Figure 4
From: Learning physical properties of liquid crystals with deep convolutional neural networks
Predicting the sample temperature of E7 liquid crystals with convolutional neural networks. (A) Examples of experimental textures obtained from polarized optical microscope imaging for different samples at different temperatures (indicated within the images in degree Celsius). (B) Schematic representation of the network architecture used for the regression task of predicting the sample temperature. This network architecture is slightly different from all others we have used so far; it is composed of three blocks of eight 4 × 4 convolutions followed by other eight 4 × 4 convolutions and by eight 3 × 3 max-poling layers. After the last max-poling operations (S3), we have two fully connected layers (with 32 and 16 nodes) and an output layer with a single node and a linear activation function. All convolutional layers use a ReLU activation function. (C) Coefficient of determination estimated from the training and validation sets as a function of the number of training epochs. We separate 15% of data as test set and the remaining is divided into training (80%) and validation (20%) sets (all obtained in a stratified manner). The coefficient of determination calculated for the test set is ≈0.982. (D) Relationship between predicted and true temperature values obtained by applying the training network to the test set (the dashed line is the 1:1 relationship). The values between brackets in Panel A also indicate network predictions. (E) Coefficient of determination estimated from the test set as a function of the number of convolution (and max-pooling) blocks nb in the architecture (panel B corresponds to nb = 3). The markers represent the average values obtained from ten realizations of the training procedures, and the error bars are 95% confidence intervals.
