糖心传媒

Dr. Cameron Buckner

University of Houston

Unless you have been living under a rock for the last few years, you have heard about the incredible recent achievements of deep neural networks. These artificial intelligence methods seem to achieve "superhuman" performance on new domains weekly, from beating grandmasters at the ancient Chinese game of Go to driverless navigation to automated image classification to winning at video games, to name just a few. However, it has proven difficult to explain how or why deep neural networks perform so well.

In philosophy of mind, empiricists like Locke and Hume argued that complex cognition is based on information derived from sensory experience, often appealing to a faculty of abstraction. Rationalists have frequently complained, however, that they never adequately explained how this faculty of abstraction actually works, charging that abstraction is just empiricist "magic".

In this talk, I tie these two problems together, to the mutual benefit of both disciplines. I argue that the engineering tweaks that distinguish deep networks from their shallower forebears explains how they (and the brain) implement a form of hierarchical processing that I call “transformational abstraction”. Transformational abstraction iteratively converts sensory-based representations of category exemplars into more generic formats that are increasingly tolerant to “nuisance variation”, one of the main systematic challenges that cognition must overcome to reliably succeed in the natural world. I here illustrate the power of transformational abstraction through a series of examples, including classic philosophical conundrums from Locke and Hume as well as more contemporary applications like Go strategy and abstract arts."