DIKU Bits - How do you know your machine has really learned?

Foredrag — It is no secret that the compute requirements for Machine Learning (ML) are growing at a rapid rate. A tremendous amount of effort has gone into making these ML systems more efficient by means of model compression, and with it reducing energy, memory, and compute resources. But how do we explain this compressibility, and importantly, do we know if these machines really learn? In this talk, we present the case for the ''learning is compression'' hypothesis and how it can be measured.