Location and Venue: Monday, 12.9.2016, 11.30-12.30 in the the Hotel Dormero in Hanover.
Representation Learning - I've heard that one before
(University of Stuttgart):
The revival of NNs surprised some, including me. Back then I considered NNs problematic especially because of their 'representational limitations' in comparison to the explicit structure that can be represented (and learned), e.g., with graphical models, or probabilistic relational models, or representing functions indirectly via optimization or planning problems, as often done in robotics. In fact, the limitation seemed not only w.r.t. representational capacity, but also w.r.t. the computational operations on such representations. It is however interesting to see that 'Representation Learning' became, again, a central research topic in the NN community. I introduce the talk discussing this controversy between the (perhaps feasible?) dream of learning everything in a generic, essentially 'no-prior' substrate ('end-to-end learning') versus the tough science of trying to identify what we believe is essential problem structure and learning relative to such priors. I mention some older work of mine as well as some newer that might seem to move away from the 'representation issue', but never really has.