Talk (Work in Progress): Cameron Beebe (MCMP/LMU)
Machine Analogies and Transfer Learning
An analogical or similarity-based notion has been argued to be fundamental to human cognition. Proponents note the 'fluid' ability to transfer problem-solving methods and draw similarities between tasks. I compare particular state-of-the-art notions in machine learning, namely transfer learning, and consider whether the protocols also satisfy the notion of analogy cited. Arguably, transfer learning as a framework for general machine intelligence is even more flexible than the fundamental role of analogy envisioned in human cognition. It would seem we are on the brink of machine intelligences that can achieve something that, for some, is fundamental to human cognition. I discuss the implications of such a view, and consider its limitations.