Σεμινάριο Τμήματος:»Conditional random fields with privileged information and its applications to human activity and facial expression recognition»
Στο πλαίσιο της διοργάνωσης των σεμιναρίων του τμήματος θα πραγματοποιηθεί την Παρασκευή 14/12/2018 και ώρα 12:00 στην αίθουσα Σεμιναρίων του Τμήματος Μηχανικών Η/Υ και Πληροφορικής, ομιλία με τίτλο «Conditional random fields with privileged information and its applications to human activity and facial expression recognition». Ομιλητής θα είναι ο κ. Μιχάλης Βρίγκας, Τμήμα Μηχανικών Η/Υ & Πληροφορικής, Πανεπιστήμιο Ιωαννίνων.
Recent advances in computer vision such as video surveillance and human-machine interactions rely on machine learning techniques trained on large-scale human annotated datasets. However, training data may not always be available during testing and learning using privileged information (LUPI) has been used to overcome this problem. The insight of privileged information is that one may have access to additional information about the training samples, which is not available during testing. The LUPI technique simulates a real-life learning condition when a student learns from his/her teacher, where the latter provides the student with additional knowledge, comments, explanations, or rewards in class. Subsequently, the student should be able to face any problem related to what he/she has learned without the help of the teacher.
A supervised probabilistic approach that integrates LUPI into a hidden conditional random field (HCRF) model for recognizing human activities will be presented. This model facilitates the training process by learning the conditional probability distribution between human activities and observations. Moreover, the method provides robustness to outliers (such as noise or missing data) by modeling the conditional distribution of the privileged information by a Student’s t-density function, which is naturally integrated into the model. Moreover, a novel probabilistic model, which incorporates the LUPI paradigm into a unified framework for recognizing facial expressions and affective states of a person, will also be presented. This model aims to indirectly transfer the knowledge from privileged to the original feature space using conditional random fields (CRFs) through a two-step classification process. First, a standard CRF model on the privileged data is trained to encode the ability of privileged information to distinguish between different class labels into the model weights. The learned privileged weights are then used to penalize the training process on the original feature space by learning the conditional probability distribution between the class labels and original observations.