Figure 1. CNN architecture (adopted from Krizhevsky et al. '12)-Measuring Customer Behavior with Deep Convolutional Neural Networks

<p>The architecture of a CNN can be described as following. A small pixel region goes to input neurons and then connects to a first convolution hidden layer (Figure1). There we can see a set of learnable filters, which are activated during the presentation some particular type of feature in pixel region in the input. On this phase, CNN does shift invariance, which is carried by feature map. Subsampling layer goes next. There we have two processes: local averaging and sampling. As a result, we get declining resolution of feature map. To correspond this task CNN needs supervised learning. Before starting the experiment, we gave a set of labeled videos with different emotional experience. The system analyses images and finds similar features. Then the system creates a map, where it arranges videos in accordance with similar features. Thereby, images with similar emotions form certain class. &nbsp;To test the system, we add other videos and correct the system when it refers them improperly. The proposed model consists of four convolutional layers, followed by max-pooling layers, and three fully-connected layers with a final classificatory presented with MLP (with six basic outputs, corresponding to basic emotions for emotion classification and two outputs for motion classification for typical and non-typical behavior). The input data was presented as infrared camera output.</p>