Marijn Stollenga - Using Guided Autoencoders on Face Recognition

afstudeercolloquium

In this master thesis we will create guided autoencoders (GAE) and apply them to face recognition. GAEs are agents that interact with the image. They perceive part of an image through a window, use an autoencoder to encode it, and react to what they see by moving the window. They are trained to move the window to find and encode specific parts of the face -- in our case the eyes, nose and mouth. We use the LFWC (cropped Labeled Faces in the Wild) dataset which is very varied and has many uncontrolled variables. We train GAEs using the CACLA reinforcement learning algorithm which can deal with continuous states and actions. To create a state, GAEs evaluate their separately trained autoencoder on what is visible through their window. They use the resulting encoding as a state to guide their actions. We show that GAEs can find the way to their parts provided they are relatively close to them. We use both shallow and deep (stacked) autoencoders. Surprisingly, deep GAEs do not outperform shallow GAEs on this task. The GAEs are finally used to classify the gender of faces and whether a person is smiling or not. Their classification performance is lower than expected due to the unstable nature of GAEs. At their current performance, GAEs are applicable for fine-tuning guesses of positions of parts of the face. We discuss possible future directions to address the unstable nature. In summary, the GAEs are not stable enough to be applied directly on classification, but their ability to find parts is promising. We think that if the stability issues are addressed, GAEs can improve automatic facerecognition and computer vision systems in general.