Colloquium Ehrenfestii: the Physics Inside Deep Learning
- Wednesday 15 January 2020
Niels Bohrweg 2
2333 CA Leiden
- De Sitterzaal
The Physics Inside Deep Learning
Physics has produced a number of very elegant elegant theories about how our world is organized. From the partial differential Navier-Stokes equations describing fluid mechanics, via Riemannian differential geometry to describe general relativity, quantum mechanics to describe atoms, to quantum field theories to describe the elementary particles.
In a seemingly completely unrelated scientific discipline, machine learning has made great strides in analyzing signals such as speech, image and video. In particular, convolutional neural nets have been extremely successful at processing these raw signals to build abstract representations from which highly accurate predictions can be made. Examples include: classifying skin cancer from smartphone photos, transcribing speech into sentences, translating these sentences to other languages and then synthesizing them back into speech, segmenting and classifying objects in a video stream in real time, and defeating the world champion of GO.
What do these two so seemingly different fields have in common? In this talk I will describe our amazing journey of incorporating the principles of physics into these deep learning architectures. From incorporating global symmetries into neural networks (group-CNNs) to deep learning on curved manifold resulting (gauge-CNNs). From using the ideas of quantum mechanics, such as entanglement to describe a new class of "quantum neural networks" (QubeNet) to hybrid systems that combine learned convolutions with PDEs.
I will leave the audience with the intriguing question if these synergies are coincidental or whether there is a deeper level to understand this as information processing systems? As John Archibald Wheeler would put it to Claude Shannon: Is it all 'it from bit'?