[Colloquium] The epistemology of deep learning by Tom Sterkenburg

Epistemology is the branch of philosophy that is concerned with knowledge. It is concerned, for instance, with the fundamental limitations of automated methods for inferring knowledge from data: methods as developed within machine learning. A central insight, arrived at by philosophers and machine learning researchers alike, is that every learning algorithm must implement a particular inductive bias, or restrictive inductive assumptions about the learning situation at hand.
Deep learning constitutes perhaps the most successful exponent of modern machine learning. Recently, much discussion within the machine learning community has centered around what has been called an "apparent paradox" of deep learning: the fact that the empirical generalization performance of deep neural networks is astonishingly good, much better than our best mathematical theory can actually explain. To put it differently, we do not understand the inductive biases that deep neural networks implement.
In my talk I will discuss this debate from a philosophical perspective. Specifically, I will look at the two main lines of proposed explanation: the subtle manifestation of the inductive bias in the "implicit regularization" during the learning process, and the apparent fit of this inductive bias to the particular structure of real-world data. In addition, I will put into epistemological context a related debate within the community, that concerns the perceived tension between practical success and the demand of theoretical understanding.

Wie:

Tom Sterkenburg, postdoctoral fellow at the Munich Center for Mathematical Philosophy

Wat:

Colloquium

Waar:

tba (somewhere in the centre)

Wanneer:

12 june

Kosten:

none, free drinks afterwards at janskerkhof 13!