| KEYWORD |
Integrating high-level semantics in deep learning for computer vision
Reference persons FABRIZIO LAMBERTI
Research Groups GR-09 - GRAphics and INtelligent Systems - GRAINS
Description Deep neural networks, and especially convolutional neural networks, represent the state-of-the-art in tasks such as object detection and classification. Their success relies on the ability to learn intrinsic properties (features) with good discriminative capabilities from images, by encoding them as weights of a deep neural network. However, deep networks are quite difficult to understand, and cannot easily exploit high-level information about objects and their relationships (e.g. the fact that dogs and cats are both animals, or that a car has four wheels). Neural-symbolic learning and reasoning is an area of research that integrates modern machine learning (deep neural networks) with existing techniques from the fields of relational/symbolic reasoning and knowledge representation (first order logic, ontologies, knowledge bases and so forth). In this thesis, the student(s) will work to complement existing computer vision models with Knowledge Bases, based on existing public resources (WordNet, ConceptNet). Recently published techniques such as Logic Tensor Networks will be implemented and used to extend deep convolutional networks. Skills possessed or to be acquired: programming skills (Python, Keras/Tensorflow, or other deep learning framework); good analytical and mathematical skills.
See also http://grains.polito.it/work.php
Deadline 13/02/2019
PROPONI LA TUA CANDIDATURA