46 G. P. Tsialiamanis et al. Fig. 5.4 Principal components of all samples (a), samples excepting panels 3 and 6 (b) and samples of panels 3 and 6 (c) Table 5.2 Confusion Matrix of neural network classifier trained on the first dataset, test set, total accuracy: 98.48% Predicted panel 1 2 4 5 7 8 9 Missing panel 1 65 1 0 0 0 0 0 Missing panel 2 0 63 1 0 0 0 2 Missing panel 4 1 0 65 0 0 0 0 Missing panel 5 0 0 0 66 0 0 0 Missing panel 7 0 0 0 0 66 0 0 Missing panel 8 1 0 0 0 0 65 0 Missing panel 9 1 0 0 0 0 0 65 task, as the input and hidden layers of the target task; this means that only the weights between the hidden and output layers remain to be trained for the target task. This strategy reduces the number of parameters considerably. The functional form of the network for the source task is given by, y =f0W2(f1(W1x+b1)) +b2 (5.4) where f0 and f1 are the non-linear activation functions of the output layer and the hidden layer respectively, W1,2 are the weight matrices of the transformations between the layers, b1,2 are the bias vectors of the layers, x is the input vector and y the output vector. The softmax function is chosen to be the activation function of the decision layer, as this is appropriate
RkJQdWJsaXNoZXIy MTMzNzEzMQ==