Anders C. Hansen: How intelligent is artificial intelligence? - On the surprising and mysterious secrets of deep learning. Part I
This is the first in a series of three lectures by Anders C. Hansen (Cambridge Univ. and UiO) on this topic. Vegard Antun (UiO) will also contribute to the lectures.
Artificial intelligence (AI) is changing the world in front of our eyes with applications ranging from self-driving cars, via automated diagnosis in medicine, to new methods in imaging sciences and inverse problems. A timely question is therefore: how intelligent is modern AI, and can it be trusted? The two big changes for mankind and science in general are as follows:
(1) AI will replace humans in problem solving, for example in providing medical diagnosis. This is already happening, as the US Food and Drug Administration (FDA) has approved (in April 2018) AI techniques for automated diagnosis without the input of a human clinician.
(2) AI will replace standard algorithms used in the sciences. This is also happening with full force, and imaging sciences are now full of new AI techniques promising to replace standard algorithms, for example in Magnetic Resonance Imaging (MRI).
Since its revolutionary results in computer vision in 2012, deep learning is now the leading tool in state-of-the-art AI. However, recent results reveal surprising and mysterious phenomena of deep learning. In particular, deep learning typically provides completely unstable algorithms, and this seems to be a universal phenomenon. This phenomenon has the following consequences:
(1) AI replacing human activity may have mysterious non-human behaviour. For example, two images of a mole, that to the human doctor look identical, may by the AI algorithm be classified completely differently: one with 99.9% certainty that it is benign and the other with 99.9% certainty that it is cancerous. A sym-phony may be classified by the AI algorithm as someone's voice, and a stop sign may be viewed as a 45 mph sign.
(2) AI replacing standard algorithms may have very surprising behaviour. For example, an AI algorithm used in MRI may remove an important detail in the MRI image, for example a brain tumour, or add a non-existing detail causing an incorrect diagnosis.
The above phenomena suggest the question: how intelligent is modern AI? We will discuss this fundamental question and also a new radical theory explaining why the above puzzling phenomena happen and that a cure is not in anyway close. This suggests some rather delicate ethical and moral questions about the use deep learning based methods and modern AI in society.