Towards adversarial learning principles in statistical relational artificial intelligence

 Many aspects of today’s life generate and store an increasing amount of data. This data has supported the accelerated developments in machine learning, a field that makes predictions based on existing data. On the forefront of these developments is deep learning, which powered many of the breakthroughs. The examples include DeepMind’s AlphaGo, Google’s driverless cars and Alexa’s speech recognition system. These advances were made possible due to deep learning's capability to process enormous amounts of data and extract useful patterns that can aid future predictions.  Despite the impressive progress, deep learning is still limited. Most importantly, it lacks the ability to perform reasoning, which prevents it from applying the extracted patterns beyond the situations it has previously experienced. This is often unacceptable: we would expect a driverless car to avoid a wall, even if it has not previously seen the exact same type of wall.  This research project aims to overcome this limitation by combining deep learning with machine learning methods based on mathematical logic. These methods are well known for their reasoning capabilities. This is a challenging problem, but with a great promise: the resulting methods capable of extracting complex patterns and reasoning would allow the state of the art machine learning to tackle problems currently out of reach.