Philosophies | Special Issue : Philosophy and Epistemology of Deep Learning

Call For Papers:
Deadline for manuscript submissions: 15 March 2019
Guest Editors:
Dr. Hector Zenil
Prof. Dr. Selmer Bringsjord
Current popular approaches to Machine Learning (ML), Deep Learning (DL) and Artificial Intelligence (AI) are mostly statistical in nature, and are not well equipped to deal with abstraction and explanation. In particular, they cannot generate candidate models or make generalizations directly from data to discover possible causal mechanisms. One method that researchers are resorting to in order to discover how deep learning algorithms work involves using what are called ‘generative models’ (a possible misnomer). They train a learning algorithm and handicap it systematically whilst asking it to generate examples. By observing the resulting examples they are able to make inferences about what may be happening in the algorithm at some level. 
However, current trends and methods are widely considered black-box approaches that have worked amazingly well in classification tasks, but provide little to no understanding of causation and are unable to deal with forms of symbolic computation such as logical inference and explanation. As a consequence, they also fail to be scalable in domains they have not been trained for, and require tons of data to be trained on, before they can do anything interesting—-and they require training every time they are presented with (even slightly) different data. 
Furthermore, how other cognitive features, such as human consciousness, may be related to current and future directions in deep learning, and whether such features may prove advantageous or disadvantageous remains an open question.
The aim of this special issue is thus to attempt to ask the right questions and shed some light on the achievements, limitations and future directions in reinforcement/deep learning approaches and differentiable programming. Its particular focus will be on the interplay of data and model-driven approaches that go beyond current ones, which for the most part are  based on traditional statistics. It will attempt to ascertain whether a fundamental theory is needed or whether one already exists, and to explore the implications of current and future technologies based on deep learning and differentiable programming for science, technology and society.

Special issue website: