volumes, labeled and unlabeled, to a template brain scan. The first volume in the dataset was arbitrar- ily chosen to be the template. This alignment makes it possible to use position features in the posterior classification. ITK (▇▇▇.▇▇▇.▇▇▇ ) was used to optimize an affine transform using a mutual information metric and a multi-resolution scheme. Using a nonlinear registration method could make the classifier rely too much on the registration through location features, making the method less robust. The next step is to preprocess the unlabeled volumes. The brain is first seg- mented using BET and FreeSurfer. The binary outputs of the two methods are then “softened” using a signed distance transform (positive if inside the brain, negative if outside). The distance map is mapped to the template space using the transforms from the registration step. The warped maps are used to calculate preliminary brain masks in the unlabeled scans by averaging the two maps for each volume and thresholding the result at zero, and they will also be used in the posterior semi-supervised learning step. The third step in the training stage is feature extraction. A pool of 58 image features is used in this study: (x, y, z) position, Gaussian derivatives of order up to two at five different scales (σ = 1.0, 2.0, 4.0, 8.0, 16.0 , in mm), and gradient magnitudes at the same scales. A subset of voxels from the training volumes is randomly selected for training purposes under the constraints that: 1) all scans contribute the same number of voxels; 2) 50% of the voxels have to be positives according to the annotated boundary (for the labeled scans) or the preliminary mask from the previous step (for the unlabeled); and 3) 50% of the voxels have to lie within 5mm of the boundary and 75% within 25mm. The features are normalized to zero mean and unit variance. Finally, a classifier can be trained using the labeled and unlabeled data. ▇▇▇▇▇▇▇’s random forests[12] were used as the base classifier because they Fig. 1. Flowchart for the training and test stages of the proposed algorithm compare favorably to other state-of-the-art algorithms[13]. Feature selection is performed through training a preliminary classifier with all the features and 20% of the available dat...

Proposed Method. The architecture mainly comprises of two units: TA and vehicle nodes. TA is responsible to authenticate and allow to register all vehicles in the VANET. As It is a trustable unit and not liable to attacks. Vehicle has to register itself with TA before join the network. Moreover, a wired connection network is being used to connect RSUs and LE. TA is also responsible to answer queries coming from any object. The vehicle nodes in VANET consists of LE, TV and MV, all are equipped with OBU or OBE. LEs are the authorized vehicles, acts as mobile TA. A MV is also considered as TV once it gets authenticated successfully with either LE. Figure 1 shows the process of becoming MV into TV. Let, three vehicles are present in a VANET, LE and 2 MVs with integrated OBUs (𝑂𝐵𝑈𝑖 and 𝑂𝐵𝑈𝑘 ). First vehicle completed the authentication process with LE to become TV and obtain adequate authorized parameters to authorize MVs, i.e., it acts like LE temporarily to authenticate the vehicle 𝑂𝐵𝑈𝑘. IEEE standard 1609.2 protocol is used by vehicle nodes to establish communication with LE, RSU and TVs preserve identification and few variables received from TS. It also, stores public/ private keys along with session keys used for communication with other VANET infrastructure components. Authentication OBU𝑖 (Mistrustful) OBU𝑘 (Mistrustful) Authentication LE(trustful) OBU𝑘 (Mistrustful) LE(trustful) OBUi (trustful) OBU𝑘 (trustful) OBUi (trustful) LE(trustful) Figure 1. MV to TV conversion procedure

Proposed Method. Based on the covariate shift assumption [167], we assume that there exist two distri- butions S(I; L) and T (I; L), where I and L donate images and labels, respectively. There are referred as the source distribution and the target distribution. Both dis- tributions are assumed to be very complex and unknown, and furthermore similar but different. In order to obtain a similar ▇▇▇▇ performance on both target domain without labels and source domain with labels, we should constrain two distributions (i.e., S and T ) to be similar. Unfortunately the distributions are unknown and can be very complex, which makes this problem difficult to solve. So, we reversely con- sider this problem that making two distributions as different as possible is equivalent to classifying them. With the help of gradient reversal layer (GRL), we can transfer the classified supervised signal to an indiscriminate (domain-invariant) supervised signal, which is formulated as, LDC (I) = −|I| (ΓI Σ I∈I log P(I) + (1 − ΓI) log(1 − P(I))) 1 , I ∈ It ΓI 0 , I ∈ Is , where ΓI is an indicator function to index which domain image I belongs to. During the backpropagation processing, GRL makes the gradient negative and feeds it ▇▇▇▇ to next layer, which is formulated as, θ → θ − α(∂LO − ∂LDC ), (5.4) where LO represents the loss functions other than LDC, θ are all the parameters of the whole nueral network and α is the step size of SGD. Eventually, the backbone network learns to generate domain-invariant representa- tion. Note that the use of GRL during the several initial epochs of training is not stable. This is because the backbone network is struggling to find the optimal path in the beginning, due to the entangled supervision signal yielded by GRL. After model has acquired robust representative capability, the GRL guides the representation towards better generalization.

Proposed Method. We propose Graph Agreement Models (GAM), a novel approach that aims to resolve the main limitation of label propagation methods while leveraging their strengths. Instead of using the edge weights as a fixed measure of how much the labels of two nodes should agree, GAM learns the probability of agreement. To achieve this, we introduce an agreement model, g, that takes as input the features of two nodes and (optionally) the weight of the edge between them, and predicts the probability that they have the same label. The predicted agreement probabilities are then used when training the classification model, f , to prevent overfitting.

Proposed Method calculation of residuals SA( T ) p,q S E p,q p,q r N p,q

Proposed Method take the high-level feedback signal into feature extraction instead of extracting fea- ture vectors based on one-pass of the data through the network forward.